mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2026-02-12 14:03:20 +02:00
Compare commits
69 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
73cd5e1b97 | ||
|
|
8ee538ce73 | ||
|
|
6d95707827 | ||
|
|
89181c0b6d | ||
|
|
ceaa89b786 | ||
|
|
2cce9fddb7 | ||
|
|
612db61886 | ||
|
|
57487a64c8 | ||
|
|
fc0fe40049 | ||
|
|
9a96352729 | ||
|
|
c03a5a46f0 | ||
|
|
6948adc90d | ||
|
|
854b09f0d7 | ||
|
|
66d403c480 | ||
|
|
f0bfe54f55 | ||
|
|
52e38faf8c | ||
|
|
a0d585537c | ||
|
|
98e57ca422 | ||
|
|
262364e31d | ||
|
|
820ebfa6f4 | ||
|
|
292f6908cd | ||
|
|
81ddc60cb3 | ||
|
|
972f323e73 | ||
|
|
f5e7734ff2 | ||
|
|
1e8924fd65 | ||
|
|
39bf692af1 | ||
|
|
e06088da0f | ||
|
|
5fa1c190d9 | ||
|
|
eb449cdfa4 | ||
|
|
5999b50eb0 | ||
|
|
9a5f57795c | ||
|
|
96441c955e | ||
|
|
8872ad2125 | ||
|
|
34ba7b5a2f | ||
|
|
b83111815e | ||
|
|
3228e77287 | ||
|
|
7fbd36c50c | ||
|
|
537eadb1b9 | ||
|
|
db6adb3c88 | ||
|
|
dfde5993ea | ||
|
|
06bf3796f4 | ||
|
|
3688c4f504 | ||
|
|
1946e46f4c | ||
|
|
f9bd518a6b | ||
|
|
7fcf1ef45d | ||
|
|
e696cfc016 | ||
|
|
3e21647666 | ||
|
|
22cae83218 | ||
|
|
449ec2ab07 | ||
|
|
3795cc1e89 | ||
|
|
b828e18c75 | ||
|
|
a4ea7a188f | ||
|
|
7a4f97d196 | ||
|
|
a498c75ad1 | ||
|
|
3409ab842d | ||
|
|
c342c3b93d | ||
|
|
af252d0758 | ||
|
|
11fb327bf3 | ||
|
|
e6e934c5ea | ||
|
|
b536eb0233 | ||
|
|
e0c93af2a0 | ||
|
|
423bee462b | ||
|
|
8abcc70a74 | ||
|
|
eaba92c3dc | ||
|
|
6ab881b7c3 | ||
|
|
d838c22bb3 | ||
|
|
25f40ca65f | ||
|
|
015deb9048 | ||
|
|
2ceda3f662 |
@@ -54,6 +54,7 @@ RUN apt-get update \
|
||||
build-essential \
|
||||
git \
|
||||
python3 \
|
||||
python3-dev \
|
||||
python3-pip \
|
||||
python3-wheel \
|
||||
&& pip install --break-system-packages --upgrade setuptools \
|
||||
|
||||
6
.github/workflows/build.yml
vendored
6
.github/workflows/build.yml
vendored
@@ -293,7 +293,9 @@ jobs:
|
||||
cmake -B build \
|
||||
-DLLAMA_FATAL_WARNINGS=ON \
|
||||
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
|
||||
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
|
||||
|
||||
- name: Build (no OpenMP)
|
||||
@@ -303,8 +305,10 @@ jobs:
|
||||
cmake -B build \
|
||||
-DLLAMA_FATAL_WARNINGS=ON \
|
||||
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
|
||||
-DGGML_OPENMP=OFF
|
||||
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
|
||||
|
||||
- name: Test
|
||||
@@ -466,7 +470,7 @@ jobs:
|
||||
export GGML_VK_VISIBLE_DEVICES=0
|
||||
export GGML_VK_DISABLE_F16=1
|
||||
# This is using llvmpipe and runs slower than other backends
|
||||
ctest -L main --verbose --timeout 4200
|
||||
ctest -L main --verbose --timeout 4800
|
||||
|
||||
ubuntu-24-cmake-webgpu:
|
||||
runs-on: ubuntu-24.04
|
||||
|
||||
73
.github/workflows/server-metal.yml
vendored
Normal file
73
.github/workflows/server-metal.yml
vendored
Normal file
@@ -0,0 +1,73 @@
|
||||
name: Server-Metal
|
||||
|
||||
on:
|
||||
workflow_dispatch: # allows manual triggering
|
||||
inputs:
|
||||
sha:
|
||||
description: 'Commit SHA1 to build'
|
||||
required: false
|
||||
type: string
|
||||
slow_tests:
|
||||
description: 'Run slow tests'
|
||||
required: true
|
||||
type: boolean
|
||||
push:
|
||||
branches:
|
||||
- master
|
||||
paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
|
||||
|
||||
env:
|
||||
LLAMA_LOG_COLORS: 1
|
||||
LLAMA_LOG_PREFIX: 1
|
||||
LLAMA_LOG_TIMESTAMPS: 1
|
||||
LLAMA_LOG_VERBOSITY: 10
|
||||
|
||||
concurrency:
|
||||
group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
|
||||
cancel-in-progress: true
|
||||
|
||||
jobs:
|
||||
server-metal:
|
||||
runs-on: [self-hosted, macOS, ARM64]
|
||||
|
||||
name: server-metal (${{ matrix.wf_name }})
|
||||
strategy:
|
||||
matrix:
|
||||
build_type: [Release]
|
||||
wf_name: ["GPUx1"]
|
||||
include:
|
||||
- build_type: Release
|
||||
extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
|
||||
wf_name: "GPUx1, backend-sampling"
|
||||
- build_type: Release
|
||||
extra_args: "GGML_METAL_DEVICES=2"
|
||||
wf_name: "GPUx2"
|
||||
- build_type: Release
|
||||
extra_args: "GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1"
|
||||
wf_name: "GPUx2, backend-sampling"
|
||||
fail-fast: false
|
||||
|
||||
steps:
|
||||
- name: Clone
|
||||
id: checkout
|
||||
uses: actions/checkout@v6
|
||||
with:
|
||||
fetch-depth: 0
|
||||
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
|
||||
|
||||
- name: Build
|
||||
id: cmake_build
|
||||
run: |
|
||||
cmake -B build -DGGML_SCHED_NO_REALLOC=ON
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
|
||||
|
||||
- name: Tests
|
||||
id: server_integration_tests
|
||||
if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
|
||||
run: |
|
||||
cd tools/server/tests
|
||||
python3 -m venv venv
|
||||
source venv/bin/activate
|
||||
pip install -r requirements.txt
|
||||
export ${{ matrix.extra_args }}
|
||||
pytest -v -x -m "not slow"
|
||||
120
.github/workflows/server-webui.yml
vendored
120
.github/workflows/server-webui.yml
vendored
@@ -8,10 +8,6 @@ on:
|
||||
description: 'Commit SHA1 to build'
|
||||
required: false
|
||||
type: string
|
||||
slow_tests:
|
||||
description: 'Run slow tests'
|
||||
required: true
|
||||
type: boolean
|
||||
push:
|
||||
branches:
|
||||
- master
|
||||
@@ -101,119 +97,3 @@ jobs:
|
||||
if: ${{ always() && steps.playwright.conclusion == 'success' }}
|
||||
run: npm run test:e2e
|
||||
working-directory: tools/server/webui
|
||||
|
||||
server-build:
|
||||
runs-on: ubuntu-latest
|
||||
|
||||
strategy:
|
||||
matrix:
|
||||
sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
|
||||
build_type: [RelWithDebInfo]
|
||||
include:
|
||||
- build_type: Release
|
||||
sanitizer: ""
|
||||
fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
|
||||
|
||||
steps:
|
||||
- name: Dependencies
|
||||
id: depends
|
||||
run: |
|
||||
sudo apt-get update
|
||||
sudo apt-get -y install \
|
||||
build-essential \
|
||||
xxd \
|
||||
git \
|
||||
cmake \
|
||||
curl \
|
||||
wget \
|
||||
language-pack-en \
|
||||
libssl-dev
|
||||
|
||||
- name: Clone
|
||||
id: checkout
|
||||
uses: actions/checkout@v6
|
||||
with:
|
||||
fetch-depth: 0
|
||||
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
|
||||
|
||||
- name: Python setup
|
||||
id: setup_python
|
||||
uses: actions/setup-python@v6
|
||||
with:
|
||||
python-version: '3.11'
|
||||
|
||||
- name: Tests dependencies
|
||||
id: test_dependencies
|
||||
run: |
|
||||
pip install -r tools/server/tests/requirements.txt
|
||||
|
||||
- name: Setup Node.js for WebUI
|
||||
uses: actions/setup-node@v6
|
||||
with:
|
||||
node-version: "22"
|
||||
cache: "npm"
|
||||
cache-dependency-path: "tools/server/webui/package-lock.json"
|
||||
|
||||
- name: Install WebUI dependencies
|
||||
run: npm ci
|
||||
working-directory: tools/server/webui
|
||||
|
||||
- name: Build WebUI
|
||||
run: npm run build
|
||||
working-directory: tools/server/webui
|
||||
|
||||
- name: Build (no OpenMP)
|
||||
id: cmake_build_no_openmp
|
||||
if: ${{ matrix.sanitizer == 'THREAD' }}
|
||||
run: |
|
||||
cmake -B build \
|
||||
-DGGML_NATIVE=OFF \
|
||||
-DLLAMA_BUILD_SERVER=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
|
||||
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DGGML_OPENMP=OFF ;
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
|
||||
|
||||
- name: Build (sanitizers)
|
||||
id: cmake_build_sanitizers
|
||||
if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
|
||||
run: |
|
||||
cmake -B build \
|
||||
-DGGML_NATIVE=OFF \
|
||||
-DLLAMA_BUILD_SERVER=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
|
||||
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
|
||||
|
||||
- name: Build (sanitizers)
|
||||
id: cmake_build
|
||||
if: ${{ matrix.sanitizer == '' }}
|
||||
run: |
|
||||
cmake -B build \
|
||||
-DGGML_NATIVE=OFF \
|
||||
-DLLAMA_BUILD_SERVER=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
|
||||
|
||||
- name: Tests
|
||||
id: server_integration_tests
|
||||
if: ${{ matrix.sanitizer == '' }}
|
||||
env:
|
||||
GITHUB_ACTIONS: "true"
|
||||
run: |
|
||||
cd tools/server/tests
|
||||
./tests.sh
|
||||
|
||||
- name: Tests (sanitizers)
|
||||
id: server_integration_tests_sanitizers
|
||||
if: ${{ matrix.sanitizer != '' }}
|
||||
run: |
|
||||
cd tools/server/tests
|
||||
LLAMA_SANITIZE=1 ./tests.sh
|
||||
|
||||
- name: Slow tests
|
||||
id: server_integration_tests_slow
|
||||
if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
|
||||
run: |
|
||||
cd tools/server/tests
|
||||
SLOW_TESTS=1 ./tests.sh
|
||||
|
||||
22
.github/workflows/server.yml
vendored
22
.github/workflows/server.yml
vendored
@@ -81,18 +81,14 @@ jobs:
|
||||
-DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
|
||||
-DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
|
||||
-DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
|
||||
cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
|
||||
|
||||
- name: Python setup
|
||||
id: setup_python
|
||||
uses: actions/setup-python@v6
|
||||
with:
|
||||
python-version: '3.11'
|
||||
|
||||
- name: Tests dependencies
|
||||
id: test_dependencies
|
||||
run: |
|
||||
pip install -r tools/server/tests/requirements.txt
|
||||
pip-install: -r tools/server/tests/requirements.txt
|
||||
|
||||
- name: Tests
|
||||
id: server_integration_tests
|
||||
@@ -102,6 +98,14 @@ jobs:
|
||||
export ${{ matrix.extra_args }}
|
||||
pytest -v -x -m "not slow"
|
||||
|
||||
- name: Slow tests
|
||||
id: server_integration_tests_slow
|
||||
if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
|
||||
run: |
|
||||
cd tools/server/tests
|
||||
export ${{ matrix.extra_args }}
|
||||
SLOW_TESTS=1 pytest -v -x
|
||||
|
||||
server-windows:
|
||||
runs-on: windows-2022
|
||||
|
||||
@@ -124,11 +128,7 @@ jobs:
|
||||
uses: actions/setup-python@v6
|
||||
with:
|
||||
python-version: '3.11'
|
||||
|
||||
- name: Tests dependencies
|
||||
id: test_dependencies
|
||||
run: |
|
||||
pip install -r tools/server/tests/requirements.txt
|
||||
pip-install: -r tools/server/tests/requirements.txt
|
||||
|
||||
- name: Tests
|
||||
id: server_integration_tests
|
||||
|
||||
@@ -109,6 +109,7 @@ option(LLAMA_BUILD_TOOLS "llama: build tools" ${LLAMA_STANDALONE})
|
||||
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
|
||||
option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE})
|
||||
option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
|
||||
option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
|
||||
|
||||
# 3rd party libs
|
||||
option(LLAMA_HTTPLIB "llama: httplib for downloading functionality" ON)
|
||||
|
||||
@@ -27,6 +27,7 @@
|
||||
/examples/batched.swift/ @ggerganov
|
||||
/examples/batched/ @ggerganov
|
||||
/examples/convert-llama2c-to-ggml/ @ggerganov
|
||||
/examples/debug/ @danbev @pwilkin
|
||||
/examples/deprecation-warning/ @ggerganov
|
||||
/examples/diffusion/ @am17an
|
||||
/examples/embedding/ @ggerganov
|
||||
|
||||
@@ -288,6 +288,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
|
||||
| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
|
||||
| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
|
||||
| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
|
||||
| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
|
||||
|
||||
## Obtaining and quantizing models
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ g++ --version
|
||||
g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
|
||||
|
||||
nvidia-smi
|
||||
Sun Nov 2 10:43:25 2025
|
||||
Thu Feb 5 13:49:40 2026
|
||||
+-----------------------------------------------------------------------------------------+
|
||||
| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 |
|
||||
+-----------------------------------------+------------------------+----------------------+
|
||||
@@ -17,7 +17,7 @@ Sun Nov 2 10:43:25 2025
|
||||
| | | MIG M. |
|
||||
|=========================================+========================+======================|
|
||||
| 0 NVIDIA GB10 On | 0000000F:01:00.0 Off | N/A |
|
||||
| N/A 35C P8 4W / N/A | Not Supported | 0% Default |
|
||||
| N/A 47C P0 13W / N/A | Not Supported | 0% Default |
|
||||
| | | N/A |
|
||||
+-----------------------------------------+------------------------+----------------------+
|
||||
```
|
||||
@@ -29,46 +29,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.374 | 1369.01 | 0.383 | 83.64 | 0.757 | 719.01 |
|
||||
| 512 | 32 | 2 | 1088 | 0.274 | 3741.35 | 0.659 | 97.14 | 0.933 | 1166.66 |
|
||||
| 512 | 32 | 4 | 2176 | 0.526 | 3896.47 | 0.817 | 156.73 | 1.342 | 1621.08 |
|
||||
| 512 | 32 | 8 | 4352 | 1.044 | 3925.10 | 0.987 | 259.44 | 2.030 | 2143.56 |
|
||||
| 512 | 32 | 16 | 8704 | 2.076 | 3945.84 | 1.248 | 410.32 | 3.324 | 2618.60 |
|
||||
| 512 | 32 | 32 | 17408 | 4.170 | 3929.28 | 1.630 | 628.40 | 5.799 | 3001.76 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.083 | 3782.66 | 0.394 | 81.21 | 1.477 | 2795.13 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.166 | 3782.72 | 0.725 | 88.28 | 2.891 | 2856.14 |
|
||||
| 4096 | 32 | 4 | 16512 | 4.333 | 3780.88 | 0.896 | 142.82 | 5.230 | 3157.38 |
|
||||
| 4096 | 32 | 8 | 33024 | 8.618 | 3802.14 | 1.155 | 221.69 | 9.773 | 3379.08 |
|
||||
| 4096 | 32 | 16 | 66048 | 17.330 | 3781.73 | 1.598 | 320.34 | 18.928 | 3489.45 |
|
||||
| 4096 | 32 | 32 | 132096 | 34.671 | 3780.48 | 2.336 | 438.35 | 37.007 | 3569.51 |
|
||||
| 8192 | 32 | 1 | 8224 | 2.233 | 3668.56 | 0.438 | 72.98 | 2.671 | 3078.44 |
|
||||
| 8192 | 32 | 2 | 16448 | 4.425 | 3702.95 | 0.756 | 84.66 | 5.181 | 3174.95 |
|
||||
| 8192 | 32 | 4 | 32896 | 8.859 | 3698.64 | 0.967 | 132.38 | 9.826 | 3347.72 |
|
||||
| 8192 | 32 | 8 | 65792 | 17.714 | 3699.57 | 1.277 | 200.52 | 18.991 | 3464.35 |
|
||||
| 8192 | 32 | 16 | 131584 | 35.494 | 3692.84 | 1.841 | 278.12 | 37.335 | 3524.46 |
|
||||
| 8192 | 32 | 32 | 263168 | 70.949 | 3694.82 | 2.798 | 365.99 | 73.747 | 3568.53 |
|
||||
| 512 | 32 | 1 | 544 | 0.270 | 1895.57 | 0.399 | 80.13 | 0.669 | 812.60 |
|
||||
| 512 | 32 | 2 | 1088 | 0.230 | 4451.23 | 0.583 | 109.71 | 0.813 | 1337.56 |
|
||||
| 512 | 32 | 4 | 2176 | 0.437 | 4688.87 | 0.820 | 156.03 | 1.257 | 1730.91 |
|
||||
| 512 | 32 | 8 | 4352 | 0.863 | 4744.23 | 0.942 | 271.79 | 1.805 | 2410.73 |
|
||||
| 512 | 32 | 16 | 8704 | 1.725 | 4748.19 | 1.173 | 436.38 | 2.899 | 3002.85 |
|
||||
| 512 | 32 | 32 | 17408 | 3.437 | 4767.38 | 1.503 | 681.49 | 4.939 | 3524.40 |
|
||||
| 4096 | 32 | 1 | 4128 | 0.907 | 4513.91 | 0.407 | 78.54 | 1.315 | 3139.56 |
|
||||
| 4096 | 32 | 2 | 8256 | 1.796 | 4560.42 | 0.625 | 102.37 | 2.422 | 3409.45 |
|
||||
| 4096 | 32 | 4 | 16512 | 3.596 | 4555.66 | 0.888 | 144.11 | 4.485 | 3681.93 |
|
||||
| 4096 | 32 | 8 | 33024 | 7.184 | 4561.44 | 1.098 | 233.11 | 8.282 | 3987.51 |
|
||||
| 4096 | 32 | 16 | 66048 | 14.369 | 4560.82 | 1.503 | 340.74 | 15.872 | 4161.30 |
|
||||
| 4096 | 32 | 32 | 132096 | 28.760 | 4557.52 | 2.162 | 473.59 | 30.922 | 4271.95 |
|
||||
| 8192 | 32 | 1 | 8224 | 1.859 | 4405.59 | 0.430 | 74.36 | 2.290 | 3591.61 |
|
||||
| 8192 | 32 | 2 | 16448 | 3.698 | 4430.02 | 0.656 | 97.59 | 4.354 | 3777.47 |
|
||||
| 8192 | 32 | 4 | 32896 | 7.403 | 4426.10 | 0.957 | 133.82 | 8.360 | 3934.97 |
|
||||
| 8192 | 32 | 8 | 65792 | 14.802 | 4427.63 | 1.222 | 209.44 | 16.024 | 4105.87 |
|
||||
| 8192 | 32 | 16 | 131584 | 29.596 | 4428.67 | 1.741 | 294.13 | 31.337 | 4199.00 |
|
||||
| 8192 | 32 | 32 | 263168 | 59.169 | 4430.42 | 2.619 | 390.92 | 61.789 | 4259.17 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 3714.25 ± 20.36 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 86.58 ± 0.43 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 3445.17 ± 17.85 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 81.72 ± 0.53 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 3218.78 ± 11.34 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 74.86 ± 0.64 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 2732.83 ± 7.17 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 71.57 ± 0.51 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 2119.75 ± 12.81 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 62.33 ± 0.24 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 4505.82 ± 12.90 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 83.43 ± 0.59 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 4158.34 ± 18.84 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 79.22 ± 0.60 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 3993.81 ± 17.55 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 75.22 ± 1.05 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 3449.98 ± 12.13 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 70.36 ± 0.37 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 2689.42 ± 18.89 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 61.65 ± 0.30 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/gpt-oss-120b-GGUF
|
||||
|
||||
@@ -77,46 +77,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.571 | 897.18 | 0.543 | 58.96 | 1.113 | 488.60 |
|
||||
| 512 | 32 | 2 | 1088 | 0.593 | 1725.37 | 1.041 | 61.45 | 1.635 | 665.48 |
|
||||
| 512 | 32 | 4 | 2176 | 1.043 | 1963.15 | 1.334 | 95.95 | 2.377 | 915.36 |
|
||||
| 512 | 32 | 8 | 4352 | 2.099 | 1951.63 | 1.717 | 149.07 | 3.816 | 1140.45 |
|
||||
| 512 | 32 | 16 | 8704 | 4.207 | 1947.12 | 2.311 | 221.56 | 6.518 | 1335.35 |
|
||||
| 512 | 32 | 32 | 17408 | 8.422 | 1945.36 | 3.298 | 310.46 | 11.720 | 1485.27 |
|
||||
| 4096 | 32 | 1 | 4128 | 2.138 | 1915.88 | 0.571 | 56.09 | 2.708 | 1524.12 |
|
||||
| 4096 | 32 | 2 | 8256 | 4.266 | 1920.25 | 1.137 | 56.27 | 5.404 | 1527.90 |
|
||||
| 4096 | 32 | 4 | 16512 | 8.564 | 1913.02 | 1.471 | 86.99 | 10.036 | 1645.29 |
|
||||
| 4096 | 32 | 8 | 33024 | 17.092 | 1917.19 | 1.979 | 129.33 | 19.071 | 1731.63 |
|
||||
| 4096 | 32 | 16 | 66048 | 34.211 | 1915.65 | 2.850 | 179.66 | 37.061 | 1782.15 |
|
||||
| 4096 | 32 | 32 | 132096 | 68.394 | 1916.44 | 4.381 | 233.72 | 72.775 | 1815.13 |
|
||||
| 8192 | 32 | 1 | 8224 | 4.349 | 1883.45 | 0.620 | 51.65 | 4.969 | 1655.04 |
|
||||
| 8192 | 32 | 2 | 16448 | 8.674 | 1888.83 | 1.178 | 54.33 | 9.852 | 1669.48 |
|
||||
| 8192 | 32 | 4 | 32896 | 17.351 | 1888.55 | 1.580 | 81.01 | 18.931 | 1737.68 |
|
||||
| 8192 | 32 | 8 | 65792 | 34.743 | 1886.31 | 2.173 | 117.80 | 36.916 | 1782.20 |
|
||||
| 8192 | 32 | 16 | 131584 | 69.413 | 1888.29 | 3.297 | 155.28 | 72.710 | 1809.70 |
|
||||
| 8192 | 32 | 32 | 263168 | 138.903 | 1887.24 | 5.004 | 204.63 | 143.907 | 1828.73 |
|
||||
| 512 | 32 | 1 | 544 | 0.445 | 1151.80 | 0.560 | 57.14 | 1.005 | 541.53 |
|
||||
| 512 | 32 | 2 | 1088 | 0.472 | 2169.85 | 0.874 | 73.27 | 1.345 | 808.65 |
|
||||
| 512 | 32 | 4 | 2176 | 0.826 | 2480.33 | 1.299 | 98.51 | 2.125 | 1023.94 |
|
||||
| 512 | 32 | 8 | 4352 | 1.644 | 2491.67 | 1.608 | 159.18 | 3.252 | 1338.20 |
|
||||
| 512 | 32 | 16 | 8704 | 3.292 | 2488.35 | 2.117 | 241.85 | 5.409 | 1609.13 |
|
||||
| 512 | 32 | 32 | 17408 | 6.604 | 2481.07 | 2.898 | 353.31 | 9.502 | 1832.04 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.698 | 2412.65 | 0.580 | 55.21 | 2.277 | 1812.66 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.399 | 2409.88 | 0.934 | 68.53 | 4.333 | 1905.27 |
|
||||
| 4096 | 32 | 4 | 16512 | 6.823 | 2401.21 | 1.411 | 90.72 | 8.234 | 2005.30 |
|
||||
| 4096 | 32 | 8 | 33024 | 13.574 | 2413.97 | 1.841 | 139.07 | 15.415 | 2142.31 |
|
||||
| 4096 | 32 | 16 | 66048 | 27.176 | 2411.52 | 2.609 | 196.26 | 29.785 | 2217.49 |
|
||||
| 4096 | 32 | 32 | 132096 | 54.359 | 2411.23 | 3.905 | 262.20 | 58.264 | 2267.19 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.491 | 2346.81 | 0.613 | 52.23 | 4.103 | 2004.21 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.939 | 2361.03 | 0.981 | 65.21 | 7.921 | 2076.56 |
|
||||
| 8192 | 32 | 4 | 32896 | 13.888 | 2359.40 | 1.511 | 84.71 | 15.399 | 2136.21 |
|
||||
| 8192 | 32 | 8 | 65792 | 27.756 | 2361.18 | 2.034 | 125.86 | 29.790 | 2208.56 |
|
||||
| 8192 | 32 | 16 | 131584 | 55.554 | 2359.34 | 3.021 | 169.49 | 58.575 | 2246.41 |
|
||||
| 8192 | 32 | 32 | 263168 | 111.036 | 2360.89 | 4.537 | 225.72 | 115.573 | 2277.08 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 1919.36 ± 5.01 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 60.40 ± 0.30 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 1825.30 ± 6.37 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 56.94 ± 0.29 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 1739.19 ± 6.00 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 52.51 ± 0.42 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 1536.75 ± 4.27 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 49.33 ± 0.27 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 1255.85 ± 3.26 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 42.99 ± 0.18 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 2443.91 ± 7.47 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 58.72 ± 0.20 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 2309.84 ± 3.63 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 55.67 ± 0.35 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 2216.68 ± 10.16 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 52.87 ± 0.43 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 1956.31 ± 6.39 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 49.45 ± 0.20 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 1567.08 ± 11.79 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 42.76 ± 0.14 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
|
||||
@@ -125,46 +125,46 @@ Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.398 | 1285.90 | 0.530 | 60.41 | 0.928 | 586.27 |
|
||||
| 512 | 32 | 2 | 1088 | 0.386 | 2651.65 | 0.948 | 67.50 | 1.334 | 815.38 |
|
||||
| 512 | 32 | 4 | 2176 | 0.666 | 3076.37 | 1.209 | 105.87 | 1.875 | 1160.71 |
|
||||
| 512 | 32 | 8 | 4352 | 1.325 | 3091.39 | 1.610 | 158.98 | 2.935 | 1482.65 |
|
||||
| 512 | 32 | 16 | 8704 | 2.664 | 3075.58 | 2.150 | 238.19 | 4.813 | 1808.39 |
|
||||
| 512 | 32 | 32 | 17408 | 5.336 | 3070.31 | 2.904 | 352.59 | 8.240 | 2112.50 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.444 | 2836.81 | 0.581 | 55.09 | 2.025 | 2038.81 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.872 | 2852.14 | 1.084 | 59.06 | 3.956 | 2086.99 |
|
||||
| 4096 | 32 | 4 | 16512 | 5.744 | 2852.32 | 1.440 | 88.90 | 7.184 | 2298.47 |
|
||||
| 4096 | 32 | 8 | 33024 | 11.463 | 2858.68 | 2.068 | 123.78 | 13.531 | 2440.65 |
|
||||
| 4096 | 32 | 16 | 66048 | 22.915 | 2859.95 | 3.018 | 169.67 | 25.933 | 2546.90 |
|
||||
| 4096 | 32 | 32 | 132096 | 45.956 | 2852.10 | 4.609 | 222.18 | 50.565 | 2612.39 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.063 | 2674.72 | 0.693 | 46.20 | 3.755 | 2189.92 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.109 | 2681.87 | 1.214 | 52.71 | 7.323 | 2245.98 |
|
||||
| 8192 | 32 | 4 | 32896 | 12.197 | 2686.63 | 1.682 | 76.11 | 13.878 | 2370.30 |
|
||||
| 8192 | 32 | 8 | 65792 | 24.409 | 2684.94 | 2.556 | 100.17 | 26.965 | 2439.95 |
|
||||
| 8192 | 32 | 16 | 131584 | 48.753 | 2688.50 | 3.994 | 128.20 | 52.747 | 2494.64 |
|
||||
| 8192 | 32 | 32 | 263168 | 97.508 | 2688.42 | 6.528 | 156.86 | 104.037 | 2529.57 |
|
||||
| 512 | 32 | 1 | 544 | 0.393 | 1303.73 | 0.548 | 58.36 | 0.941 | 578.10 |
|
||||
| 512 | 32 | 2 | 1088 | 0.387 | 2648.68 | 0.910 | 70.35 | 1.296 | 839.27 |
|
||||
| 512 | 32 | 4 | 2176 | 0.659 | 3107.63 | 1.302 | 98.33 | 1.961 | 1109.77 |
|
||||
| 512 | 32 | 8 | 4352 | 1.322 | 3099.35 | 1.669 | 153.42 | 2.990 | 1455.43 |
|
||||
| 512 | 32 | 16 | 8704 | 2.639 | 3104.63 | 2.212 | 231.44 | 4.851 | 1794.32 |
|
||||
| 512 | 32 | 32 | 17408 | 5.284 | 3100.80 | 2.955 | 346.53 | 8.239 | 2112.93 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.417 | 2890.36 | 0.598 | 53.51 | 2.015 | 2048.45 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.829 | 2895.62 | 1.019 | 62.82 | 3.848 | 2145.60 |
|
||||
| 4096 | 32 | 4 | 16512 | 5.656 | 2896.96 | 1.528 | 83.79 | 7.183 | 2298.71 |
|
||||
| 4096 | 32 | 8 | 33024 | 11.338 | 2890.02 | 2.127 | 120.36 | 13.465 | 2452.53 |
|
||||
| 4096 | 32 | 16 | 66048 | 22.709 | 2885.96 | 3.104 | 164.97 | 25.812 | 2558.79 |
|
||||
| 4096 | 32 | 32 | 132096 | 45.301 | 2893.35 | 4.723 | 216.80 | 50.024 | 2640.63 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.022 | 2711.09 | 0.678 | 47.20 | 3.700 | 2222.89 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.039 | 2713.01 | 1.149 | 55.70 | 7.188 | 2288.21 |
|
||||
| 8192 | 32 | 4 | 32896 | 12.050 | 2719.35 | 1.785 | 71.69 | 13.835 | 2377.67 |
|
||||
| 8192 | 32 | 8 | 65792 | 24.113 | 2717.90 | 2.629 | 97.39 | 26.741 | 2460.31 |
|
||||
| 8192 | 32 | 16 | 131584 | 48.178 | 2720.58 | 4.099 | 124.91 | 52.277 | 2517.06 |
|
||||
| 8192 | 32 | 32 | 263168 | 96.401 | 2719.31 | 6.696 | 152.93 | 103.097 | 2552.63 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 2925.55 ± 4.25 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 62.80 ± 0.27 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 2531.01 ± 6.79 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 55.86 ± 0.33 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 2244.39 ± 5.33 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 45.95 ± 0.33 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 1783.17 ± 3.68 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 39.07 ± 0.10 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 1241.90 ± 3.13 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 29.92 ± 0.06 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 2986.97 ± 18.87 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 61.06 ± 0.23 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 2633.45 ± 6.26 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 54.77 ± 0.28 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 2354.14 ± 3.84 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 48.02 ± 0.40 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 1908.86 ± 4.25 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 40.23 ± 0.10 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 1348.17 ± 2.00 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 30.21 ± 0.04 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
|
||||
@@ -173,46 +173,46 @@ Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.211 | 2421.57 | 1.055 | 30.33 | 1.266 | 429.57 |
|
||||
| 512 | 32 | 2 | 1088 | 0.419 | 2441.34 | 1.130 | 56.65 | 1.549 | 702.32 |
|
||||
| 512 | 32 | 4 | 2176 | 0.873 | 2345.54 | 1.174 | 108.99 | 2.048 | 1062.74 |
|
||||
| 512 | 32 | 8 | 4352 | 1.727 | 2371.85 | 1.254 | 204.22 | 2.980 | 1460.19 |
|
||||
| 512 | 32 | 16 | 8704 | 3.452 | 2373.22 | 1.492 | 343.16 | 4.944 | 1760.56 |
|
||||
| 512 | 32 | 32 | 17408 | 6.916 | 2368.93 | 1.675 | 611.51 | 8.591 | 2026.36 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.799 | 2277.26 | 1.084 | 29.51 | 2.883 | 1431.91 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.577 | 2290.01 | 1.196 | 53.50 | 4.774 | 1729.51 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.172 | 2284.36 | 1.313 | 97.50 | 8.485 | 1946.00 |
|
||||
| 4096 | 32 | 8 | 33024 | 14.341 | 2284.96 | 1.520 | 168.46 | 15.860 | 2082.18 |
|
||||
| 4096 | 32 | 16 | 66048 | 28.675 | 2285.44 | 1.983 | 258.21 | 30.658 | 2154.33 |
|
||||
| 4096 | 32 | 32 | 132096 | 57.354 | 2285.32 | 2.640 | 387.87 | 59.994 | 2201.82 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.701 | 2213.75 | 1.119 | 28.59 | 4.820 | 1706.34 |
|
||||
| 8192 | 32 | 2 | 16448 | 7.410 | 2211.19 | 1.272 | 50.31 | 8.682 | 1894.56 |
|
||||
| 8192 | 32 | 4 | 32896 | 14.802 | 2213.83 | 1.460 | 87.68 | 16.261 | 2022.96 |
|
||||
| 8192 | 32 | 8 | 65792 | 29.609 | 2213.35 | 1.781 | 143.74 | 31.390 | 2095.93 |
|
||||
| 8192 | 32 | 16 | 131584 | 59.229 | 2212.96 | 2.495 | 205.17 | 61.725 | 2131.79 |
|
||||
| 8192 | 32 | 32 | 263168 | 118.449 | 2213.15 | 3.714 | 275.75 | 122.162 | 2154.25 |
|
||||
| 512 | 32 | 1 | 544 | 0.212 | 2420.12 | 1.100 | 29.10 | 1.311 | 414.85 |
|
||||
| 512 | 32 | 2 | 1088 | 0.428 | 2393.89 | 1.185 | 54.00 | 1.613 | 674.56 |
|
||||
| 512 | 32 | 4 | 2176 | 0.894 | 2290.41 | 1.229 | 104.17 | 2.123 | 1025.02 |
|
||||
| 512 | 32 | 8 | 4352 | 1.758 | 2330.36 | 1.319 | 194.15 | 3.076 | 1414.70 |
|
||||
| 512 | 32 | 16 | 8704 | 3.508 | 2335.21 | 1.543 | 331.90 | 5.051 | 1723.33 |
|
||||
| 512 | 32 | 32 | 17408 | 7.035 | 2328.93 | 1.738 | 589.21 | 8.773 | 1984.29 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.831 | 2237.25 | 1.125 | 28.44 | 2.956 | 1396.42 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.642 | 2249.48 | 1.253 | 51.07 | 4.895 | 1686.64 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.274 | 2252.26 | 1.380 | 92.72 | 8.655 | 1907.81 |
|
||||
| 4096 | 32 | 8 | 33024 | 14.576 | 2248.09 | 1.617 | 158.29 | 16.193 | 2039.37 |
|
||||
| 4096 | 32 | 16 | 66048 | 29.138 | 2249.17 | 2.081 | 246.01 | 31.219 | 2115.63 |
|
||||
| 4096 | 32 | 32 | 132096 | 58.275 | 2249.19 | 2.814 | 363.87 | 61.089 | 2162.34 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.757 | 2180.26 | 1.184 | 27.03 | 4.941 | 1664.37 |
|
||||
| 8192 | 32 | 2 | 16448 | 7.522 | 2178.05 | 1.341 | 47.73 | 8.863 | 1855.77 |
|
||||
| 8192 | 32 | 4 | 32896 | 15.043 | 2178.25 | 1.548 | 82.69 | 16.591 | 1982.74 |
|
||||
| 8192 | 32 | 8 | 65792 | 30.111 | 2176.49 | 1.937 | 132.13 | 32.048 | 2052.90 |
|
||||
| 8192 | 32 | 16 | 131584 | 60.405 | 2169.90 | 2.706 | 189.21 | 63.111 | 2084.97 |
|
||||
| 8192 | 32 | 32 | 263168 | 120.439 | 2176.58 | 3.993 | 256.46 | 124.432 | 2114.96 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 2272.74 ± 4.68 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 30.66 ± 0.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 2107.80 ± 9.55 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 29.71 ± 0.05 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 1937.80 ± 6.75 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 28.86 ± 0.04 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 1641.12 ± 1.78 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 27.24 ± 0.04 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 1296.02 ± 2.67 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 23.78 ± 0.03 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 2250.28 ± 6.41 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 29.43 ± 0.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 2100.19 ± 8.96 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 28.61 ± 0.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 2007.56 ± 4.16 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 27.38 ± 0.09 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 1779.11 ± 6.42 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 25.72 ± 0.03 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 1471.23 ± 1.71 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 22.51 ± 0.02 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
|
||||
@@ -221,44 +221,91 @@ Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.094 | 5434.73 | 0.394 | 81.21 | 0.488 | 1114.15 |
|
||||
| 512 | 32 | 2 | 1088 | 0.168 | 6091.68 | 0.498 | 128.52 | 0.666 | 1633.41 |
|
||||
| 512 | 32 | 4 | 2176 | 0.341 | 6010.68 | 0.542 | 236.37 | 0.882 | 2466.43 |
|
||||
| 512 | 32 | 8 | 4352 | 0.665 | 6161.46 | 0.678 | 377.74 | 1.342 | 3241.72 |
|
||||
| 512 | 32 | 16 | 8704 | 1.323 | 6193.19 | 0.902 | 567.41 | 2.225 | 3911.74 |
|
||||
| 512 | 32 | 32 | 17408 | 2.642 | 6202.03 | 1.231 | 832.03 | 3.872 | 4495.36 |
|
||||
| 4096 | 32 | 1 | 4128 | 0.701 | 5840.49 | 0.439 | 72.95 | 1.140 | 3621.23 |
|
||||
| 4096 | 32 | 2 | 8256 | 1.387 | 5906.82 | 0.574 | 111.48 | 1.961 | 4210.12 |
|
||||
| 4096 | 32 | 4 | 16512 | 2.758 | 5940.33 | 0.651 | 196.58 | 3.409 | 4843.33 |
|
||||
| 4096 | 32 | 8 | 33024 | 5.491 | 5967.56 | 0.876 | 292.40 | 6.367 | 5187.12 |
|
||||
| 4096 | 32 | 16 | 66048 | 10.978 | 5969.58 | 1.275 | 401.69 | 12.253 | 5390.38 |
|
||||
| 4096 | 32 | 32 | 132096 | 21.944 | 5972.93 | 1.992 | 514.16 | 23.936 | 5518.73 |
|
||||
| 8192 | 32 | 1 | 8224 | 1.402 | 5841.91 | 0.452 | 70.73 | 1.855 | 4434.12 |
|
||||
| 8192 | 32 | 2 | 16448 | 2.793 | 5865.34 | 0.637 | 100.55 | 3.430 | 4795.51 |
|
||||
| 8192 | 32 | 4 | 32896 | 5.564 | 5889.64 | 0.770 | 166.26 | 6.334 | 5193.95 |
|
||||
| 8192 | 32 | 8 | 65792 | 11.114 | 5896.44 | 1.122 | 228.07 | 12.237 | 5376.51 |
|
||||
| 8192 | 32 | 16 | 131584 | 22.210 | 5901.38 | 1.789 | 286.15 | 24.000 | 5482.74 |
|
||||
| 8192 | 32 | 32 | 263168 | 44.382 | 5906.56 | 3.044 | 336.38 | 47.426 | 5549.02 |
|
||||
| 512 | 32 | 1 | 544 | 0.092 | 5566.97 | 0.412 | 77.63 | 0.504 | 1078.95 |
|
||||
| 512 | 32 | 2 | 1088 | 0.161 | 6345.67 | 0.522 | 122.70 | 0.683 | 1593.06 |
|
||||
| 512 | 32 | 4 | 2176 | 0.325 | 6309.87 | 0.562 | 227.68 | 0.887 | 2453.87 |
|
||||
| 512 | 32 | 8 | 4352 | 0.643 | 6374.42 | 0.685 | 373.67 | 1.328 | 3277.94 |
|
||||
| 512 | 32 | 16 | 8704 | 1.277 | 6413.64 | 0.915 | 559.47 | 2.192 | 3970.01 |
|
||||
| 512 | 32 | 32 | 17408 | 2.518 | 6506.57 | 1.249 | 819.61 | 3.767 | 4620.64 |
|
||||
| 4096 | 32 | 1 | 4128 | 0.674 | 6079.68 | 0.453 | 70.60 | 1.127 | 3662.88 |
|
||||
| 4096 | 32 | 2 | 8256 | 1.335 | 6137.82 | 0.627 | 102.03 | 1.962 | 4208.11 |
|
||||
| 4096 | 32 | 4 | 16512 | 2.657 | 6167.35 | 0.749 | 170.92 | 3.405 | 4848.71 |
|
||||
| 4096 | 32 | 8 | 33024 | 5.307 | 6173.91 | 0.974 | 262.89 | 6.281 | 5257.53 |
|
||||
| 4096 | 32 | 16 | 66048 | 10.610 | 6176.96 | 1.379 | 371.42 | 11.988 | 5509.40 |
|
||||
| 4096 | 32 | 32 | 132096 | 21.213 | 6178.89 | 2.122 | 482.50 | 23.335 | 5660.82 |
|
||||
| 8192 | 32 | 1 | 8224 | 1.359 | 6027.34 | 0.467 | 68.52 | 1.826 | 4503.48 |
|
||||
| 8192 | 32 | 2 | 16448 | 2.699 | 6069.68 | 0.653 | 98.03 | 3.352 | 4906.68 |
|
||||
| 8192 | 32 | 4 | 32896 | 5.366 | 6106.74 | 0.818 | 156.55 | 6.184 | 5319.96 |
|
||||
| 8192 | 32 | 8 | 65792 | 10.755 | 6093.50 | 1.174 | 218.04 | 11.929 | 5515.22 |
|
||||
| 8192 | 32 | 16 | 131584 | 21.484 | 6100.82 | 1.829 | 279.90 | 23.314 | 5644.11 |
|
||||
| 8192 | 32 | 32 | 263168 | 42.950 | 6103.40 | 3.058 | 334.91 | 46.008 | 5720.05 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 5810.04 ± 21.71 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 84.54 ± 0.18 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 5288.04 ± 3.54 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 78.82 ± 1.37 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 4960.43 ± 16.64 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 74.13 ± 0.30 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 4495.92 ± 31.11 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 72.37 ± 0.29 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 3746.90 ± 40.01 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 63.02 ± 0.20 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 5948.74 ± 10.61 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 81.05 ± 0.20 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 5652.69 ± 34.29 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 76.37 ± 0.58 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 5509.57 ± 40.69 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 71.61 ± 0.80 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 5340.86 ± 36.92 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 70.89 ± 0.34 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 5023.30 ± 13.52 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 62.28 ± 0.30 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.433 | 1181.83 | 0.693 | 46.16 | 1.126 | 482.94 |
|
||||
| 512 | 32 | 2 | 1088 | 0.439 | 2334.46 | 1.034 | 61.89 | 1.473 | 738.75 |
|
||||
| 512 | 32 | 4 | 2176 | 0.772 | 2654.46 | 1.459 | 87.76 | 2.230 | 975.77 |
|
||||
| 512 | 32 | 8 | 4352 | 1.541 | 2658.78 | 2.043 | 125.31 | 3.583 | 1214.47 |
|
||||
| 512 | 32 | 16 | 8704 | 3.083 | 2656.91 | 2.675 | 191.42 | 5.758 | 1511.62 |
|
||||
| 512 | 32 | 32 | 17408 | 6.159 | 2660.12 | 3.615 | 283.24 | 9.774 | 1780.98 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.915 | 2139.30 | 0.725 | 44.14 | 2.640 | 1563.83 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.834 | 2136.40 | 1.119 | 57.21 | 4.953 | 1666.81 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.636 | 2145.72 | 1.631 | 78.49 | 9.266 | 1781.93 |
|
||||
| 4096 | 32 | 8 | 33024 | 15.295 | 2142.40 | 2.344 | 109.21 | 17.639 | 1872.20 |
|
||||
| 4096 | 32 | 16 | 66048 | 30.573 | 2143.62 | 3.773 | 135.70 | 34.346 | 1923.04 |
|
||||
| 4096 | 32 | 32 | 132096 | 61.282 | 2138.82 | 5.795 | 176.71 | 67.077 | 1969.31 |
|
||||
| 8192 | 32 | 1 | 8224 | 4.510 | 1816.24 | 0.760 | 42.11 | 5.270 | 1560.44 |
|
||||
| 8192 | 32 | 2 | 16448 | 9.036 | 1813.19 | 1.206 | 53.06 | 10.242 | 1605.91 |
|
||||
| 8192 | 32 | 4 | 32896 | 18.070 | 1813.43 | 1.783 | 71.80 | 19.852 | 1657.03 |
|
||||
| 8192 | 32 | 8 | 65792 | 36.125 | 1814.15 | 2.635 | 97.14 | 38.760 | 1697.41 |
|
||||
| 8192 | 32 | 16 | 131584 | 72.367 | 1811.20 | 4.954 | 103.34 | 77.322 | 1701.77 |
|
||||
| 8192 | 32 | 32 | 263168 | 144.501 | 1814.13 | 8.103 | 126.37 | 152.604 | 1724.51 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | --------------: | -------------------: |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 | 2364.18 ± 11.43 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 | 48.68 ± 0.12 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d4096 | 1684.13 ± 1.24 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d4096 | 44.62 ± 0.22 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d8192 | 1314.68 ± 1.41 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d8192 | 42.59 ± 0.11 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d16384 | 914.05 ± 3.32 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d16384 | 38.72 ± 0.13 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d32768 | 567.20 ± 0.90 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d32768 | 32.65 ± 0.09 |
|
||||
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
298
benches/mac-m2-ultra/mac-m2-ultra.md
Normal file
298
benches/mac-m2-ultra/mac-m2-ultra.md
Normal file
@@ -0,0 +1,298 @@
|
||||
## System info
|
||||
|
||||
```bash
|
||||
uname -a
|
||||
Darwin gg-studio 25.2.0 Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 arm64
|
||||
|
||||
g++ --version
|
||||
Apple clang version 17.0.0 (clang-1700.3.19.1)
|
||||
Target: arm64-apple-darwin25.2.0
|
||||
```
|
||||
|
||||
## ggml-org/gpt-oss-20b-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.215 | 2381.35 | 0.245 | 130.45 | 0.460 | 1181.81 |
|
||||
| 512 | 32 | 2 | 1088 | 0.379 | 2701.43 | 0.382 | 167.56 | 0.761 | 1429.67 |
|
||||
| 512 | 32 | 4 | 2176 | 0.721 | 2839.27 | 0.604 | 211.76 | 1.326 | 1641.32 |
|
||||
| 512 | 32 | 8 | 4352 | 1.433 | 2858.30 | 1.033 | 247.75 | 2.466 | 1764.57 |
|
||||
| 512 | 32 | 16 | 8704 | 2.853 | 2871.12 | 1.570 | 326.11 | 4.423 | 1967.77 |
|
||||
| 512 | 32 | 32 | 17408 | 5.699 | 2874.95 | 1.910 | 536.15 | 7.609 | 2287.88 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.552 | 2638.56 | 0.334 | 95.72 | 1.887 | 2188.00 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.084 | 2655.88 | 0.404 | 158.54 | 3.488 | 2366.86 |
|
||||
| 4096 | 32 | 4 | 16512 | 6.151 | 2663.78 | 0.652 | 196.39 | 6.802 | 2427.37 |
|
||||
| 4096 | 32 | 8 | 33024 | 12.288 | 2666.77 | 1.135 | 225.47 | 13.423 | 2460.27 |
|
||||
| 4096 | 32 | 16 | 66048 | 24.563 | 2668.12 | 1.762 | 290.55 | 26.325 | 2508.97 |
|
||||
| 4096 | 32 | 32 | 132096 | 49.114 | 2668.73 | 2.398 | 426.94 | 51.512 | 2564.35 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.345 | 2448.78 | 0.275 | 116.46 | 3.620 | 2271.76 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.665 | 2458.11 | 0.425 | 150.71 | 7.090 | 2319.91 |
|
||||
| 8192 | 32 | 4 | 32896 | 13.315 | 2460.92 | 0.691 | 185.21 | 14.006 | 2348.63 |
|
||||
| 8192 | 32 | 8 | 65792 | 26.611 | 2462.73 | 1.212 | 211.16 | 27.823 | 2364.62 |
|
||||
| 8192 | 32 | 16 | 131584 | 53.232 | 2462.27 | 1.919 | 266.83 | 55.151 | 2385.88 |
|
||||
| 8192 | 32 | 32 | 263168 | 110.455 | 2373.30 | 2.752 | 372.03 | 113.208 | 2324.64 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 2713.40 ± 3.56 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 129.97 ± 3.90 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 2324.59 ± 3.01 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 123.38 ± 0.17 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1989.82 ± 30.11 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 117.39 ± 0.33 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 1556.54 ± 6.22 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 109.75 ± 0.42 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 1122.63 ± 1.45 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 98.25 ± 0.08 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/gpt-oss-120b-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.426 | 1200.92 | 0.361 | 88.56 | 0.788 | 690.64 |
|
||||
| 512 | 32 | 2 | 1088 | 0.683 | 1500.14 | 0.545 | 117.35 | 1.228 | 886.02 |
|
||||
| 512 | 32 | 4 | 2176 | 1.204 | 1701.56 | 0.847 | 151.19 | 2.050 | 1061.34 |
|
||||
| 512 | 32 | 8 | 4352 | 2.402 | 1705.20 | 1.455 | 176.00 | 3.857 | 1128.45 |
|
||||
| 512 | 32 | 16 | 8704 | 4.802 | 1705.90 | 2.349 | 217.93 | 7.152 | 1217.08 |
|
||||
| 512 | 32 | 32 | 17408 | 9.593 | 1707.85 | 3.665 | 279.42 | 13.258 | 1313.01 |
|
||||
| 4096 | 32 | 1 | 4128 | 2.581 | 1587.08 | 0.390 | 82.12 | 2.970 | 1389.67 |
|
||||
| 4096 | 32 | 2 | 8256 | 5.124 | 1598.79 | 0.589 | 108.62 | 5.713 | 1445.10 |
|
||||
| 4096 | 32 | 4 | 16512 | 10.231 | 1601.47 | 0.928 | 137.98 | 11.158 | 1479.80 |
|
||||
| 4096 | 32 | 8 | 33024 | 20.468 | 1600.94 | 1.606 | 159.38 | 22.074 | 1496.04 |
|
||||
| 4096 | 32 | 16 | 66048 | 40.924 | 1601.42 | 2.639 | 193.99 | 43.563 | 1516.15 |
|
||||
| 4096 | 32 | 32 | 132096 | 81.819 | 1601.98 | 4.466 | 229.29 | 86.284 | 1530.94 |
|
||||
| 8192 | 32 | 1 | 8224 | 5.517 | 1484.74 | 0.409 | 78.16 | 5.927 | 1387.58 |
|
||||
| 8192 | 32 | 2 | 16448 | 11.008 | 1488.43 | 0.622 | 102.92 | 11.629 | 1414.34 |
|
||||
| 8192 | 32 | 4 | 32896 | 22.002 | 1489.29 | 0.987 | 129.66 | 22.990 | 1430.90 |
|
||||
| 8192 | 32 | 8 | 65792 | 46.051 | 1423.11 | 1.858 | 137.79 | 47.909 | 1373.27 |
|
||||
| 8192 | 32 | 16 | 131584 | 97.680 | 1341.85 | 2.872 | 178.28 | 100.552 | 1308.62 |
|
||||
| 8192 | 32 | 32 | 263168 | 176.407 | 1486.02 | 5.048 | 202.85 | 181.455 | 1450.32 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 1648.69 ± 1.80 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 85.60 ± 0.52 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 1429.86 ± 1.01 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 82.03 ± 0.12 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1257.90 ± 1.81 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 78.23 ± 0.33 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 1013.49 ± 0.70 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 73.20 ± 0.28 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 721.11 ± 0.58 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 65.52 ± 0.10 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.243 | 2109.23 | 0.419 | 76.34 | 0.662 | 821.84 |
|
||||
| 512 | 32 | 2 | 1088 | 0.406 | 2521.40 | 0.575 | 111.36 | 0.981 | 1109.27 |
|
||||
| 512 | 32 | 4 | 2176 | 0.744 | 2751.65 | 0.841 | 152.22 | 1.585 | 1372.71 |
|
||||
| 512 | 32 | 8 | 4352 | 1.479 | 2770.20 | 1.330 | 192.48 | 2.809 | 1549.53 |
|
||||
| 512 | 32 | 16 | 8704 | 2.951 | 2776.20 | 2.572 | 199.05 | 5.523 | 1575.93 |
|
||||
| 512 | 32 | 32 | 17408 | 5.899 | 2777.64 | 2.603 | 393.34 | 8.502 | 2047.54 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.901 | 2154.15 | 0.474 | 67.58 | 2.375 | 1738.14 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.788 | 2162.89 | 0.652 | 98.17 | 4.439 | 1859.69 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.564 | 2166.18 | 0.990 | 129.24 | 8.554 | 1930.34 |
|
||||
| 4096 | 32 | 8 | 33024 | 15.121 | 2166.98 | 1.632 | 156.82 | 16.754 | 1971.12 |
|
||||
| 4096 | 32 | 16 | 66048 | 30.241 | 2167.09 | 3.166 | 161.72 | 33.407 | 1977.04 |
|
||||
| 4096 | 32 | 32 | 132096 | 60.474 | 2167.42 | 3.780 | 270.93 | 64.254 | 2055.86 |
|
||||
| 8192 | 32 | 1 | 8224 | 4.733 | 1730.92 | 0.483 | 66.29 | 5.215 | 1576.85 |
|
||||
| 8192 | 32 | 2 | 16448 | 9.459 | 1732.09 | 0.722 | 88.58 | 10.182 | 1615.46 |
|
||||
| 8192 | 32 | 4 | 32896 | 18.912 | 1732.65 | 1.120 | 114.26 | 20.032 | 1642.14 |
|
||||
| 8192 | 32 | 8 | 65792 | 37.797 | 1733.91 | 1.873 | 136.67 | 39.670 | 1658.49 |
|
||||
| 8192 | 32 | 16 | 131584 | 84.133 | 1557.92 | 3.718 | 137.72 | 87.850 | 1497.82 |
|
||||
| 8192 | 32 | 32 | 263168 | 157.550 | 1663.88 | 4.854 | 210.98 | 162.403 | 1620.46 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 2453.11 ± 1.70 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 78.97 ± 0.46 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 1569.46 ± 1.97 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 71.18 ± 0.37 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1145.51 ± 1.16 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 65.11 ± 0.36 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 741.04 ± 0.74 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 56.87 ± 0.14 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 431.31 ± 0.31 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 45.26 ± 0.11 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.339 | 1509.22 | 0.409 | 78.17 | 0.749 | 726.67 |
|
||||
| 512 | 32 | 2 | 1088 | 0.646 | 1584.93 | 0.483 | 132.45 | 1.129 | 963.45 |
|
||||
| 512 | 32 | 4 | 2176 | 1.258 | 1627.50 | 0.585 | 218.67 | 1.844 | 1180.21 |
|
||||
| 512 | 32 | 8 | 4352 | 2.506 | 1634.41 | 1.005 | 254.83 | 3.511 | 1239.64 |
|
||||
| 512 | 32 | 16 | 8704 | 5.007 | 1635.99 | 1.595 | 321.07 | 6.602 | 1318.38 |
|
||||
| 512 | 32 | 32 | 17408 | 10.007 | 1637.19 | 1.676 | 611.12 | 11.683 | 1490.03 |
|
||||
| 4096 | 32 | 1 | 4128 | 2.730 | 1500.46 | 0.431 | 74.31 | 3.160 | 1306.12 |
|
||||
| 4096 | 32 | 2 | 8256 | 5.446 | 1504.33 | 0.524 | 122.04 | 5.970 | 1382.91 |
|
||||
| 4096 | 32 | 4 | 16512 | 10.875 | 1506.59 | 0.662 | 193.45 | 11.537 | 1431.28 |
|
||||
| 4096 | 32 | 8 | 33024 | 21.749 | 1506.61 | 1.158 | 221.11 | 22.907 | 1441.64 |
|
||||
| 4096 | 32 | 16 | 66048 | 43.477 | 1507.36 | 1.901 | 269.32 | 45.378 | 1455.49 |
|
||||
| 4096 | 32 | 32 | 132096 | 86.954 | 1507.37 | 2.325 | 440.42 | 89.279 | 1479.59 |
|
||||
| 8192 | 32 | 1 | 8224 | 5.940 | 1379.21 | 0.449 | 71.20 | 6.389 | 1287.20 |
|
||||
| 8192 | 32 | 2 | 16448 | 11.865 | 1380.84 | 0.559 | 114.59 | 12.424 | 1323.92 |
|
||||
| 8192 | 32 | 4 | 32896 | 23.723 | 1381.25 | 0.728 | 175.80 | 24.452 | 1345.35 |
|
||||
| 8192 | 32 | 8 | 65792 | 47.434 | 1381.63 | 1.279 | 200.09 | 48.713 | 1350.60 |
|
||||
| 8192 | 32 | 16 | 131584 | 94.864 | 1381.69 | 2.198 | 232.97 | 97.061 | 1355.68 |
|
||||
| 8192 | 32 | 32 | 263168 | 189.743 | 1381.57 | 3.052 | 335.50 | 192.795 | 1365.01 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 1565.91 ± 0.86 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 79.68 ± 0.39 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 1317.41 ± 1.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 74.70 ± 0.04 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1134.65 ± 0.76 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 71.31 ± 0.12 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 886.46 ± 0.78 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 65.93 ± 0.06 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 612.21 ± 0.30 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 56.83 ± 0.02 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.186 | 2748.06 | 0.235 | 136.28 | 0.421 | 1291.78 |
|
||||
| 512 | 32 | 2 | 1088 | 0.342 | 2990.95 | 0.312 | 204.99 | 0.655 | 1662.15 |
|
||||
| 512 | 32 | 4 | 2176 | 0.662 | 3092.69 | 0.404 | 316.97 | 1.066 | 2041.21 |
|
||||
| 512 | 32 | 8 | 4352 | 1.317 | 3110.41 | 0.579 | 441.80 | 1.896 | 2294.97 |
|
||||
| 512 | 32 | 16 | 8704 | 2.625 | 3120.23 | 1.207 | 424.08 | 3.833 | 2270.93 |
|
||||
| 512 | 32 | 32 | 17408 | 5.242 | 3125.34 | 1.299 | 788.23 | 6.541 | 2661.19 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.408 | 2909.90 | 0.296 | 108.07 | 1.704 | 2422.95 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.793 | 2933.40 | 0.325 | 197.00 | 3.118 | 2648.25 |
|
||||
| 4096 | 32 | 4 | 16512 | 5.567 | 2943.22 | 0.440 | 291.07 | 6.006 | 2749.05 |
|
||||
| 4096 | 32 | 8 | 33024 | 11.114 | 2948.23 | 0.640 | 400.26 | 11.754 | 2809.59 |
|
||||
| 4096 | 32 | 16 | 66048 | 22.217 | 2949.76 | 1.327 | 385.83 | 23.544 | 2805.26 |
|
||||
| 4096 | 32 | 32 | 132096 | 44.420 | 2950.77 | 1.553 | 659.30 | 45.973 | 2873.36 |
|
||||
| 8192 | 32 | 1 | 8224 | 2.860 | 2864.58 | 0.250 | 127.90 | 3.110 | 2644.42 |
|
||||
| 8192 | 32 | 2 | 16448 | 5.702 | 2873.63 | 0.335 | 191.07 | 6.036 | 2724.77 |
|
||||
| 8192 | 32 | 4 | 32896 | 11.383 | 2878.69 | 0.456 | 280.72 | 11.839 | 2778.63 |
|
||||
| 8192 | 32 | 8 | 65792 | 22.750 | 2880.75 | 0.671 | 381.48 | 23.421 | 2809.14 |
|
||||
| 8192 | 32 | 16 | 131584 | 45.484 | 2881.74 | 1.406 | 364.04 | 46.890 | 2806.22 |
|
||||
| 8192 | 32 | 32 | 263168 | 90.956 | 2882.10 | 1.793 | 570.98 | 92.749 | 2837.41 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 2923.59 ± 3.10 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 134.28 ± 1.29 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 2748.21 ± 3.05 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 133.11 ± 0.08 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 2641.45 ± 2.31 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 125.85 ± 0.35 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 2446.20 ± 2.94 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 125.00 ± 0.12 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 2129.18 ± 7.43 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 113.14 ± 0.10 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.326 | 1568.69 | 0.522 | 61.28 | 0.849 | 641.09 |
|
||||
| 512 | 32 | 2 | 1088 | 0.528 | 1939.42 | 0.744 | 86.07 | 1.272 | 855.63 |
|
||||
| 512 | 32 | 4 | 2176 | 0.968 | 2114.85 | 1.105 | 115.85 | 2.073 | 1049.56 |
|
||||
| 512 | 32 | 8 | 4352 | 1.928 | 2124.62 | 1.684 | 151.99 | 3.612 | 1204.82 |
|
||||
| 512 | 32 | 16 | 8704 | 3.844 | 2131.34 | 3.141 | 162.99 | 6.985 | 1246.11 |
|
||||
| 512 | 32 | 32 | 17408 | 7.683 | 2132.38 | 3.924 | 260.95 | 11.608 | 1499.71 |
|
||||
| 4096 | 32 | 1 | 4128 | 3.280 | 1248.75 | 0.723 | 44.29 | 4.003 | 1031.33 |
|
||||
| 4096 | 32 | 2 | 8256 | 6.545 | 1251.63 | 0.930 | 68.85 | 7.475 | 1104.53 |
|
||||
| 4096 | 32 | 4 | 16512 | 13.080 | 1252.64 | 1.454 | 88.03 | 14.534 | 1136.12 |
|
||||
| 4096 | 32 | 8 | 33024 | 26.154 | 1252.90 | 2.388 | 107.20 | 28.542 | 1157.04 |
|
||||
| 4096 | 32 | 16 | 66048 | 52.297 | 1253.14 | 4.724 | 108.37 | 57.022 | 1158.30 |
|
||||
| 4096 | 32 | 32 | 132096 | 104.578 | 1253.34 | 7.266 | 140.93 | 111.844 | 1181.08 |
|
||||
| 8192 | 32 | 1 | 8224 | 9.623 | 851.31 | 0.767 | 41.72 | 10.390 | 791.54 |
|
||||
| 8192 | 32 | 2 | 16448 | 20.916 | 783.32 | 1.148 | 55.74 | 22.064 | 745.45 |
|
||||
| 8192 | 32 | 4 | 32896 | 43.509 | 753.14 | 1.833 | 69.82 | 45.342 | 725.51 |
|
||||
| 8192 | 32 | 8 | 65792 | 79.621 | 823.10 | 3.180 | 80.50 | 82.801 | 794.58 |
|
||||
| 8192 | 32 | 16 | 131584 | 153.770 | 852.39 | 6.502 | 78.74 | 160.272 | 821.00 |
|
||||
| 8192 | 32 | 32 | 263168 | 307.539 | 852.39 | 10.839 | 94.48 | 318.378 | 826.59 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 1629.33 ± 0.27 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 59.58 ± 0.13 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 732.67 ± 0.42 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 47.44 ± 0.15 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 474.33 ± 0.33 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 40.20 ± 0.20 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 277.46 ± 0.09 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 31.50 ± 0.93 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 151.44 ± 0.05 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 21.81 ± 0.01 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
@@ -3437,16 +3437,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
|
||||
params.speculative.ngram_size_m = value;
|
||||
}
|
||||
).set_examples({LLAMA_EXAMPLE_SERVER}));
|
||||
add_opt(common_arg(
|
||||
{"--spec-ngram-check-rate"}, "N",
|
||||
string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
|
||||
[](common_params & params, int value) {
|
||||
if (value < 1) {
|
||||
throw std::invalid_argument("ngram check rate must be at least 1");
|
||||
}
|
||||
params.speculative.ngram_check_rate = value;
|
||||
}
|
||||
).set_examples({LLAMA_EXAMPLE_SERVER}));
|
||||
add_opt(common_arg(
|
||||
{"--spec-ngram-min-hits"}, "N",
|
||||
string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),
|
||||
|
||||
@@ -380,15 +380,46 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
|
||||
return msgs;
|
||||
}
|
||||
|
||||
json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
|
||||
static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
|
||||
if (!c.supports_string_content && !c.supports_typed_content) {
|
||||
LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
|
||||
}
|
||||
|
||||
bool only_string_accepted = c.supports_string_content && !c.supports_typed_content;
|
||||
bool only_typed_accepted = !c.supports_string_content && c.supports_typed_content;
|
||||
|
||||
json messages = json::array();
|
||||
for (const auto & msg : msgs) {
|
||||
json jmsg = msg.to_json_oaicompat(concat_typed_text);
|
||||
messages.push_back(jmsg);
|
||||
if (only_string_accepted) {
|
||||
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
|
||||
messages.push_back(jmsg);
|
||||
} else if (only_typed_accepted) {
|
||||
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
|
||||
if (jmsg.at("content").is_string()) {
|
||||
jmsg["content"] = json::array({
|
||||
json{
|
||||
{"type", "text"},
|
||||
{"text", jmsg.at("content").get<std::string>()},
|
||||
}
|
||||
});
|
||||
}
|
||||
messages.push_back(jmsg);
|
||||
} else {
|
||||
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
|
||||
messages.push_back(jmsg);
|
||||
}
|
||||
}
|
||||
return messages;
|
||||
}
|
||||
|
||||
// DEPRECATED: only used in tests
|
||||
json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
|
||||
jinja::caps c;
|
||||
c.supports_string_content = true;
|
||||
c.supports_typed_content = !concat_typed_text;
|
||||
return render_message_to_json(msgs, c);
|
||||
}
|
||||
|
||||
std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
|
||||
std::vector<common_chat_tool> result;
|
||||
|
||||
@@ -3020,7 +3051,7 @@ static common_chat_params common_chat_templates_apply_jinja(
|
||||
: *tmpls->template_default;
|
||||
const auto & src = tmpl.source();
|
||||
const auto & caps = tmpl.original_caps();
|
||||
params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
|
||||
params.messages = render_message_to_json(inputs.messages, tmpl.original_caps());
|
||||
params.add_generation_prompt = inputs.add_generation_prompt;
|
||||
params.tool_choice = inputs.tool_choice;
|
||||
params.reasoning_format = inputs.reasoning_format;
|
||||
|
||||
@@ -240,6 +240,8 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
|
||||
|
||||
// Parses a JSON array of messages in OpenAI's chat completion API format.
|
||||
std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
|
||||
|
||||
// DEPRECATED: only used in tests
|
||||
nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
|
||||
|
||||
std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
|
||||
|
||||
@@ -269,7 +269,6 @@ struct common_params_speculative {
|
||||
|
||||
uint16_t ngram_size_n = 12; // ngram size for lookup
|
||||
uint16_t ngram_size_m = 48; // mgram size for speculative tokens
|
||||
uint16_t ngram_check_rate = 1; // check rate for ngram lookup
|
||||
uint16_t ngram_min_hits = 1; // minimum hits at ngram/mgram lookup for mgram to be proposed
|
||||
|
||||
std::shared_ptr<common_ngram_mod> ngram_mod;
|
||||
|
||||
@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
|
||||
return v;
|
||||
}
|
||||
|
||||
#define INDENT " "
|
||||
|
||||
template <bool abort>
|
||||
void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
|
||||
GGML_ASSERT(n > 0);
|
||||
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
|
||||
}
|
||||
}
|
||||
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
|
||||
LOG_ERR(" [\n");
|
||||
LOG(INDENT "[\n");
|
||||
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
|
||||
if (i2 == n && ne[2] > 2 * n) {
|
||||
LOG_ERR(" ..., \n");
|
||||
LOG(INDENT INDENT "..., \n");
|
||||
i2 = ne[2] - n;
|
||||
}
|
||||
LOG_ERR(" [\n");
|
||||
LOG(INDENT INDENT "[\n");
|
||||
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
|
||||
if (i1 == n && ne[1] > 2 * n) {
|
||||
LOG_ERR(" ..., \n");
|
||||
LOG(INDENT INDENT INDENT "..., \n");
|
||||
i1 = ne[1] - n;
|
||||
}
|
||||
LOG_ERR(" [");
|
||||
LOG(INDENT INDENT INDENT "[");
|
||||
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
|
||||
if (i0 == n && ne[0] > 2 * n) {
|
||||
LOG_ERR("..., ");
|
||||
LOG(" ..., ");
|
||||
i0 = ne[0] - n;
|
||||
}
|
||||
const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
|
||||
LOG_ERR("%12.4f", v);
|
||||
LOG("%12.4f", v);
|
||||
if (i0 < ne[0] - 1) {
|
||||
LOG_ERR(", ");
|
||||
LOG(", ");
|
||||
}
|
||||
}
|
||||
LOG_ERR("],\n");
|
||||
LOG(" ],\n");
|
||||
}
|
||||
LOG_ERR(" ],\n");
|
||||
LOG(INDENT INDENT "],\n");
|
||||
}
|
||||
LOG_ERR(" ]\n");
|
||||
LOG_ERR(" sum = %f\n", sum);
|
||||
LOG(INDENT "]\n");
|
||||
LOG(INDENT "sum = %f\n", sum);
|
||||
}
|
||||
|
||||
if constexpr (abort) {
|
||||
if (std::isnan(sum)) {
|
||||
LOG_ERR("encountered NaN - aborting\n");
|
||||
LOG("encountered NaN - aborting\n");
|
||||
exit(0);
|
||||
}
|
||||
}
|
||||
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
|
||||
}
|
||||
|
||||
if (matches_filter) {
|
||||
LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
|
||||
ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
|
||||
common_ggml_ne_string(t).c_str());
|
||||
LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
|
||||
ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
|
||||
common_ggml_ne_string(t).c_str());
|
||||
}
|
||||
|
||||
const bool is_host = ggml_backend_buffer_is_host(t->buffer);
|
||||
|
||||
@@ -63,7 +63,8 @@ static void caps_print_stats(value & v, const std::string & path) {
|
||||
|
||||
std::map<std::string, bool> caps::to_map() const {
|
||||
return {
|
||||
{"requires_typed_content", requires_typed_content},
|
||||
{"supports_string_content", supports_string_content},
|
||||
{"supports_typed_content", supports_typed_content},
|
||||
{"supports_tools", supports_tools},
|
||||
{"supports_tool_calls", supports_tool_calls},
|
||||
{"supports_parallel_tool_calls", supports_parallel_tool_calls},
|
||||
@@ -89,7 +90,7 @@ caps caps_get(jinja::program & prog) {
|
||||
return v->stats.ops.find(op_name) != v->stats.ops.end();
|
||||
};
|
||||
|
||||
// case: typed content requirement
|
||||
// case: typed content support
|
||||
caps_try_execute(
|
||||
prog,
|
||||
[&]() {
|
||||
@@ -105,12 +106,16 @@ caps caps_get(jinja::program & prog) {
|
||||
// tools
|
||||
return json{nullptr};
|
||||
},
|
||||
[&](bool, value & messages, value &) {
|
||||
[&](bool success, value & messages, value &) {
|
||||
auto & content = messages->at(0)->at("content");
|
||||
caps_print_stats(content, "messages[0].content");
|
||||
if (has_op(content, "selectattr") || has_op(content, "array_access")) {
|
||||
// accessed as an array
|
||||
result.requires_typed_content = true;
|
||||
result.supports_typed_content = true;
|
||||
}
|
||||
if (!success) {
|
||||
// failed to execute with content as string
|
||||
result.supports_string_content = false;
|
||||
}
|
||||
}
|
||||
);
|
||||
|
||||
@@ -14,7 +14,9 @@ struct caps {
|
||||
bool supports_parallel_tool_calls = true;
|
||||
bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
|
||||
|
||||
bool requires_typed_content = false; // default: use string content
|
||||
// one of the 2 content capabilities must be true
|
||||
bool supports_string_content = true;
|
||||
bool supports_typed_content = false;
|
||||
|
||||
// for reporting on server
|
||||
std::map<std::string, bool> to_map() const;
|
||||
|
||||
@@ -446,6 +446,12 @@ value for_statement::execute_impl(context & ctx) {
|
||||
|
||||
value iterable_val = iter_expr->execute(scope);
|
||||
|
||||
// mark the variable being iterated as used for stats
|
||||
if (ctx.is_get_stats) {
|
||||
iterable_val->stats.used = true;
|
||||
iterable_val->stats.ops.insert("array_access");
|
||||
}
|
||||
|
||||
if (iterable_val->is_undefined()) {
|
||||
JJ_DEBUG("%s", "For loop iterable is undefined, skipping loop");
|
||||
iterable_val = mk_val<value_array>();
|
||||
|
||||
@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
|
||||
* @return Vector of draft tokens, empty if no matching pattern is found
|
||||
*/
|
||||
llama_tokens common_ngram_simple_draft(
|
||||
common_ngram_simple_state & state,
|
||||
const common_ngram_simple_config & config,
|
||||
const llama_tokens & tokens, llama_token sampled) {
|
||||
|
||||
// Simple implementation of self-speculative decoding without a draft model.
|
||||
//
|
||||
const size_t cur_len = tokens.size();
|
||||
// Only check every check_rate tokens to save compute
|
||||
// i.e., perform check if (cur_len - idx_last_check) >= check_rate
|
||||
if (state.idx_last_check + state.config.check_rate > cur_len) {
|
||||
llama_tokens draft_tokens;
|
||||
return draft_tokens;
|
||||
}
|
||||
|
||||
size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
|
||||
size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
|
||||
const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
|
||||
const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft
|
||||
|
||||
// vector for tokens we want to verify.
|
||||
// return empty vector if there is no match.
|
||||
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
|
||||
}
|
||||
pattern.push_back(sampled); // add the last token to the pattern
|
||||
|
||||
// We do a search in the token history.
|
||||
state.idx_last_check = cur_len;
|
||||
|
||||
size_t match_pos = 0; // we ignore position 0, position 0 == no match
|
||||
// search backwards, but skip the current match (we are currently there)
|
||||
for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
|
||||
@@ -240,10 +231,9 @@ void common_ngram_map_draft(common_ngram_map & map,
|
||||
GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
|
||||
}
|
||||
|
||||
// Only check every check_rate tokens to save compute
|
||||
// i.e., perform check if (cur_len - idx_last_check) >= check_rate
|
||||
if (map.idx_last_check + map.check_rate > cur_len) {
|
||||
return;
|
||||
if (map.idx_last_check > cur_len) {
|
||||
// Should not happen because of common_ngram_map_begin().
|
||||
GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
|
||||
}
|
||||
map.idx_last_check = cur_len;
|
||||
|
||||
@@ -471,7 +461,7 @@ void common_ngram_map_draft(common_ngram_map & map,
|
||||
slot_max = v;
|
||||
}
|
||||
}
|
||||
// What is sum of the other occurences?
|
||||
// What is sum of the other occurrences?
|
||||
uint32_t sum_occur = 0;
|
||||
for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
|
||||
if (v == slot_max) {
|
||||
|
||||
@@ -24,26 +24,11 @@
|
||||
struct common_ngram_simple_config {
|
||||
uint16_t size_ngram; // size of n-grams to lookup in self-mode
|
||||
uint16_t size_mgram; // size of m-grams to draft in self-mode
|
||||
uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
|
||||
};
|
||||
|
||||
// current state (and config) of n-gram simple.
|
||||
struct common_ngram_simple_state {
|
||||
common_ngram_simple_config config;
|
||||
|
||||
size_t idx_last_check = 0; // index of last check in context history (mutable)
|
||||
|
||||
common_ngram_simple_state(const common_ngram_simple_config & config)
|
||||
: config(config) {}
|
||||
};
|
||||
|
||||
// Searches for a n-gram in the history and checks whether a draft sequence should be generated.
|
||||
// state: the ngram simple state to search in.
|
||||
// inp: the tokens generated so far.
|
||||
// sampled: the token that was just sampled.
|
||||
// draft: vector to store the draft tokens, initially empty.
|
||||
llama_tokens common_ngram_simple_draft(
|
||||
common_ngram_simple_state & state,
|
||||
const common_ngram_simple_config & config,
|
||||
const llama_tokens & tokens, llama_token sampled);
|
||||
|
||||
|
||||
@@ -59,7 +44,7 @@ llama_tokens common_ngram_simple_draft(
|
||||
// statistics of a m-gram after a known n-gram
|
||||
struct common_ngram_map_value {
|
||||
size_t value_idx = 0; // index of value m-gram in token-history (0 if unused)
|
||||
uint16_t value_num = 0; // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
|
||||
uint16_t value_num = 0; // number of occurrences of this value m-gram after the key n-gram (0 in an unused values-slot)
|
||||
int16_t n_accepted = -1; // number of accepted tokens at last draft (-1 if unused)
|
||||
};
|
||||
|
||||
@@ -68,7 +53,7 @@ struct common_ngram_map_key {
|
||||
size_t key_idx; // index of key n-gram in token-history
|
||||
size_t stat_idx; // index of last token of stastistics computation (key_num, values)
|
||||
|
||||
uint16_t key_num; // number of occurences of this key n-gram in token-history
|
||||
uint16_t key_num; // number of occurrences of this key n-gram in token-history
|
||||
common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
|
||||
};
|
||||
|
||||
@@ -80,15 +65,14 @@ struct common_ngram_map {
|
||||
bool key_only; // true if only key n-grams are used, no values.
|
||||
|
||||
std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
|
||||
uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
|
||||
uint16_t min_hits; // minimum number of key hits to consider a draft
|
||||
|
||||
bool show_key_map_stats = false; // true, if statitics of the key_map should be printed.
|
||||
bool show_key_map_stats = false; // true, if statistics of the key_map should be printed.
|
||||
|
||||
common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
|
||||
uint16_t check_rate, uint16_t min_hits)
|
||||
uint16_t min_hits)
|
||||
: size_key(sz_key), size_value(sz_value), key_only(only_keys),
|
||||
check_rate(check_rate), min_hits(min_hits) {
|
||||
min_hits(min_hits) {
|
||||
key_map.resize(COMMON_NGRAM_HASH_MAP_SIZE); // 2^18 hash entries, 0 entries if key_map shouldn't be used
|
||||
}
|
||||
|
||||
|
||||
@@ -113,13 +113,14 @@ static bool common_speculative_are_compatible(
|
||||
struct common_speculative_state {
|
||||
const enum common_speculative_type type;
|
||||
|
||||
// TODO: rename to n_call_draft, n_gen_drafts, n_acc_drafts, n_gen_tokens, n_acc_tokens
|
||||
// TODO: add n_call_begin, n_call_accept
|
||||
size_t drafts_call_count = 0; // number of times this implementation was called.
|
||||
size_t drafts_generated_count = 0; // number of times a draft or part was generated by this implementation.
|
||||
size_t drafts_accepted_count = 0; // number of times a draft or part was accepted by the target model.
|
||||
size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
|
||||
size_t drafts_accepted_tokens = 0; // number of tokens accepted by the target model.
|
||||
size_t n_call_begin = 0; // number of times this implementation was called for refresh.
|
||||
size_t n_call_draft = 0; // number of times this implementation was called for generation.
|
||||
size_t n_call_accept = 0; // number of times this implementation was called for accumulation.
|
||||
|
||||
size_t n_gen_drafts = 0; // number of times a draft or part was generated by this implementation.
|
||||
size_t n_acc_drafts = 0; // number of times a draft or part was accepted by the target model.
|
||||
size_t n_gen_tokens = 0; // number of tokens generated by this implementation.
|
||||
size_t n_acc_tokens = 0; // number of tokens accepted by the target model.
|
||||
|
||||
// TODO: track performance of most recent calls
|
||||
const bool gen_perf = true; // whether to generate performance stats.
|
||||
@@ -463,12 +464,12 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
|
||||
|
||||
// state of self-speculation (simple implementation, not ngram-map)
|
||||
struct common_speculative_state_ngram_simple : public common_speculative_state {
|
||||
common_ngram_simple_state state;
|
||||
common_ngram_simple_config config;
|
||||
|
||||
common_speculative_state_ngram_simple(
|
||||
enum common_speculative_type type,
|
||||
common_ngram_simple_state state)
|
||||
: common_speculative_state(type), state(state) {}
|
||||
common_ngram_simple_config config)
|
||||
: common_speculative_state(type), config(config) {}
|
||||
|
||||
void begin(const llama_tokens & prompt) override {
|
||||
GGML_UNUSED(prompt);
|
||||
@@ -479,7 +480,8 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
|
||||
const llama_tokens & prompt_tgt,
|
||||
llama_token id_last,
|
||||
llama_tokens & result) override {
|
||||
result = common_ngram_simple_draft(state, prompt_tgt, id_last);
|
||||
|
||||
result = common_ngram_simple_draft(config, prompt_tgt, id_last);
|
||||
GGML_UNUSED(params);
|
||||
}
|
||||
|
||||
@@ -744,10 +746,9 @@ static common_ngram_map get_common_ngram_map(const common_speculative_config & c
|
||||
uint16_t size_key = config.params.ngram_size_n;
|
||||
uint16_t size_value = config.params.ngram_size_m;
|
||||
bool key_only = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
|
||||
uint16_t check_rate = config.params.ngram_check_rate;
|
||||
uint16_t min_hits = config.params.ngram_min_hits;
|
||||
|
||||
return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
|
||||
return common_ngram_map(size_key, size_value, key_only, min_hits);
|
||||
}
|
||||
|
||||
static common_speculative_state_ngram_cache create_state_ngram_cache(
|
||||
@@ -797,6 +798,42 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
|
||||
return it->second;
|
||||
}
|
||||
|
||||
bool common_speculative_is_compat(llama_context * ctx_tgt) {
|
||||
auto * mem = llama_get_memory(ctx_tgt);
|
||||
if (mem == nullptr) {
|
||||
return false;
|
||||
}
|
||||
|
||||
bool res = true;
|
||||
|
||||
llama_memory_clear(mem, true);
|
||||
|
||||
// eval 2 tokens to check if the context is compatible
|
||||
std::vector<llama_token> tmp;
|
||||
tmp.push_back(0);
|
||||
tmp.push_back(0);
|
||||
|
||||
int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
|
||||
if (ret != 0) {
|
||||
LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
|
||||
res = false;
|
||||
goto done;
|
||||
}
|
||||
|
||||
// try to remove the last tokens
|
||||
if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
|
||||
LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
|
||||
res = false;
|
||||
goto done;
|
||||
}
|
||||
|
||||
done:
|
||||
llama_memory_clear(mem, true);
|
||||
llama_synchronize(ctx_tgt);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
// initialization of the speculative decoding system
|
||||
//
|
||||
common_speculative * common_speculative_init(
|
||||
@@ -887,16 +924,14 @@ common_speculative * common_speculative_init(
|
||||
|
||||
uint16_t ngram_size_key = ngram_map.size_key;
|
||||
uint16_t mgram_size_value = ngram_map.size_value;
|
||||
uint16_t check_rate = ngram_map.check_rate;
|
||||
|
||||
auto config_simple = common_ngram_simple_config{
|
||||
auto config_simple = common_ngram_simple_config {
|
||||
/* .size_ngram = */ ngram_size_key,
|
||||
/* .size_mgram = */ mgram_size_value,
|
||||
/* .check_rate = */ check_rate
|
||||
/* .size_mgram = */ mgram_size_value
|
||||
};
|
||||
auto state = std::make_unique<common_speculative_state_ngram_simple>(
|
||||
/* .type = */ config.type,
|
||||
/* .state = */ common_ngram_simple_state(config_simple)
|
||||
/* .state = */ config_simple
|
||||
);
|
||||
impls.push_back(std::move(state));
|
||||
break;
|
||||
@@ -953,6 +988,7 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
|
||||
for (auto & impl : spec->impls) {
|
||||
common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
|
||||
impl->begin(prompt);
|
||||
impl->n_call_begin++;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -969,17 +1005,17 @@ llama_tokens common_speculative_draft(
|
||||
{
|
||||
common_time_meas tm(impl->t_draft_us, !impl->gen_perf);
|
||||
impl->draft(params, prompt_tgt, id_last, result);
|
||||
impl->drafts_call_count++;
|
||||
impl->n_call_draft++;
|
||||
}
|
||||
|
||||
if (!result.empty()) {
|
||||
LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
|
||||
common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
|
||||
impl.get()->drafts_call_count, result.size());
|
||||
impl.get()->n_call_draft, result.size());
|
||||
|
||||
spec->curr_impl = impl.get(); // set current implementation for stats
|
||||
impl->drafts_generated_count++;
|
||||
impl->drafts_generated_tokens += result.size();
|
||||
impl->n_gen_drafts++;
|
||||
impl->n_gen_tokens += result.size();
|
||||
|
||||
break; // We have a draft, so break out of the loop and return it.
|
||||
}
|
||||
@@ -1000,11 +1036,12 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
|
||||
{
|
||||
common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
|
||||
if (n_accepted > 0) {
|
||||
impl->drafts_accepted_count++;
|
||||
impl->drafts_accepted_tokens += n_accepted;
|
||||
impl->n_acc_drafts++;
|
||||
impl->n_acc_tokens += n_accepted;
|
||||
}
|
||||
|
||||
impl->accept(n_accepted);
|
||||
impl->n_call_accept++;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1025,13 +1062,13 @@ void common_speculative_print_stats(const common_speculative * spec) {
|
||||
str_perf = "";
|
||||
}
|
||||
|
||||
LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
|
||||
LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
|
||||
common_speculative_type_to_str(impl->type).c_str(),
|
||||
impl->drafts_call_count,
|
||||
impl->drafts_generated_count,
|
||||
impl->drafts_accepted_count,
|
||||
impl->drafts_generated_tokens,
|
||||
impl->drafts_accepted_tokens,
|
||||
impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
|
||||
impl->n_gen_drafts,
|
||||
impl->n_acc_drafts,
|
||||
impl->n_gen_tokens,
|
||||
impl->n_acc_tokens,
|
||||
str_perf.c_str());
|
||||
}
|
||||
}
|
||||
|
||||
@@ -14,6 +14,10 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
|
||||
// convert type to string
|
||||
std::string common_speculative_type_to_str(enum common_speculative_type type);
|
||||
|
||||
// check if the llama_context is compatible for speculative decoding
|
||||
// note: clears the memory of the context
|
||||
bool common_speculative_is_compat(llama_context * ctx_tgt);
|
||||
|
||||
common_speculative * common_speculative_init(
|
||||
common_params_speculative & params,
|
||||
llama_context * ctx_tgt);
|
||||
|
||||
@@ -586,6 +586,10 @@ class ModelBase:
|
||||
gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
|
||||
gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
|
||||
gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
|
||||
# Kimi KDA conv weights should be F32
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_Q,
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_K,
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_V,
|
||||
)
|
||||
)
|
||||
or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
|
||||
@@ -903,10 +907,10 @@ class TextModel(ModelBase):
|
||||
if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
|
||||
self.gguf_writer.add_layer_norm_eps(f_norm_eps)
|
||||
logger.info(f"gguf: layer norm epsilon = {f_norm_eps}")
|
||||
if (n_experts := self.hparams.get("num_local_experts")) is not None:
|
||||
if (n_experts := self.find_hparam(["num_local_experts", "num_experts"], optional=True)) is not None:
|
||||
self.gguf_writer.add_expert_count(n_experts)
|
||||
logger.info(f"gguf: expert count = {n_experts}")
|
||||
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
|
||||
if (n_experts_used := self.find_hparam(["num_experts_per_tok", "num_experts_per_token"], optional=True)) is not None:
|
||||
self.gguf_writer.add_expert_used_count(n_experts_used)
|
||||
logger.info(f"gguf: experts used count = {n_experts_used}")
|
||||
if (n_expert_groups := self.hparams.get("n_group")) is not None:
|
||||
@@ -916,7 +920,7 @@ class TextModel(ModelBase):
|
||||
self.gguf_writer.add_expert_group_used_count(n_group_used)
|
||||
logger.info(f"gguf: expert groups used count = {n_group_used}")
|
||||
|
||||
if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func"], optional=True)) is not None:
|
||||
if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func"], optional=True)) is not None:
|
||||
if score_func == "sigmoid":
|
||||
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
|
||||
elif score_func == "softmax":
|
||||
@@ -1257,6 +1261,9 @@ class TextModel(ModelBase):
|
||||
if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
|
||||
# ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
|
||||
res = "exaone-moe"
|
||||
if chkhsh == "d30d75d9059f1aa2c19359de71047b3ae408c70875e8a3ccf8c5fba56c9d8af4":
|
||||
# ref: https://huggingface.co/Qwen/Qwen3.5-9B-Instruct
|
||||
res = "qwen35"
|
||||
|
||||
if res is None:
|
||||
logger.warning("\n")
|
||||
@@ -4105,37 +4112,29 @@ class Qwen2MoeModel(TextModel):
|
||||
# Expected GGML ne: {n_embd, n_ff_exp, n_expert} for gate/up, {n_ff_exp, n_embd, n_expert} for down
|
||||
if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
|
||||
mapped = f"{name}.weight" if not name.endswith(".weight") else name
|
||||
# Input: (n_expert=128, n_ff_exp=768, n_embd=2048)
|
||||
# Want GGML ne: {n_ff_exp, n_embd, n_expert} = {768, 2048, 128}
|
||||
# Need PyTorch: (128, 2048, 768) [reversed of GGML]
|
||||
# So: permute(0, 2, 1): (128, 768, 2048) -> (128, 2048, 768)
|
||||
permuted = data_torch.permute(0, 2, 1).contiguous()
|
||||
yield from super().modify_tensors(permuted, mapped, bid)
|
||||
# HF: [n_expert, n_embd, n_ff] -> GGML: {n_ff, n_embd, n_expert}
|
||||
yield from super().modify_tensors(data_torch, mapped, bid)
|
||||
return
|
||||
|
||||
if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
|
||||
if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
|
||||
if data_torch.ndim < 3 or data_torch.shape[-2] % 2 != 0:
|
||||
raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
|
||||
split_dim = data_torch.shape[-1] // 2
|
||||
gate = data_torch[..., :split_dim].contiguous()
|
||||
up = data_torch[..., split_dim:].contiguous()
|
||||
# Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
|
||||
# Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
|
||||
# Need PyTorch: (128, 768, 2048) [reversed of GGML]
|
||||
# So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
|
||||
base_name = name.removesuffix(".weight")
|
||||
base = base_name.rsplit('.', 1)[0]
|
||||
mapped_gate = f"{base}.gate_proj.weight"
|
||||
mapped_up = f"{base}.up_proj.weight"
|
||||
perm_gate = gate.permute(0, 2, 1).contiguous()
|
||||
perm_up = up.permute(0, 2, 1).contiguous()
|
||||
yield from super().modify_tensors(perm_gate, mapped_gate, bid)
|
||||
yield from super().modify_tensors(perm_up, mapped_up, bid)
|
||||
# HF: [n_expert, 2*n_ff, n_embd] -> split on dim=-2
|
||||
n_ff = data_torch.shape[-2] // 2
|
||||
gate = data_torch[..., :n_ff, :].contiguous()
|
||||
up = data_torch[..., n_ff:, :].contiguous()
|
||||
# gate/up: [n_expert, n_ff, n_embd] -> GGML: {n_embd, n_ff, n_expert}
|
||||
base_name = name.removesuffix(".weight").removesuffix(".gate_up_proj")
|
||||
mapped_gate = f"{base_name}.gate_proj.weight"
|
||||
mapped_up = f"{base_name}.up_proj.weight"
|
||||
yield from super().modify_tensors(gate, mapped_gate, bid)
|
||||
yield from super().modify_tensors(up, mapped_up, bid)
|
||||
return
|
||||
|
||||
if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") or name.startswith("model.visual"):
|
||||
# skip visual tensors
|
||||
return
|
||||
|
||||
if name.find("experts") != -1:
|
||||
n_experts = self.hparams["num_experts"]
|
||||
assert bid is not None
|
||||
@@ -4291,6 +4290,7 @@ class Qwen3NextModel(Qwen2MoeModel):
|
||||
self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
|
||||
self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
|
||||
self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
|
||||
self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
|
||||
if (rope_dim := self.hparams.get("head_dim")) is None:
|
||||
rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
|
||||
self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))
|
||||
@@ -4355,7 +4355,7 @@ class RND1Model(Qwen2MoeModel):
|
||||
self.gguf_writer.add_mask_token_id(mask_token_id)
|
||||
|
||||
|
||||
@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
|
||||
@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration", "Qwen3_5ForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration")
|
||||
class Qwen3VLVisionModel(MmprojModel):
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
@@ -4401,6 +4401,10 @@ class Qwen3VLVisionModel(MmprojModel):
|
||||
if name.startswith("model.language_model.") or name.startswith("lm_head."):
|
||||
return
|
||||
|
||||
# Skip MTP tensors
|
||||
if name.startswith("mtp."):
|
||||
return
|
||||
|
||||
if name.startswith("model.visual."):
|
||||
name = name.replace("model.visual.", "visual.", 1)
|
||||
|
||||
@@ -4531,9 +4535,125 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel):
|
||||
if name.startswith("model.visual."):
|
||||
return
|
||||
|
||||
# Qwen3VL has transposed packed tensors, so we treat it differently from general Qwen2MoE packed tensors
|
||||
if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
|
||||
name = name.replace("language_model.", "")
|
||||
mapped = f"{name}.weight" if not name.endswith(".weight") else name
|
||||
permuted = data_torch.permute(0, 2, 1).contiguous()
|
||||
yield from ModelBase.modify_tensors(self, permuted, mapped, bid)
|
||||
return
|
||||
|
||||
if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
|
||||
name = name.replace("language_model.", "")
|
||||
if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
|
||||
raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
|
||||
split_dim = data_torch.shape[-1] // 2
|
||||
gate = data_torch[..., :split_dim].contiguous()
|
||||
up = data_torch[..., split_dim:].contiguous()
|
||||
# Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
|
||||
# Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
|
||||
# Need PyTorch: (128, 768, 2048) [reversed of GGML]
|
||||
# So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
|
||||
base_name = name.removesuffix(".weight")
|
||||
base = base_name.rsplit('.', 1)[0]
|
||||
mapped_gate = f"{base}.gate_proj.weight"
|
||||
mapped_up = f"{base}.up_proj.weight"
|
||||
perm_gate = gate.permute(0, 2, 1).contiguous()
|
||||
perm_up = up.permute(0, 2, 1).contiguous()
|
||||
yield from ModelBase.modify_tensors(self, perm_gate, mapped_gate, bid)
|
||||
yield from ModelBase.modify_tensors(self, perm_up, mapped_up, bid)
|
||||
return
|
||||
|
||||
yield from super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
|
||||
class _LinearAttentionVReorderBase(Qwen3NextModel):
|
||||
model_arch = gguf.MODEL_ARCH.QWEN3NEXT # overridden by subclasses
|
||||
"""reorders V heads from grouped to tiled order for ggml broadcast
|
||||
|
||||
see https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306
|
||||
|
||||
Linear attention may has num_k_heads < num_v_heads. The HF weights store
|
||||
V heads grouped by K head: [G0_v0..v{r-1}, G1_v0..v{r-1}, ...].
|
||||
ggml binary ops use tiled broadcast: [K0, K1, ..., K0, K1, ...].
|
||||
We reorder V heads to tiled order so ggml_repeat can replace the expensive
|
||||
interleaved repeat: [G0_v0, G1_v0, ..., G0_v1, G1_v1, ...].
|
||||
"""
|
||||
|
||||
@staticmethod
|
||||
def _reorder_v_heads(tensor: Tensor, dim: int, num_k_heads: int, num_v_per_k: int, head_dim: int) -> Tensor:
|
||||
"""Reorder V heads from grouped (by K head) to tiled order along the given dimension."""
|
||||
shape = list(tensor.shape)
|
||||
if dim < 0:
|
||||
dim += len(shape)
|
||||
new_shape = shape[:dim] + [num_k_heads, num_v_per_k, head_dim] + shape[dim + 1:]
|
||||
tensor = tensor.reshape(*new_shape)
|
||||
perm = list(range(len(new_shape)))
|
||||
perm[dim], perm[dim + 1] = perm[dim + 1], perm[dim]
|
||||
return tensor.permute(*perm).contiguous().reshape(*shape)
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
||||
num_k_heads = self.hparams.get("linear_num_key_heads", 0)
|
||||
num_v_heads = self.hparams.get("linear_num_value_heads", 0)
|
||||
|
||||
if num_k_heads > 0 and num_v_heads > 0 and num_k_heads != num_v_heads and "linear_attn." in name:
|
||||
head_k_dim = self.hparams["linear_key_head_dim"]
|
||||
head_v_dim = self.hparams["linear_value_head_dim"]
|
||||
num_v_per_k = num_v_heads // num_k_heads
|
||||
|
||||
if ".in_proj_qkv." in name:
|
||||
# QKV weight: reorder only the V rows
|
||||
q_dim = head_k_dim * num_k_heads
|
||||
k_dim = head_k_dim * num_k_heads
|
||||
q = data_torch[:q_dim]
|
||||
k = data_torch[q_dim:q_dim + k_dim]
|
||||
v = data_torch[q_dim + k_dim:]
|
||||
v = self._reorder_v_heads(v, 0, num_k_heads, num_v_per_k, head_v_dim)
|
||||
data_torch = torch.cat([q, k, v], dim=0)
|
||||
|
||||
elif ".in_proj_z." in name:
|
||||
# Z gate weight: reorder rows (num_v_heads * head_v_dim)
|
||||
data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, head_v_dim)
|
||||
|
||||
elif ".in_proj_b." in name or ".in_proj_a." in name:
|
||||
# Beta/Alpha weight: reorder rows (num_v_heads, head_dim=1)
|
||||
data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, 1)
|
||||
|
||||
elif ".A_log" in name or ".dt_bias" in name or ".dt_proj" in name:
|
||||
# A_log / dt_bias: 1D parameters with num_v_heads elements
|
||||
if data_torch.ndim == 1:
|
||||
data_torch = self._reorder_v_heads(
|
||||
data_torch.unsqueeze(-1), 0, num_k_heads, num_v_per_k, 1
|
||||
).squeeze(-1)
|
||||
else:
|
||||
data_torch = self._reorder_v_heads(data_torch, -1, num_k_heads, num_v_per_k, 1)
|
||||
|
||||
elif ".conv1d" in name:
|
||||
# Conv1d kernel: reorder only the V channel portion
|
||||
data = data_torch.squeeze()
|
||||
qk_channels = head_k_dim * num_k_heads * 2
|
||||
qk_part = data[:qk_channels]
|
||||
v_part = data[qk_channels:]
|
||||
v_part = self._reorder_v_heads(v_part, 0, num_k_heads, num_v_per_k, head_v_dim)
|
||||
data_torch = torch.cat([qk_part, v_part], dim=0)
|
||||
|
||||
elif ".out_proj." in name:
|
||||
# Out projection weight: reorder columns (input dimension)
|
||||
data_torch = self._reorder_v_heads(data_torch, 1, num_k_heads, num_v_per_k, head_v_dim)
|
||||
|
||||
yield from super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
|
||||
@ModelBase.register("Qwen3_5ForConditionalGeneration")
|
||||
class Qwen3_5TextModel(_LinearAttentionVReorderBase):
|
||||
model_arch = gguf.MODEL_ARCH.QWEN35
|
||||
|
||||
|
||||
@ModelBase.register("Qwen3_5MoeForConditionalGeneration")
|
||||
class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase):
|
||||
model_arch = gguf.MODEL_ARCH.QWEN35MOE
|
||||
|
||||
|
||||
@ModelBase.register("GPT2LMHeadModel")
|
||||
class GPT2Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.GPT2
|
||||
@@ -5013,6 +5133,221 @@ class CodeShellModel(TextModel):
|
||||
self.gguf_writer.add_rope_scaling_factor(1.0)
|
||||
|
||||
|
||||
@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")
|
||||
class KimiLinearModel(TextModel):
|
||||
"""Kimi-Linear model with hybrid MLA+KDA architecture"""
|
||||
model_arch = gguf.MODEL_ARCH.KIMI_LINEAR
|
||||
|
||||
_experts: list[dict[str, Tensor]] | None = None
|
||||
|
||||
def set_vocab(self):
|
||||
try:
|
||||
self._set_vocab_gpt2()
|
||||
return
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
from transformers import AutoTokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
|
||||
tokpre = self.get_vocab_base_pre(tokenizer)
|
||||
|
||||
if tokpre == "kimi-k2":
|
||||
# Build merges list using the approach similar to HunYuanMoE
|
||||
merges = []
|
||||
vocab = {}
|
||||
mergeable_ranks = tokenizer.model._mergeable_ranks
|
||||
for token, rank in mergeable_ranks.items():
|
||||
vocab[QwenModel.token_bytes_to_string(token)] = rank
|
||||
if len(token) == 1:
|
||||
continue
|
||||
merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
|
||||
if len(merged) == 2:
|
||||
merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
|
||||
# Build token list
|
||||
vocab_size = self.hparams["vocab_size"]
|
||||
special_tokens = tokenizer.special_tokens
|
||||
reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
|
||||
tokens: list[str] = []
|
||||
toktypes: list[int] = []
|
||||
|
||||
for i in range(vocab_size):
|
||||
if i not in reverse_vocab:
|
||||
tokens.append(f"[PAD{i}]")
|
||||
toktypes.append(gguf.TokenType.UNUSED)
|
||||
else:
|
||||
token = reverse_vocab[i]
|
||||
tokens.append(token)
|
||||
if i in special_tokens.values():
|
||||
toktypes.append(gguf.TokenType.CONTROL)
|
||||
else:
|
||||
toktypes.append(gguf.TokenType.NORMAL)
|
||||
|
||||
self.gguf_writer.add_tokenizer_model("gpt2")
|
||||
self.gguf_writer.add_tokenizer_pre(tokpre)
|
||||
self.gguf_writer.add_token_list(tokens)
|
||||
self.gguf_writer.add_token_types(toktypes)
|
||||
self.gguf_writer.add_token_merges(merges)
|
||||
|
||||
special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
|
||||
special_vocab.add_to_gguf(self.gguf_writer)
|
||||
# override eos id in config.json with tiktoken eos id
|
||||
self.gguf_writer.add_eos_token_id(tokenizer.eos_id)
|
||||
else:
|
||||
raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
# note: To enable MLA KV cache, attention needs to be converted into MQA (ie: GQA with 1 group)
|
||||
self.hparams["num_key_value_heads"] = 1
|
||||
|
||||
super().set_gguf_parameters()
|
||||
self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
|
||||
|
||||
# KDA & MLA params
|
||||
# Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv
|
||||
linear_attn_config = self.hparams["linear_attn_config"]
|
||||
# n_head == 0 for KDA layers, n_head > 0 for MLA layers
|
||||
# full_attention_layers list will be used to distingush layer type
|
||||
_num_kv_heads = list()
|
||||
_full_attn_layers = linear_attn_config["full_attn_layers"]
|
||||
for il in range(self.hparams["num_hidden_layers"]):
|
||||
if il + 1 in _full_attn_layers:
|
||||
_num_kv_heads.append(self.hparams["num_key_value_heads"])
|
||||
else:
|
||||
_num_kv_heads.append(0)
|
||||
assert len(_num_kv_heads) == self.hparams["num_hidden_layers"]
|
||||
self.gguf_writer.add_head_count_kv(_num_kv_heads)
|
||||
|
||||
if (ssm_d_conv := linear_attn_config.get("short_conv_kernel_size")) is not None:
|
||||
self.gguf_writer.add_ssm_conv_kernel(ssm_d_conv)
|
||||
if (kda_head_dim := linear_attn_config.get("head_dim")) is not None:
|
||||
self.gguf_writer.add_kda_head_dim(kda_head_dim)
|
||||
|
||||
# MLA params - use add_* methods that handle arch substitution
|
||||
# Support both HuggingFace naming (q_lora_rank, kv_lora_rank) and internal naming (n_lora_q, n_lora_kv)
|
||||
if (q_lora_rank := self.find_hparam(["q_lora_rank", "n_lora_q"], optional=True)) is not None:
|
||||
self.gguf_writer.add_q_lora_rank(q_lora_rank)
|
||||
# To enable MLA KV cache, MLA needs to be converted into MQA with larger heads, then decompresses to MHA
|
||||
kv_lora_rank = self.find_hparam(["kv_lora_rank", "n_lora_kv"], optional=False)
|
||||
self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
|
||||
|
||||
# MLA head dimensions
|
||||
# Support HuggingFace naming: qk_nope_head_dim, qk_rope_head_dim, v_head_dim
|
||||
qk_nope_head_dim = self.hparams.get("qk_nope_head_dim")
|
||||
# Rotation - use qk_rope_head_dim for Kimi
|
||||
qk_rope_head_dim = self.find_hparam(["qk_rope_head_dim", "n_rot"], optional=False)
|
||||
self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
|
||||
self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
|
||||
v_head_dim = self.hparams.get("v_head_dim")
|
||||
|
||||
# Calculate n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
|
||||
if (n_embd_head_k_mla := self.find_hparam(["n_embd_head_k_mla"], optional=True)) is not None:
|
||||
self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
|
||||
elif qk_nope_head_dim is not None:
|
||||
n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
|
||||
self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
|
||||
|
||||
# n_embd_head_v_mla = v_head_dim
|
||||
if (n_embd_head_v_mla := self.hparams.get("n_embd_head_v_mla")) is not None:
|
||||
self.gguf_writer.add_value_length_mla(n_embd_head_v_mla)
|
||||
elif v_head_dim is not None:
|
||||
self.gguf_writer.add_value_length_mla(v_head_dim)
|
||||
|
||||
# moe_intermediate_size (1024 for Kimi)
|
||||
self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
|
||||
# num_shared_experts (1 for Kimi)
|
||||
self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
|
||||
# first_k_dense_replace (1 for Kimi - first layer uses dense MLP)
|
||||
self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
|
||||
# Routed scaling factor (expert_weights_scale = 2.446 for Kimi)
|
||||
self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
|
||||
|
||||
def prepare_tensors(self):
|
||||
super().prepare_tensors()
|
||||
if self._experts is not None:
|
||||
experts = [k for d in self._experts for k in d.keys()]
|
||||
if len(experts) > 0:
|
||||
raise ValueError(f"Unprocessed experts: {experts}")
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
||||
logger.info(f"Processing {name}: shape before = {tuple(data_torch.shape)}")
|
||||
|
||||
# Handle KDA conv1d weights
|
||||
# HuggingFace/vLLM stores as [d_inner, d_conv] (2D), memory layout: conv_step changes fastest
|
||||
# llama.cpp expects ggml ne = [d_conv, 1, d_inner, 1], memory layout: ne[0]=d_conv changes fastest
|
||||
# GGUF reverses numpy shape when writing, so numpy (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
|
||||
# Memory layouts match: both have conv_step (d_conv) changing fastest
|
||||
if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):
|
||||
# HF shape: [d_inner, d_conv] e.g. [4096, 4]
|
||||
# Target numpy shape: (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
|
||||
if data_torch.ndim == 2:
|
||||
d_inner, d_conv = data_torch.shape
|
||||
# Reshape to (1, d_inner, 1, d_conv) - memory layout preserved (d_conv fastest)
|
||||
data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
|
||||
logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
|
||||
elif data_torch.ndim == 3:
|
||||
# Already 3D [d_inner, 1, d_conv] from unsqueeze
|
||||
d_inner, _, d_conv = data_torch.shape
|
||||
data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
|
||||
logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, 1, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
|
||||
|
||||
# Kimi specific bias
|
||||
if name.endswith("e_score_correction_bias"):
|
||||
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
|
||||
|
||||
# Handle A_log: iHF stores as [1, 1, num_heads, 1]
|
||||
# llama.cpp expects ggml ne = [1, num_heads, 1, 1]
|
||||
# GGUF reverses numpy shape: numpy (1, 1, num_heads, 1) -> ggml ne = [1, num_heads, 1, 1]
|
||||
if name.endswith(".A_log"):
|
||||
data_torch = -torch.exp(data_torch)
|
||||
if name.endswith(".dt_bias"):
|
||||
name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
|
||||
logger.info("Changed dt_bias to dt_proj.bias")
|
||||
|
||||
# process the experts separately
|
||||
if name.find("block_sparse_moe.experts") != -1:
|
||||
n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=False)
|
||||
assert bid is not None
|
||||
|
||||
if self._experts is None:
|
||||
self._experts = [{} for _ in range(self.block_count)]
|
||||
|
||||
self._experts[bid][name] = data_torch
|
||||
|
||||
if len(self._experts[bid]) >= n_experts * 3:
|
||||
# merge the experts into a single 3d tensor
|
||||
# w1: gate, w2: down, w3: up
|
||||
for wid, tname in [("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),
|
||||
("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),
|
||||
("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)]:
|
||||
datas: list[Tensor] = []
|
||||
for xid in range(n_experts):
|
||||
ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
|
||||
datas.append(self._experts[bid][ename])
|
||||
del self._experts[bid][ename]
|
||||
data_torch = torch.stack(datas, dim=0)
|
||||
new_name = self.format_tensor_name(tname, bid)
|
||||
yield from super().modify_tensors(data_torch, new_name, bid)
|
||||
return
|
||||
|
||||
# note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
|
||||
if name.endswith("kv_b_proj.weight"):
|
||||
name_kb = name.replace("kv_b_proj", "k_b_proj")
|
||||
name_vb = name.replace("kv_b_proj", "v_b_proj")
|
||||
n_head_kv = self.hparams["num_key_value_heads"]
|
||||
v_head_dim = self.find_hparam(["n_embd_head_v_mla", "v_head_dim"], optional=False)
|
||||
qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
|
||||
logger.info("Split kv_b n_head_kv %d\n" % n_head_kv)
|
||||
assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
|
||||
kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
|
||||
k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
|
||||
k_b = k_b.transpose(1, 2)
|
||||
yield from super().modify_tensors(k_b, name_kb, bid)
|
||||
yield from super().modify_tensors(v_b, name_vb, bid)
|
||||
return
|
||||
|
||||
yield from super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
|
||||
@ModelBase.register("InternLM2ForCausalLM")
|
||||
class InternLM2Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.INTERNLM2
|
||||
@@ -7693,6 +8028,135 @@ class MimoV2Model(TextModel):
|
||||
raise ValueError(f"Unprocessed experts: {experts}")
|
||||
|
||||
|
||||
@ModelBase.register("Step3p5ForCausalLM")
|
||||
class Step35Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.STEP35
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
rope_theta = self.hparams.get("rope_theta")
|
||||
if isinstance(rope_theta, list):
|
||||
self.hparams["rope_theta"] = float(rope_theta[0])
|
||||
self.hparams["local_rope_theta"] = float(rope_theta[1])
|
||||
self.rope_parameters["rope_theta"] = self.hparams["rope_theta"]
|
||||
self.rope_parameters["sliding_attention"] = {"rope_theta": self.hparams["local_rope_theta"]}
|
||||
|
||||
super().set_gguf_parameters()
|
||||
|
||||
layer_types = self.hparams.get("layer_types") or []
|
||||
partial_rotary_factors = self.hparams.get("partial_rotary_factors") or []
|
||||
attn_other = self.hparams.get("attention_other_setting") or {}
|
||||
|
||||
n_head_base = self.hparams["num_attention_heads"]
|
||||
n_kv_base = self.hparams["num_attention_groups"]
|
||||
|
||||
n_head_swa = attn_other.get("num_attention_heads", n_head_base)
|
||||
n_kv_swa = attn_other.get("num_attention_groups", n_kv_base)
|
||||
|
||||
layer_types = layer_types[: self.block_count]
|
||||
partial_rotary_factors = partial_rotary_factors[: self.block_count]
|
||||
assert [1.0 if lt == "sliding_attention" else 0.5 for lt in layer_types] == partial_rotary_factors
|
||||
head_arr = [n_head_swa if lt == "sliding_attention" else n_head_base for lt in layer_types]
|
||||
kv_arr = [n_kv_swa if lt == "sliding_attention" else n_kv_base for lt in layer_types]
|
||||
swa_pat = [lt == "sliding_attention" for lt in layer_types]
|
||||
|
||||
self.gguf_writer.add_head_count(head_arr)
|
||||
self.gguf_writer.add_head_count_kv(kv_arr)
|
||||
|
||||
self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
|
||||
self.gguf_writer.add_sliding_window_pattern(swa_pat)
|
||||
|
||||
self.gguf_writer.add_value_length(self.hparams["head_dim"])
|
||||
|
||||
# MoE params
|
||||
self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"])
|
||||
self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
|
||||
self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
|
||||
self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["share_expert_dim"])
|
||||
|
||||
if (moe_router_scaling_factor := self.hparams.get("moe_router_scaling_factor")) is not None:
|
||||
self.gguf_writer.add_expert_weights_scale(moe_router_scaling_factor)
|
||||
if (norm_expert_weight := self.hparams.get("norm_expert_weight")) is not None:
|
||||
self.gguf_writer.add_expert_weights_norm(norm_expert_weight)
|
||||
|
||||
# leading dense blocks
|
||||
leading_dense = 0
|
||||
moe_layers_enum = self.hparams.get("moe_layers_enum")
|
||||
if isinstance(moe_layers_enum, str) and moe_layers_enum.strip():
|
||||
moe_layers = sorted(int(i) for i in moe_layers_enum.strip().split(","))
|
||||
if moe_layers:
|
||||
leading_dense = max(0, moe_layers[0])
|
||||
self.gguf_writer.add_leading_dense_block_count(leading_dense)
|
||||
self.gguf_writer.add_moe_every_n_layers(int(self.hparams.get("moe_every_n_layer", 1)))
|
||||
|
||||
self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-5))
|
||||
|
||||
# Optional per-layer SwiGLU clamps.
|
||||
if (limits := self.hparams.get("swiglu_limits")) is not None:
|
||||
limits_f = [0.0 if v is None else float(v) for v in limits[: self.block_count]]
|
||||
self.gguf_writer.add_swiglu_clamp_exp(limits_f)
|
||||
if (limits_shared := self.hparams.get("swiglu_limits_shared")) is not None:
|
||||
limits_shared_f = [0.0 if v is None else float(v) for v in limits_shared[: self.block_count]]
|
||||
self.gguf_writer.add_swiglu_clamp_shexp(limits_shared_f)
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
|
||||
# remove mtp layers
|
||||
if (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None:
|
||||
il = int(m.group(1))
|
||||
n_main = int(self.hparams.get("num_hidden_layers", self.block_count))
|
||||
if il >= n_main:
|
||||
return
|
||||
if name.endswith("norm.weight"):
|
||||
data_torch += 1.0
|
||||
# Map router bias (expert selection bias) to a GGUF bias tensor
|
||||
if name.endswith(".moe.router_bias"):
|
||||
name += ".bias"
|
||||
|
||||
if name.endswith((".self_attn.g_proj.weight", ".moe.gate.weight", ".moe.up_proj.weight", ".moe.gate_proj.weight", ".moe.down_proj.weight")):
|
||||
data_torch = data_torch.squeeze().contiguous()
|
||||
|
||||
yield from super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
|
||||
# Step35 can optionally use Llama-3 style RoPE scaling (HF: rope_scaling.rope_type == "llama3").
|
||||
# llama.cpp represents this via a single extra tensor: "rope_freqs.weight" (aka MODEL_TENSOR.ROPE_FREQS).
|
||||
rope_params = self.rope_parameters.get("full_attention", self.rope_parameters)
|
||||
rope_type = rope_params.get("rope_type") or ""
|
||||
if rope_type.lower() != "llama3":
|
||||
return
|
||||
|
||||
# Step35 configs can carry per-layer rope_theta as a list; for llama3 rope factors we use the base value.
|
||||
rope_theta = self.hparams.get("rope_theta", 10000.0)
|
||||
if isinstance(rope_theta, list):
|
||||
rope_theta = rope_theta[0]
|
||||
base = float(rope_theta)
|
||||
if (dim := self.hparams.get("head_dim")) is None:
|
||||
dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
|
||||
dim = int(dim)
|
||||
|
||||
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
|
||||
|
||||
factor = float(rope_params.get("factor", 8.0))
|
||||
low_freq_factor = float(rope_params.get("low_freq_factor", 1.0))
|
||||
high_freq_factor = float(rope_params.get("high_freq_factor", 4.0))
|
||||
old_context_len = int(rope_params.get("original_max_position_embeddings", self.hparams.get("original_max_position_embeddings", 8192)))
|
||||
|
||||
low_freq_wavelen = old_context_len / low_freq_factor
|
||||
high_freq_wavelen = old_context_len / high_freq_factor
|
||||
|
||||
rope_factors: list[float] = []
|
||||
for freq in freqs:
|
||||
wavelen = 2 * math.pi / float(freq)
|
||||
if wavelen < high_freq_wavelen:
|
||||
rope_factors.append(1.0)
|
||||
elif wavelen > low_freq_wavelen:
|
||||
rope_factors.append(factor)
|
||||
else:
|
||||
smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
|
||||
rope_factors.append(1.0 / ((1.0 - smooth) / factor + smooth))
|
||||
|
||||
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
|
||||
|
||||
|
||||
@ModelBase.register("PanguEmbeddedForCausalLM")
|
||||
class PanguEmbeddedModel(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.PANGU_EMBED
|
||||
|
||||
@@ -148,6 +148,7 @@ models = [
|
||||
{"name": "youtu", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
|
||||
{"name": "solar-open", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
|
||||
{"name": "exaone-moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
|
||||
{"name": "qwen35", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", }
|
||||
]
|
||||
|
||||
# some models are known to be broken upstream, so we will skip them as exceptions
|
||||
|
||||
180
docs/backend/VirtGPU.md
Normal file
180
docs/backend/VirtGPU.md
Normal file
@@ -0,0 +1,180 @@
|
||||
# GGML-VirtGPU Backend
|
||||
|
||||
The GGML-VirtGPU backend enables GGML applications to run machine
|
||||
learning computations on host hardware while the application itself
|
||||
runs inside a virtual machine. It uses host-guest shared memory to
|
||||
efficiently share data buffers between the two sides.
|
||||
|
||||
This backend relies on the virtio-gpu, and VirglRenderer API Remoting
|
||||
(APIR) component. The backend is split into two libraries:
|
||||
- a GGML implementation (the "remoting frontend"), running in the
|
||||
guest and interacting with the virtgpu device
|
||||
- a VirglRenderer APIR compatible library (the "remoting backend"),
|
||||
running in the host and interacting with Virglrenderer and an actual
|
||||
GGML device backend.
|
||||
|
||||
## OS support
|
||||
|
||||
| OS | Status | Backend | CI testing | Notes
|
||||
| -------- | ----------------- | ----------- | ----------- | -----
|
||||
| MacOS 14 | Supported | ggml-metal | X | Working when compiled on MacOS 14
|
||||
| MacOS 15 | Supported | ggml-metal | X | Working when compiled on MacOS 14 or MacOS 15
|
||||
| MacOS 26 | Not tested | | |
|
||||
| Linux | Under development | ggml-vulkan | not working | Working locally, CI running into deadlocks
|
||||
|
||||
|
||||
## Architecture Overview
|
||||
|
||||
The GGML-VirtGPU backend consists of three main components:
|
||||
|
||||
```mermaid
|
||||
graph TD
|
||||
%% Nodes
|
||||
|
||||
subgraph GuestVM ["Guest VM - Frontend"]
|
||||
App([GGML Application<br/>llama.cpp, etc.])
|
||||
|
||||
direction TB
|
||||
Interface[GGML Backend Interface]
|
||||
Comm["GGML-VirtGPU<br/>(hypercalls + shared mem)"]
|
||||
|
||||
App --> Interface
|
||||
Interface --> Comm
|
||||
end
|
||||
|
||||
API[virtio-gpu / virglrenderer API]
|
||||
|
||||
subgraph HostSystem [Host System - Backend]
|
||||
direction TB
|
||||
Dispatcher[GGML-VirtGPU-Backend]
|
||||
BackendLib[GGML Backend library<br/>Metal / Vulkan / CPU / ...]
|
||||
|
||||
Dispatcher --> BackendLib
|
||||
end
|
||||
|
||||
%% Connections
|
||||
Comm --> API
|
||||
API --> HostSystem
|
||||
```
|
||||
|
||||
### Key Components
|
||||
|
||||
1. **Guest-side Frontend** (`ggml-virtgpu/`): Implements the GGML backend interface and forwards operations to the host
|
||||
2. **Host-side Backend** (`ggml-virtgpu/backend/`): Receives forwarded operations and executes them on actual hardware backends
|
||||
3. **Communication Layer**: Uses virtio-gpu hypercalls and shared memory for efficient data transfer
|
||||
|
||||
## Features
|
||||
|
||||
- **Dynamic backend loading** on the host side (CPU, CUDA, Metal, etc.)
|
||||
- **Zero-copy data transfer** via host-guest shared memory pages
|
||||
|
||||
## Communication Protocol
|
||||
|
||||
### Hypercalls and Shared Memory
|
||||
|
||||
The backend uses two primary communication mechanisms:
|
||||
|
||||
1. **Hypercalls (`DRM_IOCTL_VIRTGPU_EXECBUFFER`)**: Trigger remote execution from guest to host
|
||||
2. **Shared Memory Pages**: Zero-copy data transfer for tensors and parameters
|
||||
|
||||
#### Shared Memory Layout
|
||||
|
||||
Each connection uses two shared memory buffers:
|
||||
|
||||
- **Data Buffer** (24 MiB): For command/response data and tensor transfers
|
||||
- **Reply Buffer** (16 KiB): For command replies and status information
|
||||
- **Data Buffers**: Dynamically allocated host-guest shared buffers
|
||||
served as GGML buffers.
|
||||
|
||||
### APIR Protocol
|
||||
|
||||
The Virglrender API Remoting protocol defines three command types:
|
||||
|
||||
- `HANDSHAKE`: Protocol version negotiation and capability discovery
|
||||
- `LOADLIBRARY`: Dynamic loading of backend libraries on the host
|
||||
- `FORWARD`: API function call forwarding
|
||||
|
||||
### Binary Serialization
|
||||
|
||||
Commands and data are serialized using a custom binary protocol with:
|
||||
|
||||
- Fixed-size encoding for basic types
|
||||
- Variable-length arrays with size prefixes
|
||||
- Buffer bounds checking
|
||||
- Error recovery mechanisms
|
||||
|
||||
## Supported Operations
|
||||
|
||||
### Device Operations
|
||||
- Device enumeration and capability queries
|
||||
- Memory information (total/free)
|
||||
- Backend type detection
|
||||
|
||||
### Buffer Operations
|
||||
- Buffer allocation and deallocation
|
||||
- Tensor data transfer (host ↔ guest)
|
||||
- Memory copying and clearing
|
||||
|
||||
### Computation Operations
|
||||
- Graph execution forwarding
|
||||
|
||||
## Build Requirements
|
||||
|
||||
### Guest-side Dependencies
|
||||
- `libdrm` for DRM/virtio-gpu communication
|
||||
- C++20 compatible compiler
|
||||
- CMake 3.14+
|
||||
|
||||
### Host-side Dependencies
|
||||
- virglrenderer with APIR support (pending upstream review)
|
||||
- Target backend libraries (libggml-metal, libggml-vulkan, etc.)
|
||||
|
||||
## Configuration
|
||||
|
||||
### Environment Variables
|
||||
|
||||
- `GGML_VIRTGPU_BACKEND_LIBRARY`: Path to the host-side backend library
|
||||
- `GGML_VIRTGPU_DEBUG`: Enable debug logging
|
||||
|
||||
### Build Options
|
||||
|
||||
- `GGML_VIRTGPU`: Enable the VirtGPU backend (`ON` or `OFF`, default: `OFF`)
|
||||
- `GGML_VIRTGPU_BACKEND`: Build the host-side backend component (`ON`, `OFF` or `ONLY`, default: `OFF`)
|
||||
|
||||
### System Requirements
|
||||
|
||||
- VM with virtio-gpu support
|
||||
- VirglRenderer with APIR patches
|
||||
- Compatible backend libraries on host
|
||||
|
||||
## Limitations
|
||||
|
||||
- **VM-specific**: Only works in virtual machines with virtio-gpu support
|
||||
- **Host dependency**: Requires properly configured host-side backend
|
||||
- **Latency**: Small overhead from VM escaping for each operation
|
||||
|
||||
|
||||
* This work is pending upstream changes in the VirglRenderer
|
||||
project.
|
||||
* The backend can be tested with Virglrenderer compiled from source
|
||||
using this PR:
|
||||
https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590
|
||||
* This work is pending changes in the VMM/hypervisor running the
|
||||
virtual machine, which need to know how to route the newly
|
||||
introduced APIR capset.
|
||||
* The environment variable `VIRGL_ROUTE_VENUS_TO_APIR=1` allows
|
||||
using the Venus capset, until the relevant hypervisors have been
|
||||
patched. However, setting this flag breaks the Vulkan/Venus normal
|
||||
behavior.
|
||||
* The environment variable `GGML_REMOTING_USE_APIR_CAPSET` tells the
|
||||
`ggml-virtgpu` backend to use the APIR capset. This will become
|
||||
the default when the relevant hypervisors have been patched.
|
||||
|
||||
* This work focused on improving the performance of llama.cpp running
|
||||
on MacOS containers, and is mainly tested on this platform. The
|
||||
linux support (via `krun`) is in progress.
|
||||
|
||||
## See Also
|
||||
|
||||
- [Development and Testing](VirtGPU/development.md)
|
||||
- [Backend configuration](VirtGPU/configuration.md)
|
||||
174
docs/backend/VirtGPU/configuration.md
Normal file
174
docs/backend/VirtGPU/configuration.md
Normal file
@@ -0,0 +1,174 @@
|
||||
# GGML-VirtGPU Backend Configuration
|
||||
|
||||
This document describes the environment variables used by the ggml-virtgpu backend system, covering both the frontend (guest-side) and backend (host-side) components.
|
||||
|
||||
## Environment Variables Overview
|
||||
|
||||
The ggml-virtgpu backend uses environment variables for configuration across three main components:
|
||||
- **Frontend (Guest)**: GGML applications running in VMs
|
||||
- **Hypervisor**: Virglrenderer/APIR system
|
||||
- **Backend (Host)**: Host-side GGML backend integration
|
||||
|
||||
## Frontend (Guest-side) Configuration
|
||||
|
||||
### GGML_REMOTING_USE_APIR_CAPSET
|
||||
- **Location**: `ggml/src/ggml-virtgpu/virtgpu.cpp`
|
||||
- **Type**: Boolean flag (presence-based)
|
||||
- **Purpose**: Controls which virtio-gpu capability set to use for communication
|
||||
- **Values**:
|
||||
- Set (any value): Use the APIR capset (long-term setup)
|
||||
- Unset: Use the Venus capset (easier for testing with an unmodified hypervisor)
|
||||
- **Default**: Unset (Venus capset)
|
||||
- **Usage**:
|
||||
```bash
|
||||
export GGML_REMOTING_USE_APIR_CAPSET=1 # Use APIR capset
|
||||
# or leave unset for Venus capset
|
||||
```
|
||||
|
||||
## Hypervisor (Virglrenderer/APIR) Configuration
|
||||
|
||||
These environment variables are used during the transition phase for
|
||||
running with an unmodified hypervisor (not supporting the
|
||||
VirglRenderer APIR component). They will be removed in the future, and
|
||||
the hypervisor will instead configure VirglRenderer with the APIR
|
||||
_Configuration Key_.
|
||||
|
||||
### VIRGL_APIR_BACKEND_LIBRARY
|
||||
- **Location**: `virglrenderer/src/apir/apir-context.c`
|
||||
- **Configuration Key**: `apir.load_library.path`
|
||||
- **Type**: File path string
|
||||
- **Purpose**: Path to the APIR backend library that virglrenderer should dynamically load
|
||||
- **Required**: Yes
|
||||
- **Example**:
|
||||
```bash
|
||||
export VIRGL_APIR_BACKEND_LIBRARY="/path/to/libggml-remotingbackend.so"
|
||||
```
|
||||
|
||||
### VIRGL_ROUTE_VENUS_TO_APIR
|
||||
- **Location**: `virglrenderer/src/apir/apir-renderer.h`
|
||||
- **Type**: Boolean flag (presence-based)
|
||||
- **Purpose**: Temporary workaround to route Venus capset calls to APIR during hypervisor transition period
|
||||
- **Status**: will be removed once hypervisors support APIR natively
|
||||
- **Warning**: Breaks normal Vulkan/Venus functionality
|
||||
- **Usage**:
|
||||
```bash
|
||||
export VIRGL_ROUTE_VENUS_TO_APIR=1 # For testing with an unmodified hypervisor
|
||||
```
|
||||
|
||||
### VIRGL_APIR_LOG_TO_FILE
|
||||
- **Location**: `virglrenderer/src/apir/apir-renderer.c`
|
||||
- **Environment Variable**: `VIRGL_APIR_LOG_TO_FILE`
|
||||
- **Type**: File path string
|
||||
- **Purpose**: Enable debug logging from the VirglRenderer APIR component to specified file
|
||||
- **Required**: No (optional debugging)
|
||||
- **Default**: Logging to `stderr`
|
||||
- **Usage**:
|
||||
```bash
|
||||
export VIRGL_APIR_LOG_TO_FILE="/tmp/apir-debug.log"
|
||||
```
|
||||
|
||||
## Backend (Host-side) Configuration
|
||||
|
||||
These environment variables are used during the transition phase for
|
||||
running with an unmodified hypervisor (not supporting the
|
||||
VirglRenderer APIR component). They will be removed in the future, and
|
||||
the hypervisor will instead configure VirglRenderer with the APIR
|
||||
_Configuration Key_.
|
||||
|
||||
### APIR_LLAMA_CPP_GGML_LIBRARY_PATH
|
||||
- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
|
||||
- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
|
||||
- **Configuration Key**: `ggml.library.path`
|
||||
- **Type**: File path string
|
||||
- **Purpose**: Path to the actual GGML backend library (Metal, CUDA, Vulkan, etc.)
|
||||
- **Required**: **Yes** - backend initialization fails without this
|
||||
- **Examples**:
|
||||
```bash
|
||||
# macOS with Metal backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
|
||||
|
||||
# Linux with CUDA backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-cuda.so"
|
||||
|
||||
# macOS or Linux with Vulkan backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-vulkan.so"
|
||||
```
|
||||
|
||||
### APIR_LLAMA_CPP_GGML_LIBRARY_REG
|
||||
- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
|
||||
- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
|
||||
- **Configuration Key**: `ggml.library.reg`
|
||||
- **Type**: Function symbol name string
|
||||
- **Purpose**: Name of the backend registration function to call after loading the library
|
||||
- **Required**: No (defaults to `ggml_backend_init`)
|
||||
- **Default**: `ggml_backend_init`
|
||||
- **Examples**:
|
||||
```bash
|
||||
# Metal backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
|
||||
|
||||
# CUDA backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_cuda_reg"
|
||||
|
||||
# Vulkan backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_vulkan_reg"
|
||||
|
||||
# Generic fallback (default)
|
||||
# export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_init"
|
||||
```
|
||||
|
||||
### APIR_LLAMA_CPP_LOG_TO_FILE
|
||||
- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp:62`
|
||||
- **Environment Variable**: `APIR_LLAMA_CPP_LOG_TO_FILE`
|
||||
- **Type**: File path string
|
||||
- **Purpose**: Enable debug logging from the GGML backend to specified file
|
||||
- **Required**: No (optional debugging)
|
||||
- **Usage**:
|
||||
```bash
|
||||
export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml-backend-debug.log"
|
||||
```
|
||||
|
||||
## Configuration Flow
|
||||
|
||||
The configuration system works as follows:
|
||||
|
||||
1. **Hypervisor Setup**: Virglrenderer loads the APIR backend library specified by `VIRGL_APIR_BACKEND_LIBRARY`
|
||||
|
||||
2. **Context Creation**: When an APIR context is created, it populates a configuration table with environment variables:
|
||||
- `apir.load_library.path` ← `VIRGL_APIR_BACKEND_LIBRARY`
|
||||
- `ggml.library.path` ← `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
|
||||
- `ggml.library.reg` ← `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
|
||||
- this step will eventually be performed by the hypervisor itself, with command-line arguments instead of environment variables.
|
||||
|
||||
3. **Backend Initialization**: The backend queries the configuration via callbacks:
|
||||
- `virgl_cbs->get_config(ctx_id, "ggml.library.path")` returns the library path
|
||||
- `virgl_cbs->get_config(ctx_id, "ggml.library.reg")` returns the registration function
|
||||
|
||||
4. **Library Loading**: The backend dynamically loads and initializes the specified GGML library
|
||||
|
||||
## Error Messages
|
||||
|
||||
Common error scenarios and their messages:
|
||||
|
||||
- **Missing library path**: `"cannot open the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_PATH' not defined"`
|
||||
- **Missing registration function**: `"cannot register the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_REG' not defined"`
|
||||
|
||||
## Example Complete Configuration
|
||||
|
||||
Here's an example configuration for a macOS host with Metal backend:
|
||||
|
||||
```bash
|
||||
# Hypervisor environment
|
||||
export VIRGL_APIR_BACKEND_LIBRARY="/opt/llama.cpp/lib/libggml-virtgpu-backend.dylib"
|
||||
|
||||
# Backend configuration
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
|
||||
|
||||
# Optional logging
|
||||
export VIRGL_APIR_LOG_TO_FILE="/tmp/apir.log"
|
||||
export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml.log"
|
||||
|
||||
# Guest configuration
|
||||
export GGML_REMOTING_USE_APIR_CAPSET=1
|
||||
```
|
||||
220
docs/backend/VirtGPU/development.md
Normal file
220
docs/backend/VirtGPU/development.md
Normal file
@@ -0,0 +1,220 @@
|
||||
# Development and Testing
|
||||
|
||||
## Development
|
||||
|
||||
### Code Generation
|
||||
|
||||
The backend uses code generation from YAML configuration:
|
||||
|
||||
```bash
|
||||
# Regenerate protocol code
|
||||
cd ggml-virtgpu/
|
||||
python regenerate_remoting.py
|
||||
```
|
||||
|
||||
### Adding New Operations
|
||||
|
||||
1. Add function definition to `ggmlremoting_functions.yaml`
|
||||
2. Regenerate code with `regenerate_remoting.py`
|
||||
3. Implement guest-side forwarding in `virtgpu-forward-*.cpp`
|
||||
4. Implement host-side handling in `backend-dispatched-*.cpp`
|
||||
|
||||
## Testing
|
||||
|
||||
This document provides instructions for building and testing the GGML-VirtGPU backend on macOS with containers.
|
||||
|
||||
### Prerequisites
|
||||
|
||||
The testing setup requires:
|
||||
|
||||
- macOS host system
|
||||
- Container runtime with `libkrun` provider (podman machine)
|
||||
- Access to development patchset for VirglRenderer
|
||||
|
||||
### Required Patchsets
|
||||
|
||||
The backend requires patches that are currently under review:
|
||||
|
||||
- **Virglrenderer APIR upstream PR**: https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590 (for reference)
|
||||
- **MacOS Virglrenderer (for krunkit)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-macos
|
||||
- **Linux Virglrenderer (for krun)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-linux
|
||||
|
||||
### Build Instructions
|
||||
|
||||
#### 1. Build ggml-virtgpu-backend (Host-side, macOS)
|
||||
|
||||
```bash
|
||||
# Build the backend that runs natively on macOS
|
||||
mkdir llama.cpp
|
||||
cd llama.cpp
|
||||
git clone https://github.com/ggml-org/llama.cpp.git src
|
||||
cd src
|
||||
|
||||
LLAMA_MAC_BUILD=$PWD/build/ggml-virtgpu-backend
|
||||
|
||||
cmake -S . -B $LLAMA_MAC_BUILD \
|
||||
-DGGML_NATIVE=OFF \
|
||||
-DLLAMA_CURL=ON \
|
||||
-DGGML_REMOTINGBACKEND=ONLY \
|
||||
-DGGML_METAL=ON
|
||||
|
||||
TARGETS="ggml-metal"
|
||||
cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $TARGETS
|
||||
|
||||
# Build additional tools for native benchmarking
|
||||
EXTRA_TARGETS="llama-run llama-bench"
|
||||
cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $EXTRA_TARGETS
|
||||
```
|
||||
|
||||
#### 2. Build virglrenderer (Host-side, macOS)
|
||||
|
||||
```bash
|
||||
# Build virglrenderer with APIR support
|
||||
mkdir virglrenderer
|
||||
git clone https://gitlab.freedesktop.org/kpouget/virglrenderer -b main-macos src
|
||||
cd src
|
||||
|
||||
VIRGL_BUILD_DIR=$PWD/build
|
||||
|
||||
# -Dvenus=true and VIRGL_ROUTE_VENUS_TO_APIR=1 route the APIR requests via the Venus backend, for easier testing without a patched hypervisor
|
||||
|
||||
meson setup $VIRGL_BUILD_DIR \
|
||||
-Dvenus=true \
|
||||
-Dapir=true
|
||||
|
||||
ninja -C $VIRGL_BUILD_DIR
|
||||
```
|
||||
|
||||
#### 3. Build ggml-virtgpu (Guest-side, Linux)
|
||||
|
||||
Option A: Build from a script:
|
||||
|
||||
```bash
|
||||
# Inside a Linux container
|
||||
mkdir llama.cpp
|
||||
git clone https://github.com/ggml-org/llama.cpp.git src
|
||||
cd src
|
||||
|
||||
LLAMA_LINUX_BUILD=$PWD//build-virtgpu
|
||||
|
||||
cmake -S . -B $LLAMA_LINUX_BUILD \
|
||||
-DGGML_VIRTGPU=ON
|
||||
|
||||
ninja -C $LLAMA_LINUX_BUILD
|
||||
```
|
||||
|
||||
Option B: Build container image with frontend:
|
||||
|
||||
```bash
|
||||
cat << EOF > remoting.containerfile
|
||||
FROM quay.io/fedora/fedora:43
|
||||
USER 0
|
||||
|
||||
WORKDIR /app/remoting
|
||||
|
||||
ARG LLAMA_CPP_REPO="https://github.com/ggml-org/llama.cpp.git"
|
||||
ARG LLAMA_CPP_VERSION="master"
|
||||
ARG LLAMA_CPP_CMAKE_FLAGS="-DGGML_VIRTGPU=ON"
|
||||
ARG LLAMA_CPP_CMAKE_BUILD_FLAGS="--parallel 4"
|
||||
|
||||
RUN dnf install -y git cmake gcc gcc-c++ libcurl-devel libdrm-devel
|
||||
|
||||
RUN git clone "\${LLAMA_CPP_REPO}" src \\
|
||||
&& git -C src fetch origin \${LLAMA_CPP_VERSION} \\
|
||||
&& git -C src reset --hard FETCH_HEAD
|
||||
|
||||
RUN mkdir -p build \\
|
||||
&& cd src \\
|
||||
&& set -o pipefail \\
|
||||
&& cmake -S . -B ../build \${LLAMA_CPP_CMAKE_FLAGS} \\
|
||||
&& cmake --build ../build/ \${LLAMA_CPP_CMAKE_BUILD_FLAGS}
|
||||
|
||||
ENTRYPOINT ["/app/remoting/src/build/bin/llama-server"]
|
||||
EOF
|
||||
|
||||
mkdir -p empty_dir
|
||||
podman build -f remoting.containerfile ./empty_dir -t localhost/llama-cpp.virtgpu
|
||||
```
|
||||
|
||||
### Environment Setup
|
||||
|
||||
#### Set krunkit Environment Variables
|
||||
|
||||
```bash
|
||||
# Define the base directories (adapt these paths to your system)
|
||||
VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build
|
||||
LLAMA_MAC_BUILD=$HOME/remoting/llama.cpp/build-backend
|
||||
|
||||
# For krunkit to load the custom virglrenderer library
|
||||
export DYLD_LIBRARY_PATH=$VIRGL_BUILD_DIR/src
|
||||
|
||||
# For Virglrenderer to load the ggml-remotingbackend library
|
||||
export VIRGL_APIR_BACKEND_LIBRARY="$LLAMA_MAC_BUILD/bin/libggml-virtgpu-backend.dylib"
|
||||
|
||||
# For llama.cpp remotingbackend to load the ggml-metal backend
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="$LLAMA_MAC_BUILD/bin/libggml-metal.dylib"
|
||||
export APIR_LLAMA_CPP_GGML_LIBRARY_REG=ggml_backend_metal_reg
|
||||
```
|
||||
|
||||
#### Launch Container Environment
|
||||
|
||||
```bash
|
||||
# Set container provider to libkrun
|
||||
export CONTAINERS_MACHINE_PROVIDER=libkrun
|
||||
podman machine start
|
||||
```
|
||||
|
||||
#### Verify Environment
|
||||
|
||||
Confirm that krunkit is using the correct virglrenderer library:
|
||||
|
||||
```bash
|
||||
lsof -c krunkit | grep virglrenderer
|
||||
# Expected output:
|
||||
# krunkit 50574 user txt REG 1,14 2273912 10849442 ($VIRGL_BUILD_DIR/src)/libvirglrenderer.1.dylib
|
||||
```
|
||||
|
||||
### Running Tests
|
||||
|
||||
#### Launch Test Container
|
||||
|
||||
```bash
|
||||
# Optional model caching
|
||||
mkdir -p models
|
||||
PODMAN_CACHE_ARGS="-v models:/models --user root:root --cgroupns host --security-opt label=disable -w /models"
|
||||
|
||||
podman run $PODMAN_CACHE_ARGS -it --rm --device /dev/dri localhost/llama-cpp.virtgpu
|
||||
```
|
||||
|
||||
#### Test llama.cpp in Container
|
||||
|
||||
```bash
|
||||
|
||||
# Run performance benchmark
|
||||
/app/remoting/build/bin/llama-bench -m ./llama3.2
|
||||
```
|
||||
|
||||
Expected output (performance may vary):
|
||||
```
|
||||
| model | size | params | backend | ngl | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
|
||||
| llama 3B Q4_K - Medium | 1.87 GiB | 3.21 B | ggml-virtgpu | 99 | pp512 | 991.30 ± 0.66 |
|
||||
| llama 3B Q4_K - Medium | 1.87 GiB | 3.21 B | ggml-virtgpu | 99 | tg128 | 85.71 ± 0.11 |
|
||||
```
|
||||
|
||||
### Troubleshooting
|
||||
|
||||
#### SSH Environment Variable Issues
|
||||
|
||||
⚠️ **Warning**: Setting `DYLD_LIBRARY_PATH` from SSH doesn't work on macOS. Here is a workaround:
|
||||
|
||||
**Workaround 1: Replace system library**
|
||||
```bash
|
||||
VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build # ⚠️ adapt to your system
|
||||
BREW_VIRGL_DIR=/opt/homebrew/Cellar/virglrenderer/0.10.4d/lib
|
||||
VIRGL_LIB=libvirglrenderer.1.dylib
|
||||
|
||||
cd $BREW_VIRGL_DIR
|
||||
mv $VIRGL_LIB ${VIRGL_LIB}.orig
|
||||
ln -s $VIRGL_BUILD_DIR/src/$VIRGL_LIB
|
||||
```
|
||||
@@ -22,7 +22,7 @@ Legend:
|
||||
| ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
|
||||
| ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
|
||||
| ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
|
||||
| CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
|
||||
| CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
|
||||
| CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
|
||||
| CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
|
||||
| CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
|
||||
|
||||
@@ -77,8 +77,8 @@
|
||||
"SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
|
||||
"SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
"SYCL0","FLOOR","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
|
||||
"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
|
||||
"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
|
||||
"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
|
||||
"SYCL0","ROUND","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
"SYCL0","ROUND","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
|
||||
"SYCL0","TRUNC","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
@@ -161,8 +161,8 @@
|
||||
"SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
|
||||
"SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
"SYCL0","FLOOR","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
|
||||
"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
|
||||
"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
|
||||
"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
|
||||
"SYCL0","ROUND","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
"SYCL0","ROUND","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
|
||||
"SYCL0","TRUNC","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
|
||||
|
||||
|
Can't render this file because it is too large.
|
@@ -119,8 +119,6 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
|
||||
of lookup n-gram (default: 12)
|
||||
--spec-ngram-size-m N ngram size M for ngram-simple/ngram-map speculative decoding, length
|
||||
of draft m-gram (default: 48)
|
||||
--spec-ngram-check-rate N ngram check rate for ngram-simple/ngram-map speculative decoding
|
||||
(default: 1)
|
||||
--spec-ngram-min-hits N minimum hits for ngram-map speculative decoding (default: 1)
|
||||
```
|
||||
|
||||
@@ -153,10 +151,6 @@ Sets the size M of the draft m-gram for n-gram map based speculative decoding.
|
||||
The m-gram size determines how many tokens to draft when a match is found.
|
||||
Larger values can provide more speedup but may reduce acceptance rate.
|
||||
|
||||
### `--spec-ngram-check-rate R`
|
||||
|
||||
This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
|
||||
|
||||
### `--spec-ngram-min-hits H`
|
||||
|
||||
This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
|
||||
@@ -175,7 +169,12 @@ draft acceptance rate = 0.70312 ( 90 accepted / 128 generated)
|
||||
statistics ngram_mod: #calls = 810, #gen drafts = 15, #acc drafts = 15, #gen tokens = 960, #acc tokens = 730, dur(b,g,a) = 0.149, 0.347, 0.005 ms
|
||||
```
|
||||
|
||||
- `#calls`: number of calls of this implementations
|
||||
```
|
||||
statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts = 26, #gen tokens = 1248, #acc tokens = 968, dur(b,g,a) = 2.234, 1.427, 0.016 ms
|
||||
```
|
||||
|
||||
|
||||
- `#calls(b,g,a)`: number of calls of begin (new prompt), generation and accumulation of this implementations
|
||||
- `#gen drafts`: number of drafts generated by this implementation
|
||||
- `#acc drafts`: number of drafts accepted (partially) by the main model
|
||||
- `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)
|
||||
|
||||
159
examples/model-conversion/scripts/utils/tensor-info.py
Executable file
159
examples/model-conversion/scripts/utils/tensor-info.py
Executable file
@@ -0,0 +1,159 @@
|
||||
#!/usr/bin/env python3
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
from safetensors import safe_open
|
||||
|
||||
|
||||
MODEL_SAFETENSORS_FILE = "model.safetensors"
|
||||
MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
|
||||
|
||||
|
||||
def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
|
||||
index_file = model_path / MODEL_SAFETENSORS_INDEX
|
||||
|
||||
if index_file.exists():
|
||||
with open(index_file, 'r') as f:
|
||||
index = json.load(f)
|
||||
return index.get("weight_map", {})
|
||||
|
||||
return None
|
||||
|
||||
|
||||
def get_all_tensor_names(model_path: Path) -> list[str]:
|
||||
weight_map = get_weight_map(model_path)
|
||||
|
||||
if weight_map is not None:
|
||||
return list(weight_map.keys())
|
||||
|
||||
single_file = model_path / MODEL_SAFETENSORS_FILE
|
||||
if single_file.exists():
|
||||
try:
|
||||
with safe_open(single_file, framework="pt", device="cpu") as f:
|
||||
return list(f.keys())
|
||||
except Exception as e:
|
||||
print(f"Error reading {single_file}: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
print(f"Error: No safetensors files found in {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
|
||||
def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
|
||||
weight_map = get_weight_map(model_path)
|
||||
|
||||
if weight_map is not None:
|
||||
return weight_map.get(tensor_name)
|
||||
|
||||
single_file = model_path / MODEL_SAFETENSORS_FILE
|
||||
if single_file.exists():
|
||||
return single_file.name
|
||||
|
||||
return None
|
||||
|
||||
|
||||
def normalize_tensor_name(tensor_name: str) -> str:
|
||||
normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
|
||||
normalized = re.sub(r'\.\d+$', '.#', normalized)
|
||||
return normalized
|
||||
|
||||
|
||||
def list_all_tensors(model_path: Path, unique: bool = False):
|
||||
tensor_names = get_all_tensor_names(model_path)
|
||||
|
||||
if unique:
|
||||
seen = set()
|
||||
for tensor_name in sorted(tensor_names):
|
||||
normalized = normalize_tensor_name(tensor_name)
|
||||
if normalized not in seen:
|
||||
seen.add(normalized)
|
||||
print(normalized)
|
||||
else:
|
||||
for tensor_name in sorted(tensor_names):
|
||||
print(tensor_name)
|
||||
|
||||
|
||||
def print_tensor_info(model_path: Path, tensor_name: str):
|
||||
tensor_file = find_tensor_file(model_path, tensor_name)
|
||||
|
||||
if tensor_file is None:
|
||||
print(f"Error: Could not find tensor '{tensor_name}' in model index")
|
||||
print(f"Model path: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
file_path = model_path / tensor_file
|
||||
|
||||
try:
|
||||
with safe_open(file_path, framework="pt", device="cpu") as f:
|
||||
if tensor_name in f.keys():
|
||||
tensor_slice = f.get_slice(tensor_name)
|
||||
shape = tensor_slice.get_shape()
|
||||
print(f"Tensor: {tensor_name}")
|
||||
print(f"File: {tensor_file}")
|
||||
print(f"Shape: {shape}")
|
||||
else:
|
||||
print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
|
||||
sys.exit(1)
|
||||
|
||||
except FileNotFoundError:
|
||||
print(f"Error: The file '{file_path}' was not found.")
|
||||
sys.exit(1)
|
||||
except Exception as e:
|
||||
print(f"An error occurred: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Print tensor information from a safetensors model"
|
||||
)
|
||||
parser.add_argument(
|
||||
"tensor_name",
|
||||
nargs="?", # optional (if --list is used for example)
|
||||
help="Name of the tensor to inspect"
|
||||
)
|
||||
parser.add_argument(
|
||||
"-m", "--model-path",
|
||||
type=Path,
|
||||
help="Path to the model directory (default: MODEL_PATH environment variable)"
|
||||
)
|
||||
parser.add_argument(
|
||||
"-l", "--list",
|
||||
action="store_true",
|
||||
help="List unique tensor patterns in the model (layer numbers replaced with #)"
|
||||
)
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
model_path = args.model_path
|
||||
if model_path is None:
|
||||
model_path_str = os.environ.get("MODEL_PATH")
|
||||
if model_path_str is None:
|
||||
print("Error: --model-path not provided and MODEL_PATH environment variable not set")
|
||||
sys.exit(1)
|
||||
model_path = Path(model_path_str)
|
||||
|
||||
if not model_path.exists():
|
||||
print(f"Error: Model path does not exist: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
if not model_path.is_dir():
|
||||
print(f"Error: Model path is not a directory: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
if args.list:
|
||||
list_all_tensors(model_path, unique=True)
|
||||
else:
|
||||
if args.tensor_name is None:
|
||||
print("Error: tensor_name is required when not using --list")
|
||||
sys.exit(1)
|
||||
print_tensor_info(model_path, args.tensor_name)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -7,8 +7,6 @@
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
|
||||
|
||||
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();
|
||||
|
||||
#ifdef __cplusplus
|
||||
|
||||
@@ -471,9 +471,10 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
|
||||
|
||||
int best_score = 0;
|
||||
fs::path best_path;
|
||||
std::error_code ec;
|
||||
|
||||
for (const auto & search_path : search_paths) {
|
||||
if (std::error_code ec; !fs::exists(search_path, ec)) {
|
||||
if (!fs::exists(search_path, ec)) {
|
||||
if (ec) {
|
||||
GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
|
||||
} else {
|
||||
@@ -483,7 +484,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
|
||||
}
|
||||
fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
|
||||
for (const auto & entry : dir_it) {
|
||||
if (entry.is_regular_file()) {
|
||||
if (entry.is_regular_file(ec)) {
|
||||
auto filename = entry.path().filename();
|
||||
auto ext = entry.path().extension();
|
||||
if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
|
||||
|
||||
@@ -3286,130 +3286,223 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context & ctx, ggml_tensor
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Performs expert-specific matrix multiplication (MoE) with
|
||||
* quantized precision using the CANN backend.
|
||||
* @brief Performs quantized matrix multiplication for Mixture of Experts (MoE)
|
||||
* models using the CANN backend.
|
||||
*
|
||||
* This function executes a matrix multiplication operation tailored for
|
||||
* Mixture of Experts (MoE) models, where the input tensor is multiplied
|
||||
* with expert-specific quantized weight matrices. It leverages the CANN
|
||||
* backend to perform efficient low-precision computations and stores the
|
||||
* quantized result in the destination tensor `dst`.
|
||||
* This function implements MUL_MAT_ID operation for quantized weight matrices
|
||||
* (Q4_0 and Q8_0 formats). It selects expert-specific weight matrices based on
|
||||
* the provided expert indices, and computes matrix multiplication using CANN's
|
||||
* WeightQuantBatchMatmulV2 operator.
|
||||
*
|
||||
* Quantization techniques reduce memory footprint and improve performance
|
||||
* by using lower-bit representations (e.g., int8) instead of floating-point.
|
||||
* This function is designed to work with such formats and may incorporate
|
||||
* optimizations like identity-based fast paths or routing masks for sparse
|
||||
* expert selection.
|
||||
* The function performs the following steps:
|
||||
* 1. Converts input/output tensors to F16 format if necessary
|
||||
* 2. Uses IndexSelect to extract expert-specific weights and scales based on indices
|
||||
* 3. Performs quantized matrix multiplication for each expert using WeightQuantBatchMatmulV2
|
||||
* 4. Converts output back to the target type if needed
|
||||
*
|
||||
* @param ctx The context for executing CANN backend operations.
|
||||
* @param dst The destination tensor where the quantized MoE multiplication result
|
||||
* will be stored.
|
||||
* Tensor shapes:
|
||||
* - dst: [M, K, N, 1] - output tensor
|
||||
* - src0: [D, M, A, 1] - quantized weight matrices (Q4_0 or Q8_0)
|
||||
* - src1: [D, B, N, 1] - input activations (B = K for per-expert input, or B = 1 for broadcast)
|
||||
* - ids: [K, N] - expert indices for routing
|
||||
*
|
||||
* @note This function assumes quantized data types and is designed for
|
||||
* MoE architectures with potential sparse expert routing.
|
||||
* @param ctx The CANN backend context for operation execution.
|
||||
* @param dst The destination tensor where the multiplication result will be stored.
|
||||
*
|
||||
* @note Only Q4_0 and Q8_0 quantization formats are supported.
|
||||
* @note The function handles automatic type conversion to/from F16 as needed by the hardware.
|
||||
*/
|
||||
static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
|
||||
// TODO: Use aclnnGroupedMatMul
|
||||
//dst [M, K, N, 1]
|
||||
ggml_tensor * src0 = dst->src[0]; //src0 [D, M, A, 1]
|
||||
ggml_tensor * src1 = dst->src[1]; //src1 [D, B, N, 1], B = K or B = 1
|
||||
ggml_tensor * ids = dst->src[2]; //ids [K, N]
|
||||
// dst: [M, K, N, 1]
|
||||
// src0: [D, M, A, 1] - quantized weights
|
||||
// src1: [D, B, N, 1] - input activations, B = K or B = 1
|
||||
// ids: [K, N] - expert indices
|
||||
ggml_tensor * src0 = dst->src[0];
|
||||
ggml_tensor * src1 = dst->src[1];
|
||||
ggml_tensor * ids = dst->src[2];
|
||||
|
||||
GGML_TENSOR_BINARY_OP_LOCALS
|
||||
GGML_ASSERT(src0->ne[3] == 1);
|
||||
GGML_ASSERT(src1->ne[3] == 1);
|
||||
GGML_ASSERT(dst->ne[3] == 1);
|
||||
GGML_ASSERT(src1->ne[2] == ids->ne[1]);
|
||||
|
||||
// copy index from npu to cpu
|
||||
int64_t n_as = ne02; // A
|
||||
int64_t n_ids = ids->ne[0]; // K
|
||||
const int64_t n_batches = ids->ne[1];
|
||||
const int64_t n_select_experts = ids->ne[0];
|
||||
const enum ggml_type type = src0->type;
|
||||
|
||||
std::vector<char> ids_host(ggml_nbytes(ids));
|
||||
ACL_CHECK(aclrtMemcpyAsync(ids_host.data(), ggml_nbytes(ids), ids->data, ggml_nbytes(ids),
|
||||
ACL_MEMCPY_DEVICE_TO_HOST, ctx.stream()));
|
||||
ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
|
||||
const int32_t group_size = QK8_0; // Both Q4_0 and Q8_0 use group size of 32
|
||||
GGML_ASSERT(group_size == QK4_0);
|
||||
|
||||
char * src0_original = (char *) src0->data;
|
||||
char * src1_original = (char *) src1->data;
|
||||
char * dst_original = (char *) dst->data;
|
||||
// Calculate element size for quantized weights
|
||||
const float weight_elem_size =
|
||||
(type == GGML_TYPE_Q4_0) ? 0.5f :
|
||||
(type == GGML_TYPE_Q8_0) ? 1.0f :
|
||||
(GGML_ABORT("MUL_MAT_ID only supports Q4_0 and Q8_0"), 0.0f);
|
||||
|
||||
ggml_tensor src0_row = *src0;
|
||||
ggml_tensor src1_row = *src1;
|
||||
ggml_tensor dst_row = *dst;
|
||||
// Calculate scale offset in memory
|
||||
const size_t weight_size = src0->ne[0] * src0->ne[1] * src0->ne[2] * weight_elem_size;
|
||||
const size_t scale_elem_size = sizeof(uint16_t);
|
||||
char * scale_data = (char *) src0->data + weight_size;
|
||||
|
||||
const enum ggml_type type = dst->src[0]->type;
|
||||
float weight_elem_size;
|
||||
if (type == GGML_TYPE_Q4_0) {
|
||||
weight_elem_size = float(sizeof(uint8_t)) / 2;
|
||||
} else if (type == GGML_TYPE_Q8_0) {
|
||||
weight_elem_size = float(sizeof(uint8_t));
|
||||
} else {
|
||||
GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
|
||||
}
|
||||
// Allocate buffers for selected expert weights and scales
|
||||
const size_t selected_weight_size = src0->ne[0] * src0->ne[1] * n_select_experts * weight_elem_size;
|
||||
ggml_cann_pool_alloc selected_weight_alloc(ctx.pool(), selected_weight_size);
|
||||
void * selected_weight_buffer = selected_weight_alloc.get();
|
||||
|
||||
// src0_row [D, M, 1, 1] weight without permute
|
||||
src0_row.ne[2] = 1;
|
||||
src0_row.ne[3] = 1;
|
||||
src0_row.nb[0] = weight_elem_size;
|
||||
src0_row.nb[1] = weight_elem_size * ne00;
|
||||
src0_row.nb[2] = weight_elem_size * ne00;
|
||||
src0_row.nb[3] = weight_elem_size * ne00;
|
||||
size_t weight_stride = ne00 * ne01 * weight_elem_size;
|
||||
size_t weight_size = weight_stride * ne02 * ne03;
|
||||
const size_t selected_scale_size = (src0->ne[0] / group_size) * src0->ne[1] * n_select_experts * scale_elem_size;
|
||||
ggml_cann_pool_alloc selected_scale_alloc(ctx.pool(), selected_scale_size);
|
||||
void * selected_scale_buffer = selected_scale_alloc.get();
|
||||
|
||||
// scale [D, M, 1, 1] -> scale && permute
|
||||
size_t scale_elem_size = sizeof(uint16_t);
|
||||
size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
|
||||
// Helper lambda to allocate and cast tensor to F16 if needed
|
||||
constexpr size_t f16_elem_size = sizeof(uint16_t);
|
||||
auto prepare_f16_buffer = [&](ggml_tensor * tensor, ggml_cann_pool_alloc & allocator,
|
||||
bool need_cast = false) -> void * {
|
||||
if (tensor->type == GGML_TYPE_F16) {
|
||||
return tensor->data;
|
||||
}
|
||||
|
||||
// src1_row [D, 1, 1, 1] -> input
|
||||
src1_row.ne[1] = 1;
|
||||
src1_row.ne[2] = 1;
|
||||
src1_row.ne[3] = 1;
|
||||
src1_row.nb[2] = nb11;
|
||||
src1_row.nb[3] = nb11;
|
||||
size_t total_size = f16_elem_size;
|
||||
for (int i = 0; i < GGML_MAX_DIMS; i++) {
|
||||
total_size *= tensor->ne[i];
|
||||
}
|
||||
void * buffer = allocator.alloc(total_size);
|
||||
|
||||
// dst_row [M, 1, 1, 1] -> out
|
||||
dst_row.ne[1] = 1;
|
||||
dst_row.ne[2] = 1;
|
||||
dst_row.ne[3] = 1;
|
||||
dst_row.nb[2] = nb1;
|
||||
dst_row.nb[3] = nb1;
|
||||
if (need_cast == false) {
|
||||
return buffer;
|
||||
}
|
||||
|
||||
//create weight for one row
|
||||
ggml_cann_pool_alloc weight_allocator(ctx.pool());
|
||||
void * weight_buffer = weight_allocator.alloc(nb02);
|
||||
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
|
||||
for (int64_t id = 0; id < n_ids; id++) {
|
||||
// expert index
|
||||
int32_t i02 = *(int32_t *) (ids_host.data() + iid1 * ids->nb[1] + id * ids->nb[0]);
|
||||
GGML_ASSERT(i02 >= 0 && i02 < n_as);
|
||||
int64_t ne[GGML_MAX_DIMS];
|
||||
size_t nb[GGML_MAX_DIMS] = { f16_elem_size };
|
||||
for (int i = 0; i < GGML_MAX_DIMS; i++) {
|
||||
ne[i] = tensor->ne[i];
|
||||
if (i > 0) {
|
||||
nb[i] = nb[i - 1] * ne[i - 1];
|
||||
}
|
||||
}
|
||||
|
||||
// If B = 1 (broadcast), always use 0; otherwise, use id.
|
||||
int64_t i11 = (ne11 == 1 ? 0 : id);
|
||||
int64_t i12 = iid1;
|
||||
acl_tensor_ptr src_tensor = ggml_cann_create_tensor(tensor);
|
||||
acl_tensor_ptr f16_tensor = ggml_cann_create_tensor(buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
|
||||
aclnn_cast(ctx, src_tensor.get(), f16_tensor.get(), ACL_FLOAT16);
|
||||
|
||||
int64_t i1 = id;
|
||||
int64_t i2 = i12;
|
||||
return buffer;
|
||||
};
|
||||
|
||||
void * src0_tmp_ptr = src0_original + i02 * weight_stride;
|
||||
void * scale_tmp_ptr = src0_original + weight_size + i02 * scale_stride;
|
||||
void * src1_tmp_ptr = src1_original + i11 * nb11 + i12 * nb12;
|
||||
void * dst_tmp_ptr = dst_original + i1 * nb1 + i2 * nb2;
|
||||
// Prepare input and output buffers
|
||||
ggml_cann_pool_alloc input_alloc(ctx.pool());
|
||||
void * input_buffer = prepare_f16_buffer(src1, input_alloc, true);
|
||||
|
||||
// mem cpy
|
||||
ACL_CHECK(aclrtMemcpyAsync(weight_buffer, weight_stride, src0_tmp_ptr, weight_stride,
|
||||
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
|
||||
void * scale_buffer = (char *) weight_buffer + weight_stride;
|
||||
ACL_CHECK(aclrtMemcpyAsync(scale_buffer, scale_stride, scale_tmp_ptr, scale_stride,
|
||||
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
|
||||
ggml_cann_pool_alloc output_alloc(ctx.pool());
|
||||
void * output_buffer = prepare_f16_buffer(dst, output_alloc, false);
|
||||
|
||||
src0_row.data = weight_buffer;
|
||||
src1_row.data = src1_tmp_ptr;
|
||||
dst_row.data = dst_tmp_ptr;
|
||||
dst_row.src[0] = &src0_row;
|
||||
dst_row.src[1] = &src1_row;
|
||||
// Process each batch
|
||||
for (int64_t batch_idx = 0; batch_idx < n_batches; batch_idx++) {
|
||||
// Create index tensor for current batch
|
||||
const size_t index_offset = batch_idx * ids->nb[1];
|
||||
acl_tensor_ptr batch_indices = ggml_cann_create_tensor(ids, ids->ne, ids->nb, 1, ACL_FORMAT_ND, index_offset);
|
||||
|
||||
ggml_cann_mul_mat(ctx, &dst_row);
|
||||
// Select quantized weights using expert indices
|
||||
// Q4_0 stores 2 values per byte, Q8_0 stores 1 value per byte
|
||||
const int64_t weight_d = (type == GGML_TYPE_Q4_0) ? src0->ne[0] / 2 : src0->ne[0];
|
||||
const int64_t weight_m = src0->ne[1];
|
||||
const int64_t weight_n_experts = src0->ne[2];
|
||||
|
||||
int64_t weight_ne[3] = { weight_d, weight_m, weight_n_experts };
|
||||
size_t weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t), weight_d * weight_m * sizeof(int8_t) };
|
||||
|
||||
acl_tensor_ptr all_weights =
|
||||
ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb, 3);
|
||||
|
||||
int64_t selected_weight_ne[3] = { weight_d, weight_m, n_select_experts };
|
||||
size_t selected_weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t),
|
||||
weight_d * weight_m * sizeof(int8_t) };
|
||||
|
||||
acl_tensor_ptr selected_weights = ggml_cann_create_tensor(selected_weight_buffer, ACL_INT8, sizeof(int8_t),
|
||||
selected_weight_ne, selected_weight_nb, 3);
|
||||
|
||||
GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_weights.get(), 0, batch_indices.get(), selected_weights.get());
|
||||
|
||||
// Select scales using the same expert indices
|
||||
const int64_t scale_d = src0->ne[0] / group_size;
|
||||
int64_t scale_ne[3] = { scale_d, weight_m, weight_n_experts };
|
||||
size_t scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size, scale_d * weight_m * scale_elem_size };
|
||||
|
||||
acl_tensor_ptr all_scales =
|
||||
ggml_cann_create_tensor(scale_data, ACL_FLOAT16, scale_elem_size, scale_ne, scale_nb, 3);
|
||||
|
||||
int64_t selected_scale_ne[3] = { scale_d, weight_m, n_select_experts };
|
||||
size_t selected_scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size,
|
||||
scale_d * weight_m * scale_elem_size };
|
||||
|
||||
acl_tensor_ptr selected_scales = ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size,
|
||||
selected_scale_ne, selected_scale_nb, 3);
|
||||
|
||||
GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_scales.get(), 0, batch_indices.get(), selected_scales.get());
|
||||
|
||||
// Process each expert for current batch
|
||||
// IndexSelect output layout: [D, M, K] in contiguous format
|
||||
// WeightQuantBatchMatmulV2 expects: [M, D] with row-major stride
|
||||
for (int64_t expert_idx = 0; expert_idx < n_select_experts; expert_idx++) {
|
||||
// Determine input offset: broadcast if src1->ne[1]==1, otherwise use per-expert input
|
||||
const size_t input_offset =
|
||||
(batch_idx * src1->ne[1] + (src1->ne[1] == 1 ? 0 : expert_idx)) * src1->ne[0] * f16_elem_size;
|
||||
const size_t output_offset = (batch_idx * dst->ne[1] + expert_idx) * dst->ne[0] * f16_elem_size;
|
||||
|
||||
// Create weight view for current expert: [D, M, K] -> [M, D]
|
||||
int64_t weight_view_ne[2] = { weight_m, src0->ne[0] };
|
||||
float weight_view_nb[2] = { src0->ne[0] * weight_elem_size, weight_elem_size };
|
||||
const size_t weight_view_offset = expert_idx * selected_weight_nb[2];
|
||||
|
||||
acl_tensor_ptr weight_view =
|
||||
ggml_cann_create_tensor(selected_weight_buffer, ggml_cann_type_mapping(type), weight_elem_size,
|
||||
weight_view_ne, weight_view_nb, 2, ACL_FORMAT_ND, weight_view_offset);
|
||||
|
||||
// Create scale view for current expert: [D, M, K] -> [M, D]
|
||||
int64_t scale_view_ne[2] = { weight_m, scale_d };
|
||||
size_t scale_view_nb[2] = { selected_scale_nb[1], selected_scale_nb[0] };
|
||||
const size_t scale_view_offset = expert_idx * selected_scale_nb[2];
|
||||
|
||||
acl_tensor_ptr scale_view =
|
||||
ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size, scale_view_ne,
|
||||
scale_view_nb, 2, ACL_FORMAT_ND, scale_view_offset);
|
||||
|
||||
// Create input activation tensor [D, 1]
|
||||
int64_t input_ne[2] = { src1->ne[0], 1 };
|
||||
size_t input_nb[2] = { f16_elem_size, src1->ne[0] * f16_elem_size };
|
||||
|
||||
acl_tensor_ptr input_tensor = ggml_cann_create_tensor(input_buffer, ACL_FLOAT16, f16_elem_size, input_ne,
|
||||
input_nb, 2, ACL_FORMAT_ND, input_offset);
|
||||
|
||||
// Create output tensor [M, 1]
|
||||
int64_t output_ne[2] = { dst->ne[0], 1 };
|
||||
size_t output_nb[2] = { f16_elem_size, dst->ne[0] * f16_elem_size };
|
||||
|
||||
acl_tensor_ptr output_tensor = ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, output_ne,
|
||||
output_nb, 2, ACL_FORMAT_ND, output_offset);
|
||||
|
||||
// Perform quantized matrix multiplication
|
||||
GGML_CANN_CALL_ACLNN_OP(ctx, WeightQuantBatchMatmulV2, input_tensor.get(), weight_view.get(),
|
||||
scale_view.get(), nullptr, nullptr, nullptr, nullptr, group_size,
|
||||
output_tensor.get());
|
||||
}
|
||||
}
|
||||
return;
|
||||
|
||||
// Cast output back to original type if we used a temporary F16 buffer
|
||||
if (dst->type != GGML_TYPE_F16) {
|
||||
int64_t ne[GGML_MAX_DIMS];
|
||||
size_t nb[GGML_MAX_DIMS] = { f16_elem_size };
|
||||
for (int i = 0; i < GGML_MAX_DIMS; i++) {
|
||||
ne[i] = dst->ne[i];
|
||||
if (i > 0) {
|
||||
nb[i] = nb[i - 1] * ne[i - 1];
|
||||
}
|
||||
}
|
||||
|
||||
acl_tensor_ptr f16_output =
|
||||
ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
|
||||
acl_tensor_ptr dst_tensor = ggml_cann_create_tensor(dst);
|
||||
|
||||
aclnn_cast(ctx, f16_output.get(), dst_tensor.get(), ggml_cann_type_mapping(dst->type));
|
||||
}
|
||||
}
|
||||
|
||||
void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
|
||||
|
||||
@@ -794,19 +794,44 @@ struct ggml_backend_cann_buffer_context {
|
||||
~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
|
||||
};
|
||||
|
||||
// cann buffer type
|
||||
/**
|
||||
* @brief Check if a buffer is a CANN buffer.
|
||||
*
|
||||
* This function checks if a given buffer is a CANN buffer by comparing its
|
||||
* `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
|
||||
*
|
||||
* @param buffer The buffer to check.
|
||||
* @return true if the buffer is a CANN buffer, false otherwise.
|
||||
* @brief Structure representing context information for a specific backend
|
||||
* buffer type.
|
||||
*/
|
||||
static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
|
||||
struct ggml_backend_cann_buffer_type_context {
|
||||
int32_t device; /**< Device identifier associated with the buffer context. */
|
||||
std::string name; /**< Name associated with the buffer context. */
|
||||
};
|
||||
|
||||
static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
|
||||
return ggml_backend_buft_is_cann(buffer->buft);
|
||||
/**
|
||||
* @brief Retrieves the name associated with a CANN buffer type.
|
||||
*
|
||||
* This function returns the descriptive name associated with the specified
|
||||
* CANN buffer type context.
|
||||
*
|
||||
* @param buft Pointer to the buffer type context.
|
||||
* @return Const pointer to the C-style string containing the name.
|
||||
*/
|
||||
static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
||||
ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
|
||||
|
||||
return buft_ctx->name.c_str();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Checks if the backend buffer type is associated with the CANN backend.
|
||||
*
|
||||
* This function checks whether the provided backend buffer type is associated
|
||||
* with the CANN backend based on the comparison of its name retrieval function
|
||||
* pointer.
|
||||
*
|
||||
* @param buft Pointer to the backend buffer type to check.
|
||||
* @return bool Returns true if the buffer type is associated with the CANN
|
||||
* backend, otherwise false.
|
||||
*/
|
||||
static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
|
||||
return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -1271,7 +1296,7 @@ static void ggml_backend_cann_buffer_get_tensor(ggml_backend_buffer_t buffer,
|
||||
static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
|
||||
const ggml_tensor * src,
|
||||
ggml_tensor * dst) {
|
||||
if (ggml_backend_buffer_is_cann(src->buffer)) {
|
||||
if (ggml_backend_buft_is_cann(src->buffer->buft)) {
|
||||
ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
|
||||
ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;
|
||||
|
||||
@@ -1335,31 +1360,6 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
|
||||
/* .reset = */ NULL,
|
||||
};
|
||||
|
||||
// cann buffer type
|
||||
/**
|
||||
* @brief Structure representing context information for a specific backend
|
||||
* buffer type.
|
||||
*/
|
||||
struct ggml_backend_cann_buffer_type_context {
|
||||
int32_t device; /**< Device identifier associated with the buffer context. */
|
||||
std::string name; /**< Name associated with the buffer context. */
|
||||
};
|
||||
|
||||
/**
|
||||
* @brief Retrieves the name associated with a CANN buffer type.
|
||||
*
|
||||
* This function returns the descriptive name associated with the specified
|
||||
* CANN buffer type context.
|
||||
*
|
||||
* @param buft Pointer to the buffer type context.
|
||||
* @return Const pointer to the C-style string containing the name.
|
||||
*/
|
||||
static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
||||
ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
|
||||
|
||||
return buft_ctx->name.c_str();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Allocates a new CANN buffer of the specified type and size.
|
||||
*
|
||||
@@ -1997,7 +1997,7 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t backend_src,
|
||||
|
||||
GGML_ASSERT(!is_matmul_weight((const ggml_tensor *) src));
|
||||
|
||||
if (!ggml_backend_buffer_is_cann(src->buffer) || !ggml_backend_buffer_is_cann(dst->buffer)) {
|
||||
if (!ggml_backend_buft_is_cann(src->buffer->buft) || !ggml_backend_buft_is_cann(dst->buffer->buft)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
@@ -2523,21 +2523,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
|
||||
GGML_UNUSED(dev);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Checks if the backend buffer type is associated with the CANN backend.
|
||||
*
|
||||
* This function checks whether the provided backend buffer type is associated
|
||||
* with the CANN backend based on the comparison of its name retrieval function
|
||||
* pointer.
|
||||
*
|
||||
* @param buft Pointer to the backend buffer type to check.
|
||||
* @return bool Returns true if the buffer type is associated with the CANN
|
||||
* backend, otherwise false.
|
||||
*/
|
||||
static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
|
||||
return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Records an event on the CANN backend stream.
|
||||
*
|
||||
|
||||
@@ -43,6 +43,7 @@
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
|
||||
@@ -55,7 +56,8 @@
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
# define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
|
||||
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
|
||||
@@ -76,6 +78,7 @@
|
||||
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
|
||||
@@ -84,6 +87,7 @@
|
||||
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
|
||||
@@ -107,6 +111,7 @@
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
|
||||
@@ -119,6 +124,7 @@
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
|
||||
@@ -143,6 +149,7 @@
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
|
||||
@@ -155,6 +162,7 @@
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
|
||||
@@ -186,6 +194,7 @@
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
|
||||
@@ -197,6 +206,7 @@
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
|
||||
@@ -227,6 +237,7 @@
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
|
||||
@@ -239,6 +250,7 @@
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
|
||||
@@ -271,6 +283,7 @@
|
||||
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
|
||||
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
|
||||
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
|
||||
@@ -283,6 +296,7 @@
|
||||
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
|
||||
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
|
||||
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
|
||||
|
||||
@@ -1072,6 +1072,195 @@ void ggml_gemv_q5_K_8x8_q8_K(int n,
|
||||
ggml_gemv_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemv_q6_K_8x4_q8_K(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
const void * GGML_RESTRICT vx,
|
||||
const void * GGML_RESTRICT vy,
|
||||
int nr,
|
||||
int nc) {
|
||||
constexpr int qk = QK_K;
|
||||
const int nb = n / qk;
|
||||
|
||||
constexpr int ncols_interleaved = 8;
|
||||
constexpr int blocklen = 4;
|
||||
|
||||
assert(n % qk == 0);
|
||||
assert(nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(nb);
|
||||
UNUSED(ncols_interleaved);
|
||||
UNUSED(blocklen);
|
||||
|
||||
#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
|
||||
constexpr int col_groups = ncols_interleaved / 4;
|
||||
const uint8x16_t m4b = vdupq_n_u8(0x0f);
|
||||
const uint8x16_t mask_lo = vdupq_n_u8(0x03);
|
||||
const uint8x16_t mask_hi = vdupq_n_u8(0x30);
|
||||
|
||||
// 1x8 tile = 2 x 4
|
||||
float32x4_t acc_f32[2];
|
||||
|
||||
const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
|
||||
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
|
||||
|
||||
for (int i = 0; i < col_groups; i++) {
|
||||
acc_f32[i] = vdupq_n_f32(0);
|
||||
}
|
||||
|
||||
for (int b = 0; b < nb; b++) {
|
||||
float32x4_t q6_d_0 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d)); // d0 d1 d2 d3
|
||||
float32x4_t q6_d_1 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4)); // d4 d5 d6 d7
|
||||
float32x4_t q8_d = vdupq_n_f32(q8_ptr[b].d);
|
||||
float32x4_t sb_scale_0 = vmulq_f32(q6_d_0, q8_d);
|
||||
float32x4_t sb_scale_1 = vmulq_f32(q6_d_1, q8_d);
|
||||
|
||||
int32x4_t acc[col_groups];
|
||||
for (int i = 0; i < col_groups; i++) {
|
||||
acc[i] = vdupq_n_s32(0);
|
||||
}
|
||||
|
||||
// Load all 16 scales once and widen to int16 (Q6_K has 16 scales per block)
|
||||
// Reused for bias and dequantization later
|
||||
int16_t q6_scales[16 * 8];
|
||||
for (int i = 0; i < 16; i++) {
|
||||
int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
|
||||
vst1q_s16(q6_scales + i * 8, scales);
|
||||
}
|
||||
|
||||
// Compute bias per column using q8 bsums and preloaded scales to skip the -32 shift
|
||||
int32x4_t bias_lo = vdupq_n_s32(0);
|
||||
int32x4_t bias_hi = vdupq_n_s32(0);
|
||||
|
||||
// Load bsums in chunks of 4 to process with vectorized operations
|
||||
for (int i = 0; i < 16; i += 4) {
|
||||
int16x4_t bsums_vec = vld1_s16(q8_ptr[b].bsums + i);
|
||||
int16x4_t scales_lo_0 = vld1_s16(q6_scales + (i + 0) * 8);
|
||||
int16x4_t scales_hi_0 = vld1_s16(q6_scales + (i + 0) * 8 + 4);
|
||||
int16x4_t scales_lo_1 = vld1_s16(q6_scales + (i + 1) * 8);
|
||||
int16x4_t scales_hi_1 = vld1_s16(q6_scales + (i + 1) * 8 + 4);
|
||||
int16x4_t scales_lo_2 = vld1_s16(q6_scales + (i + 2) * 8);
|
||||
int16x4_t scales_hi_2 = vld1_s16(q6_scales + (i + 2) * 8 + 4);
|
||||
int16x4_t scales_lo_3 = vld1_s16(q6_scales + (i + 3) * 8);
|
||||
int16x4_t scales_hi_3 = vld1_s16(q6_scales + (i + 3) * 8 + 4);
|
||||
|
||||
bias_lo = vmlal_lane_s16(bias_lo, scales_lo_0, bsums_vec, 0);
|
||||
bias_hi = vmlal_lane_s16(bias_hi, scales_hi_0, bsums_vec, 0);
|
||||
bias_lo = vmlal_lane_s16(bias_lo, scales_lo_1, bsums_vec, 1);
|
||||
bias_hi = vmlal_lane_s16(bias_hi, scales_hi_1, bsums_vec, 1);
|
||||
bias_lo = vmlal_lane_s16(bias_lo, scales_lo_2, bsums_vec, 2);
|
||||
bias_hi = vmlal_lane_s16(bias_hi, scales_hi_2, bsums_vec, 2);
|
||||
bias_lo = vmlal_lane_s16(bias_lo, scales_lo_3, bsums_vec, 3);
|
||||
bias_hi = vmlal_lane_s16(bias_hi, scales_hi_3, bsums_vec, 3);
|
||||
}
|
||||
bias_lo = vshlq_n_s32(bias_lo, 5);
|
||||
bias_hi = vshlq_n_s32(bias_hi, 5);
|
||||
|
||||
// Process two 128-value halves per superblock
|
||||
for (int half = 0; half < 2; half++) {
|
||||
const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
|
||||
const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
|
||||
|
||||
// A subblock (sb) is a set of weights that share the scale
|
||||
// Since q6_K scales are per 16 elements
|
||||
// num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves)
|
||||
for (int sb = 0; sb < QK_K / 64; sb++) {
|
||||
const int8_t * q8_base_l = q8_ptr[b].qs + half * 128 + sb * 16;
|
||||
const int8_t * q8_base_h = q8_base_l + 64;
|
||||
|
||||
// Load and duplicate q8 values (each register covers four interleaved columns of q6)
|
||||
int8x16_t q8_l[4];
|
||||
int8x16_t q8_h[4];
|
||||
for (int i = 0; i < 4; i++) {
|
||||
q8_l[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_l + i * 4));
|
||||
q8_h[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_h + i * 4));
|
||||
}
|
||||
|
||||
const int ql_off_base = sb * QK_K / 2;
|
||||
const int qh_off_base = ql_off_base & 255; // wraps after 256 bytes
|
||||
|
||||
// Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
|
||||
uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
|
||||
uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
|
||||
uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
|
||||
uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);
|
||||
|
||||
// Adjust qh for subblocks 2 and 3 (shift right by 2)
|
||||
if (sb > 1) {
|
||||
q6_qh_0.val[0] = vshrq_n_u8(q6_qh_0.val[0], 2);
|
||||
q6_qh_0.val[1] = vshrq_n_u8(q6_qh_0.val[1], 2);
|
||||
q6_qh_0.val[2] = vshrq_n_u8(q6_qh_0.val[2], 2);
|
||||
q6_qh_0.val[3] = vshrq_n_u8(q6_qh_0.val[3], 2);
|
||||
q6_qh_1.val[0] = vshrq_n_u8(q6_qh_1.val[0], 2);
|
||||
q6_qh_1.val[1] = vshrq_n_u8(q6_qh_1.val[1], 2);
|
||||
q6_qh_1.val[2] = vshrq_n_u8(q6_qh_1.val[2], 2);
|
||||
q6_qh_1.val[3] = vshrq_n_u8(q6_qh_1.val[3], 2);
|
||||
}
|
||||
|
||||
const uint8x16_t q6_ql[8] = { q6_ql_0.val[0], q6_ql_0.val[1], q6_ql_0.val[2], q6_ql_0.val[3],
|
||||
q6_ql_1.val[0], q6_ql_1.val[1], q6_ql_1.val[2], q6_ql_1.val[3] };
|
||||
const uint8x16_t q6_qh[8] = { q6_qh_0.val[0], q6_qh_0.val[1], q6_qh_0.val[2], q6_qh_0.val[3],
|
||||
q6_qh_1.val[0], q6_qh_1.val[1], q6_qh_1.val[2], q6_qh_1.val[3] };
|
||||
|
||||
// Process column groups (0-3, 4-7)
|
||||
for (int g = 0; g < col_groups; g++) {
|
||||
int32x4_t sb_acc_l = vdupq_n_s32(0);
|
||||
int32x4_t sb_acc_h = vdupq_n_s32(0);
|
||||
|
||||
for (int chunk = 0; chunk < 4; chunk++) {
|
||||
const int idx = chunk * 2 + g;
|
||||
|
||||
const uint8x16_t q6_qs_l = q6_ql[idx];
|
||||
const uint8x16_t q6_qs_h = q6_qh[idx];
|
||||
|
||||
// Extract high 2 bits for upper nibble reconstruction
|
||||
const uint8x16_t q6_qs_hh = vandq_u8(q6_qs_h, mask_hi);
|
||||
|
||||
// q6 = (low4 | high2<<4), without -32 bias (handled via bsums)
|
||||
const int8x16_t q6_l =
|
||||
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_l, m4b), vandq_u8(q6_qs_h, mask_lo), 4));
|
||||
const int8x16_t q6_h = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_l, 4), q6_qs_hh));
|
||||
|
||||
sb_acc_l = vdotq_s32(sb_acc_l, q6_l, q8_l[chunk]);
|
||||
sb_acc_h = vdotq_s32(sb_acc_h, q6_h, q8_h[chunk]);
|
||||
}
|
||||
|
||||
const int scale_idx_l = half * 8 + sb;
|
||||
const int scale_idx_h = half * 8 + sb + 4;
|
||||
|
||||
const int32x4_t scale_vec_l = vmovl_s16(vld1_s16(q6_scales + scale_idx_l * 8 + g * 4));
|
||||
const int32x4_t scale_vec_h = vmovl_s16(vld1_s16(q6_scales + scale_idx_h * 8 + g * 4));
|
||||
|
||||
acc[g] = vmlaq_s32(acc[g], sb_acc_l, scale_vec_l);
|
||||
acc[g] = vmlaq_s32(acc[g], sb_acc_h, scale_vec_h);
|
||||
}
|
||||
}
|
||||
} // for half
|
||||
|
||||
// Bias correction
|
||||
acc[0] = vsubq_s32(acc[0], bias_lo);
|
||||
acc[1] = vsubq_s32(acc[1], bias_hi);
|
||||
|
||||
// Apply superblock scale (no mins for q6_K)
|
||||
// acc[g] has [c0, c1, c2, c3]
|
||||
float32x4_t w_0123 = vmulq_f32(vcvtq_f32_s32(acc[0]), sb_scale_0);
|
||||
float32x4_t w_4567 = vmulq_f32(vcvtq_f32_s32(acc[1]), sb_scale_1);
|
||||
|
||||
acc_f32[0] = vaddq_f32(acc_f32[0], w_0123);
|
||||
acc_f32[1] = vaddq_f32(acc_f32[1], w_4567);
|
||||
} // for b
|
||||
|
||||
int base = x * ncols_interleaved;
|
||||
vst1q_f32(s + base, acc_f32[0]);
|
||||
vst1q_f32(s + base + 4, acc_f32[1]);
|
||||
} // for x
|
||||
return;
|
||||
#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
|
||||
ggml_gemv_q6_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemv_q6_K_8x8_q8_K(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
@@ -1177,15 +1366,14 @@ void ggml_gemv_q6_K_8x8_q8_K(int n,
|
||||
q8_h[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_h + i * 8));
|
||||
}
|
||||
|
||||
// TODO: Test other qh repack patterns to reduce loads
|
||||
const int ql_off_base = sb * QK_K / 2;
|
||||
const int qh_off_base = ql_off_base & 255; // wraps after 256 bytes
|
||||
|
||||
// Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
|
||||
ggml_uint8x16x4_t q6_ql_0 = ggml_vld1q_u8_x4(ql_base + ql_off_base);
|
||||
ggml_uint8x16x4_t q6_ql_1 = ggml_vld1q_u8_x4(ql_base + ql_off_base + 64);
|
||||
ggml_uint8x16x4_t q6_qh_0 = ggml_vld1q_u8_x4(qh_base + qh_off_base);
|
||||
ggml_uint8x16x4_t q6_qh_1 = ggml_vld1q_u8_x4(qh_base + qh_off_base + 64);
|
||||
uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
|
||||
uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
|
||||
uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
|
||||
uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);
|
||||
|
||||
// Adjust qh for subblocks 2 and 3 (shift right by 2)
|
||||
if (sb > 1) {
|
||||
@@ -3474,6 +3662,208 @@ void ggml_gemm_q5_K_8x8_q8_K(int n,
|
||||
ggml_gemm_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemm_q6_K_8x4_q8_K(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
const void * GGML_RESTRICT vx,
|
||||
const void * GGML_RESTRICT vy,
|
||||
int nr,
|
||||
int nc) {
|
||||
constexpr int qk = QK_K;
|
||||
const int nb = n / qk;
|
||||
|
||||
constexpr int ncols_interleaved = 8;
|
||||
constexpr int blocklen = 4;
|
||||
|
||||
assert(n % qk == 0);
|
||||
assert(nr % 4 == 0);
|
||||
assert(nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(nb);
|
||||
UNUSED(ncols_interleaved);
|
||||
UNUSED(blocklen);
|
||||
|
||||
#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
|
||||
constexpr int q8_k_blocklen = 4;
|
||||
constexpr int col_groups = ncols_interleaved / 4;
|
||||
constexpr int acc_size = q8_k_blocklen * col_groups; // 4 rows, 2 column groups
|
||||
const uint8x16_t m4b = vdupq_n_u8(0x0f);
|
||||
const uint8x16_t mask_lo = vdupq_n_u8(0x03);
|
||||
const uint8x16_t mask_hi = vdupq_n_u8(0x30);
|
||||
const int8x16_t m32s = vdupq_n_s8(32);
|
||||
|
||||
float32x4_t acc_f32[acc_size];
|
||||
|
||||
for (int y = 0; y < nr / q8_k_blocklen; y++) {
|
||||
const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
|
||||
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
|
||||
|
||||
for (int i = 0; i < acc_size; i++) {
|
||||
acc_f32[i] = vdupq_n_f32(0);
|
||||
}
|
||||
|
||||
for (int b = 0; b < nb; b++) {
|
||||
float32x4_t q6_d_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));
|
||||
float32x4_t q6_d_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));
|
||||
float32x4_t q8_d_0123 = vld1q_f32(q8_ptr[b].d);
|
||||
|
||||
float32x4_t sbd_scale_0123[q8_k_blocklen];
|
||||
float32x4_t sbd_scale_4567[q8_k_blocklen];
|
||||
|
||||
sbd_scale_0123[0] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 0);
|
||||
sbd_scale_4567[0] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 0);
|
||||
sbd_scale_0123[1] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 1);
|
||||
sbd_scale_4567[1] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 1);
|
||||
sbd_scale_0123[2] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 2);
|
||||
sbd_scale_4567[2] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 2);
|
||||
sbd_scale_0123[3] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 3);
|
||||
sbd_scale_4567[3] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 3);
|
||||
|
||||
int32x4_t acc_s32[acc_size];
|
||||
for (int i = 0; i < acc_size; i++) {
|
||||
acc_s32[i] = vdupq_n_s32(0);
|
||||
}
|
||||
|
||||
int16_t q6_scales[8 * 16];
|
||||
for (int i = 0; i < 16; i++) {
|
||||
int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
|
||||
vst1q_s16(q6_scales + i * 8, scales);
|
||||
}
|
||||
|
||||
for (int half = 0; half < 2; half++) {
|
||||
const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
|
||||
const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
|
||||
|
||||
for (int sb = 0; sb < QK_K / 64; sb++) {
|
||||
int32x4_t acc_lo[acc_size];
|
||||
int32x4_t acc_hi[acc_size];
|
||||
for (int i = 0; i < acc_size; i++) {
|
||||
acc_lo[i] = vdupq_n_s32(0);
|
||||
acc_hi[i] = vdupq_n_s32(0);
|
||||
}
|
||||
|
||||
const int8_t * q8_base_l = q8_ptr[b].qs + half * 512 + sb * 64;
|
||||
const int8_t * q8_base_h = q8_ptr[b].qs + half * 512 + 256 + sb * 64;
|
||||
|
||||
// 4 rows * 16 elements per scale
|
||||
// 4 reads of 16 bytes each
|
||||
constexpr int reads_per_sb = 4;
|
||||
int8x16_t q8_l[reads_per_sb];
|
||||
int8x16_t q8_h[reads_per_sb];
|
||||
for (int k = 0; k < reads_per_sb; k++) {
|
||||
q8_l[k] = vld1q_s8(q8_base_l + 16 * k);
|
||||
q8_h[k] = vld1q_s8(q8_base_h + 16 * k);
|
||||
}
|
||||
|
||||
const int ql_off_base = sb * QK_K / 2;
|
||||
const int qh_off_base = ql_off_base & 255;
|
||||
|
||||
uint8x16_t q6_ql_0123[reads_per_sb];
|
||||
uint8x16_t q6_ql_4567[reads_per_sb];
|
||||
uint8x16_t q6_qh_0123[reads_per_sb];
|
||||
uint8x16_t q6_qh_4567[reads_per_sb];
|
||||
|
||||
for (int k = 0; k < reads_per_sb; k++) {
|
||||
q6_ql_0123[k] = vld1q_u8(ql_base + ql_off_base + k * 32);
|
||||
q6_ql_4567[k] = vld1q_u8(ql_base + ql_off_base + k * 32 + 16);
|
||||
q6_qh_0123[k] = vld1q_u8(qh_base + qh_off_base + k * 32);
|
||||
q6_qh_4567[k] = vld1q_u8(qh_base + qh_off_base + k * 32 + 16);
|
||||
}
|
||||
|
||||
if (sb > 1) {
|
||||
for (int k = 0; k < reads_per_sb; k++) {
|
||||
q6_qh_0123[k] = vshrq_n_u8(q6_qh_0123[k], 2);
|
||||
q6_qh_4567[k] = vshrq_n_u8(q6_qh_4567[k], 2);
|
||||
}
|
||||
}
|
||||
|
||||
for (int k = 0; k < reads_per_sb; k++) {
|
||||
// q = (ql | qh) - 32
|
||||
const uint8x16_t hbit_lo_0123 = vandq_u8(q6_qh_0123[k], mask_lo);
|
||||
const uint8x16_t hbit_hi_0123 = vandq_u8(q6_qh_0123[k], mask_hi);
|
||||
const uint8x16_t hbit_lo_4567 = vandq_u8(q6_qh_4567[k], mask_lo);
|
||||
const uint8x16_t hbit_hi_4567 = vandq_u8(q6_qh_4567[k], mask_hi);
|
||||
|
||||
const int8x16_t q6_0123_lo = vsubq_s8(
|
||||
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_0123[k], m4b), hbit_lo_0123, 4)), m32s);
|
||||
const int8x16_t q6_0123_hi = vsubq_s8(
|
||||
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_0123[k], 4), hbit_hi_0123)), m32s);
|
||||
|
||||
acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q6_0123_lo, q8_l[k], 0); // 0..3 r0 c0123
|
||||
acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q6_0123_lo, q8_l[k], 1); // 0..3 r1 c0123
|
||||
acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q6_0123_lo, q8_l[k], 2); // 0..3 r2 c0123
|
||||
acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q6_0123_lo, q8_l[k], 3); // 0..3 r3 c0123
|
||||
|
||||
acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q6_0123_hi, q8_h[k], 0); // 64..67 r0 c0123
|
||||
acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q6_0123_hi, q8_h[k], 1); // 64..67 r1 c0123
|
||||
acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q6_0123_hi, q8_h[k], 2); // 64..67 r2 c0123
|
||||
acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q6_0123_hi, q8_h[k], 3); // 64..67 r3 c0123
|
||||
|
||||
const int8x16_t q6_4567_lo = vsubq_s8(
|
||||
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_4567[k], m4b), hbit_lo_4567, 4)), m32s);
|
||||
const int8x16_t q6_4567_hi = vsubq_s8(
|
||||
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_4567[k], 4), hbit_hi_4567)), m32s);
|
||||
|
||||
acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q6_4567_lo, q8_l[k], 0); // 0..3 r0 c4567
|
||||
acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q6_4567_lo, q8_l[k], 1); // 0..3 r1 c4567
|
||||
acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q6_4567_lo, q8_l[k], 2); // 0..3 r2 c4567
|
||||
acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q6_4567_lo, q8_l[k], 3); // 0..3 r3 c4567
|
||||
|
||||
acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q6_4567_hi, q8_h[k], 0); // 64..67 r0 c4567
|
||||
acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q6_4567_hi, q8_h[k], 1); // 64..67 r1 c4567
|
||||
acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q6_4567_hi, q8_h[k], 2); // 64..67 r2 c4567
|
||||
acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q6_4567_hi, q8_h[k], 3); // 64..67 r3 c4567
|
||||
}
|
||||
|
||||
// Scale and bias
|
||||
const int scale_idx_l = half * 8 + sb;
|
||||
const int scale_idx_h = half * 8 + sb + 4;
|
||||
|
||||
for (int g = 0; g < col_groups; g++) {
|
||||
const int16x4_t scales_l16 = vld1_s16(q6_scales + scale_idx_l * 8 + g * 4);
|
||||
const int16x4_t scales_h16 = vld1_s16(q6_scales + scale_idx_h * 8 + g * 4);
|
||||
const int32x4_t scale_vec_l = vmovl_s16(scales_l16);
|
||||
const int32x4_t scale_vec_h = vmovl_s16(scales_h16);
|
||||
const int acc_offset = g * q8_k_blocklen;
|
||||
|
||||
for (int row = 0; row < q8_k_blocklen; row++) {
|
||||
const int idx = row * 2 + g;
|
||||
acc_s32[idx] = vmlaq_s32(acc_s32[idx], acc_lo[acc_offset + row], scale_vec_l);
|
||||
acc_s32[idx] = vmlaq_s32(acc_s32[idx], acc_hi[acc_offset + row], scale_vec_h);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Finally we apply the superblock scales
|
||||
for (int row = 0; row < q8_k_blocklen; row++) {
|
||||
const int idx0 = 2 * row;
|
||||
const int idx1 = 2 * row + 1;
|
||||
const int32x4_t acc_0123 = acc_s32[idx0];
|
||||
const int32x4_t acc_4567 = acc_s32[idx1];
|
||||
|
||||
acc_f32[idx0] = vmlaq_f32(acc_f32[idx0], vcvtq_f32_s32(acc_0123), sbd_scale_0123[row]);
|
||||
acc_f32[idx1] = vmlaq_f32(acc_f32[idx1], vcvtq_f32_s32(acc_4567), sbd_scale_4567[row]);
|
||||
}
|
||||
} // for b
|
||||
|
||||
for (int i = 0; i < q8_k_blocklen; i++) {
|
||||
int row = y * q8_k_blocklen + i;
|
||||
for (int j = 0; j < 2; j++) {
|
||||
int col = x * ncols_interleaved + j * 4;
|
||||
int offset = row * bs + col;
|
||||
vst1q_f32(s + offset, acc_f32[2 * i + j]);
|
||||
}
|
||||
}
|
||||
} // for x
|
||||
} // for y
|
||||
return;
|
||||
#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
|
||||
ggml_gemm_q6_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemm_q6_K_8x8_q8_K(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
|
||||
@@ -268,9 +268,9 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
|
||||
_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
|
||||
}
|
||||
|
||||
static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
|
||||
return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
|
||||
_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
|
||||
static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const uint8_t x1, const float y1) {
|
||||
return _mm256_set_m128(_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
|
||||
_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
|
||||
}
|
||||
#endif
|
||||
#elif defined(__SSSE3__)
|
||||
@@ -782,6 +782,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
|
||||
__m256 accum1 = _mm256_setzero_ps();
|
||||
__m256 accum2 = _mm256_setzero_ps();
|
||||
|
||||
for (; ib + 1 < nb; ib += 2) {
|
||||
const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
|
||||
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
|
||||
@@ -795,10 +796,10 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
|
||||
const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
|
||||
const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
|
||||
accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
|
||||
_mm256_cvtepi32_ps(p_1), accum1);
|
||||
accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
|
||||
_mm256_cvtepi32_ps(p_2), accum2);
|
||||
const __m256 scale0 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 0].e));
|
||||
const __m256 scale1 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 1].e));
|
||||
accum1 = _mm256_fmadd_ps(scale0, _mm256_cvtepi32_ps(p_1), accum1);
|
||||
accum2 = _mm256_fmadd_ps(scale1, _mm256_cvtepi32_ps(p_2), accum2);
|
||||
}
|
||||
|
||||
sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
|
||||
@@ -830,7 +831,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
|
||||
#endif
|
||||
for (; ib < nb; ++ib) {
|
||||
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
|
||||
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib].e);
|
||||
int sumi1 = 0;
|
||||
int sumi2 = 0;
|
||||
for (int j = 0; j < QK_MXFP4/2; ++j) {
|
||||
@@ -3817,4 +3818,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
|
||||
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
|
||||
#endif
|
||||
}
|
||||
|
||||
|
||||
@@ -59,11 +59,7 @@ static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * ds
|
||||
GGML_ASSERT(nb00 == sizeof(src0_t));
|
||||
|
||||
const auto [ir0, ir1] = get_thread_range(params, src0);
|
||||
const bool is_src1_contiguous = (nb10 == sizeof(src1_t));
|
||||
|
||||
if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous
|
||||
GGML_ASSERT(ggml_are_same_shape(src0, src1));
|
||||
}
|
||||
const bool is_src1_contiguous_rows = ggml_is_contiguous_rows(src1);
|
||||
|
||||
#ifdef GGML_USE_ACCELERATE
|
||||
vDSP_fn_t vDSP_op = nullptr;
|
||||
@@ -94,7 +90,7 @@ static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * ds
|
||||
const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
|
||||
const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
|
||||
|
||||
if (is_src1_contiguous) {
|
||||
if (is_src1_contiguous_rows) {
|
||||
// src1 is broadcastable across src0 and dst in i1, i2, i3
|
||||
const int64_t nr0 = ne00 / ne10;
|
||||
|
||||
|
||||
@@ -75,6 +75,9 @@
|
||||
// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
|
||||
float ggml_table_f32_f16[1 << 16];
|
||||
|
||||
// precomputed f32 table for e8m0 half (1 KB) (simd-mappings.h)
|
||||
float ggml_table_f32_e8m0_half[1 << 8];
|
||||
|
||||
#if defined(__ARM_ARCH)
|
||||
struct ggml_arm_arch_features_type {
|
||||
int sve_cnt;
|
||||
@@ -3681,6 +3684,11 @@ void ggml_cpu_init(void) {
|
||||
ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
|
||||
}
|
||||
|
||||
// initialize E8M0 half table (256 entries)
|
||||
for (int i = 0; i < (1 << 8); ++i) {
|
||||
ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
|
||||
}
|
||||
|
||||
const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
|
||||
|
||||
GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);
|
||||
|
||||
@@ -7629,8 +7629,7 @@ static void ggml_compute_forward_pad_f32(
|
||||
|
||||
const ggml_tensor * src0 = dst->src[0];
|
||||
|
||||
GGML_ASSERT(src0->nb[0] == sizeof(float));
|
||||
GGML_ASSERT( dst->nb[0] == sizeof(float));
|
||||
assert(dst->nb[0] == sizeof(float));
|
||||
|
||||
const int ith = params->ith;
|
||||
const int nth = params->nth;
|
||||
|
||||
@@ -256,6 +256,200 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR
|
||||
ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row);
|
||||
}
|
||||
|
||||
template <int M, int N>
|
||||
static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
const void * GGML_RESTRICT vx,
|
||||
const void * GGML_RESTRICT vy,
|
||||
int nr,
|
||||
int nc) {
|
||||
constexpr int blocklen = M;
|
||||
constexpr int ncols_interleaved = N;
|
||||
const int qk = QK_K;
|
||||
const int nb = n / qk;
|
||||
const int blocks_per_half = 64 / blocklen;
|
||||
|
||||
assert(n % qk == 0);
|
||||
assert(nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(bs);
|
||||
UNUSED(nr);
|
||||
|
||||
float sumf[8];
|
||||
|
||||
const block_q8_K * a_ptr = (const block_q8_K *) vy;
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
sumf[j] = 0.0f;
|
||||
}
|
||||
|
||||
for (int l = 0; l < nb; l++) {
|
||||
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
|
||||
const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
|
||||
const int base_h = base_l + 64;
|
||||
|
||||
const int scale_idx_l = base_l / 16;
|
||||
const int scale_idx_h = base_h / 16;
|
||||
|
||||
const int qh_shift_l = ((base_l % 128) / 32) * 2;
|
||||
const int qh_shift_h = ((base_h % 128) / 32) * 2;
|
||||
|
||||
const int qh_half_l = (base_l / 128) * 32;
|
||||
const int qh_half_h = (base_h / 128) * 32;
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
|
||||
const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
|
||||
|
||||
int sumi_l = 0;
|
||||
int sumi_h = 0;
|
||||
|
||||
for (int i = 0; i < blocklen; i++) {
|
||||
const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
|
||||
const int l_4 = b_ptr[l].ql[ql_pos] & 0xF;
|
||||
const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
|
||||
|
||||
const int qh_idx_l = qh_half_l + ((base_l + i) % 32);
|
||||
const int qh_chunk_l = qh_idx_l / blocklen;
|
||||
const int qh_pos_l = qh_idx_l % blocklen;
|
||||
const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
|
||||
const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
|
||||
|
||||
const int qh_idx_h = qh_half_h + ((base_h + i) % 32);
|
||||
const int qh_chunk_h = qh_idx_h / blocklen;
|
||||
const int qh_pos_h = qh_idx_h % blocklen;
|
||||
const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
|
||||
const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
|
||||
|
||||
const int q_l = ((hi_2_l << 4) | l_4) - 32;
|
||||
const int q_h = ((hi_2_h << 4) | hi_4) - 32;
|
||||
|
||||
const int8_t a_l = a_ptr[l].qs[base_l + i];
|
||||
const int8_t a_h = a_ptr[l].qs[base_h + i];
|
||||
|
||||
sumi_l += q_l * a_l;
|
||||
sumi_h += q_h * a_h;
|
||||
}
|
||||
|
||||
sumf[j] +=
|
||||
(sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
s[x * ncols_interleaved + j] = sumf[j];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <int M, int N>
|
||||
static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
const void * GGML_RESTRICT vx,
|
||||
const void * GGML_RESTRICT vy,
|
||||
int nr,
|
||||
int nc) {
|
||||
constexpr int blocklen = M;
|
||||
constexpr int ncols_interleaved = N;
|
||||
const int qk = QK_K;
|
||||
const int nb = n / qk;
|
||||
const int blocks_per_half = 64 / blocklen;
|
||||
const int q8_half_stride = 512;
|
||||
const int q8_low_high_step = 256;
|
||||
|
||||
assert(n % qk == 0);
|
||||
assert(nr % 4 == 0);
|
||||
assert(nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(bs);
|
||||
|
||||
float sumf[4][8];
|
||||
|
||||
for (int y = 0; y < nr / 4; y++) {
|
||||
const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
|
||||
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
sumf[m][j] = 0.0f;
|
||||
}
|
||||
}
|
||||
|
||||
for (int l = 0; l < nb; l++) {
|
||||
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
|
||||
const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
|
||||
const int base_h = base_l + 64;
|
||||
|
||||
const int scale_idx_l = base_l / 16;
|
||||
const int scale_idx_h = base_h / 16;
|
||||
|
||||
const int qh_shift_l = ((base_l % 128) / 32) * 2;
|
||||
const int qh_shift_h = ((base_h % 128) / 32) * 2;
|
||||
|
||||
const int qh_half_l = (base_l / 128) * 32;
|
||||
const int qh_half_h = (base_h / 128) * 32;
|
||||
|
||||
const int q8_base = (k / blocks_per_half) * q8_half_stride + (k % blocks_per_half) * (blocklen * 4);
|
||||
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
|
||||
const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
|
||||
|
||||
int sumi_l = 0;
|
||||
int sumi_h = 0;
|
||||
|
||||
for (int i = 0; i < blocklen; i++) {
|
||||
const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
|
||||
const int l_4 = b_ptr[l].ql[ql_pos] & 0xF;
|
||||
const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
|
||||
|
||||
const int qh_idx_l = qh_half_l + ((base_l + i) % 32);
|
||||
const int qh_chunk_l = qh_idx_l / blocklen;
|
||||
const int qh_pos_l = qh_idx_l % blocklen;
|
||||
const int qh_offset_l =
|
||||
qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
|
||||
const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
|
||||
|
||||
const int qh_idx_h = qh_half_h + ((base_h + i) % 32);
|
||||
const int qh_chunk_h = qh_idx_h / blocklen;
|
||||
const int qh_pos_h = qh_idx_h % blocklen;
|
||||
const int qh_offset_h =
|
||||
qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
|
||||
const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
|
||||
|
||||
const int q_l = ((hi_2_l << 4) | l_4) - 32;
|
||||
const int q_h = ((hi_2_h << 4) | hi_4) - 32;
|
||||
|
||||
const int8_t q8_l = a_ptr[l].qs[q8_base + m * blocklen + i];
|
||||
const int8_t q8_h = a_ptr[l].qs[q8_base + m * blocklen + i + q8_low_high_step];
|
||||
|
||||
sumi_l += q_l * q8_l;
|
||||
sumi_h += q_h * q8_h;
|
||||
}
|
||||
|
||||
sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
|
||||
a_ptr[l].d[m];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
extern "C" {
|
||||
|
||||
void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
@@ -704,94 +898,12 @@ void ggml_gemv_q5_K_8x8_q8_K_generic(int n,
|
||||
}
|
||||
|
||||
|
||||
void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
ggml_gemv_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
constexpr int qk = QK_K;
|
||||
const int nb = n / qk;
|
||||
const int ncols_interleaved = 8;
|
||||
const int blocklen = 8;
|
||||
|
||||
assert(n % qk == 0);
|
||||
assert(nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(bs);
|
||||
UNUSED(nr);
|
||||
|
||||
float sumf[8];
|
||||
|
||||
const block_q8_K * a_ptr = (const block_q8_K *) vy;
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
sumf[j] = 0.0f;
|
||||
}
|
||||
|
||||
for (int l = 0; l < nb; l++) {
|
||||
|
||||
|
||||
for (int k = 0; k < 16; k++) {
|
||||
// k = 0.. 7 weights 0-63 low, 64-127 high
|
||||
// k = 8..15 weights 128-191 low, 192-255 high
|
||||
const int base_l = (k / 8) * 128 + (k % 8) * 8;
|
||||
const int base_h = base_l + 64;
|
||||
|
||||
const int scale_idx_l = base_l / 16;
|
||||
const int scale_idx_h = base_h / 16;
|
||||
|
||||
// Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
|
||||
const int qh_shift_l = ((base_l % 128) / 32) * 2;
|
||||
const int qh_shift_h = ((base_h % 128) / 32) * 2;
|
||||
|
||||
// qh_half: offset to the correct 32-byte half (0 or 32)
|
||||
const int qh_half_l = (base_l / 128) * 32;
|
||||
const int qh_half_h = (base_h / 128) * 32;
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
// Interleaved scales
|
||||
const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
|
||||
const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
|
||||
|
||||
int sumi_l = 0;
|
||||
int sumi_h = 0;
|
||||
|
||||
for (int i = 0; i < blocklen; i++) {
|
||||
const int ql_pos = k * 64 + j * 8 + i;
|
||||
const int l_4 = b_ptr[l].ql[ql_pos] & 0xF;
|
||||
const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
|
||||
|
||||
// qh indexing with 8-byte interleaving (like q5_K)
|
||||
const int qh_byte_l = qh_half_l + ((base_l + i) % 32);
|
||||
const int qh_chunk_l = qh_byte_l / 8;
|
||||
const int qh_pos_l = qh_byte_l % 8;
|
||||
const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
|
||||
const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
|
||||
|
||||
const int qh_byte_h = qh_half_h + ((base_h + i) % 32);
|
||||
const int qh_chunk_h = qh_byte_h / 8;
|
||||
const int qh_pos_h = qh_byte_h % 8;
|
||||
const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
|
||||
const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
|
||||
|
||||
const int q_l = ((hi_2_l << 4) | l_4) - 32;
|
||||
const int q_h = ((hi_2_h << 4) | hi_4) - 32;
|
||||
|
||||
const int8_t a_l = a_ptr[l].qs[base_l + i];
|
||||
const int8_t a_h = a_ptr[l].qs[base_h + i];
|
||||
|
||||
sumi_l += q_l * a_l;
|
||||
sumi_h += q_h * a_h;
|
||||
}
|
||||
|
||||
sumf[j] +=
|
||||
(sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
s[x * ncols_interleaved + j] = sumf[j];
|
||||
}
|
||||
}
|
||||
ggml_gemv_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
@@ -1485,109 +1597,12 @@ void ggml_gemm_q5_K_8x8_q8_K_generic(int n,
|
||||
}
|
||||
}
|
||||
|
||||
void ggml_gemm_q6_K_8x8_q8_K_generic(int n,
|
||||
float * GGML_RESTRICT s,
|
||||
size_t bs,
|
||||
const void * GGML_RESTRICT vx,
|
||||
const void * GGML_RESTRICT vy,
|
||||
int nr,
|
||||
int nc) {
|
||||
const int qk = QK_K;
|
||||
const int nb = n / qk;
|
||||
const int ncols_interleaved = 8;
|
||||
const int blocklen = 8;
|
||||
void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
ggml_gemm_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
assert(n % qk == 0);
|
||||
assert(nr % 4 == 0);
|
||||
assert(nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(bs);
|
||||
|
||||
float sumf[4][8];
|
||||
|
||||
for (int y = 0; y < nr / 4; y++) {
|
||||
const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
|
||||
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
sumf[m][j] = 0.0f;
|
||||
}
|
||||
}
|
||||
|
||||
for (int l = 0; l < nb; l++) {
|
||||
for (int k = 0; k < 16; k++) {
|
||||
// k = 0.. 7 weights 0-63 low, 64-127 high
|
||||
// k = 8..15 weights 128-191 low, 192-255 high
|
||||
const int base_l = (k / 8) * 128 + (k % 8) * 8;
|
||||
const int base_h = base_l + 64;
|
||||
|
||||
const int scale_idx_l = base_l / 16;
|
||||
const int scale_idx_h = base_h / 16;
|
||||
|
||||
// Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
|
||||
const int qh_shift_l = ((base_l % 128) / 32) * 2;
|
||||
const int qh_shift_h = ((base_h % 128) / 32) * 2;
|
||||
|
||||
// qh_half: offset to the correct 32-byte half (0 or 32)
|
||||
const int qh_half_l = (base_l / 128) * 32;
|
||||
const int qh_half_h = (base_h / 128) * 32;
|
||||
|
||||
// Activation base indices for q8_Kx4 interleaved format
|
||||
// Layout: 128-value halves (k/8), then 8-value sub-blocks (k%8) with stride 32
|
||||
const int q8_base = (k / 8) * 512 + (k % 8) * 32;
|
||||
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
// Interleaved scales
|
||||
const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
|
||||
const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
|
||||
|
||||
int sumi_l = 0;
|
||||
int sumi_h = 0;
|
||||
|
||||
for (int i = 0; i < blocklen; i++) {
|
||||
const int ql_pos = k * 64 + j * 8 + i;
|
||||
const int l_4 = b_ptr[l].ql[ql_pos] & 0xF;
|
||||
const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
|
||||
|
||||
const int qh_idx_l = qh_half_l + ((base_l + i) % 32);
|
||||
const int qh_chunk_l = qh_idx_l / 8;
|
||||
const int qh_pos_l = qh_idx_l % 8;
|
||||
const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
|
||||
const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
|
||||
|
||||
const int qh_idx_h = qh_half_h + ((base_h + i) % 32);
|
||||
const int qh_chunk_h = qh_idx_h / 8;
|
||||
const int qh_pos_h = qh_idx_h % 8;
|
||||
const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
|
||||
const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
|
||||
|
||||
const int q_l = ((hi_2_l << 4) | l_4) - 32;
|
||||
const int q_h = ((hi_2_h << 4) | hi_4) - 32;
|
||||
|
||||
const int8_t q8_l = a_ptr[l].qs[q8_base + m * 8 + i];
|
||||
const int8_t q8_h = a_ptr[l].qs[q8_base + m * 8 + i + 256];
|
||||
|
||||
sumi_l += q_l * q8_l;
|
||||
sumi_h += q_h * q8_h;
|
||||
}
|
||||
|
||||
sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
|
||||
a_ptr[l].d[m];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
ggml_gemm_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
@@ -2097,18 +2112,18 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
|
||||
}
|
||||
|
||||
const int end_ls = QK_K * 4 / blck_size_interleave;
|
||||
// Interleave Q6_K quants by taking 8 bytes at a time
|
||||
// Interleave Q6_K quants by taking blck_size_interleave bytes at a time
|
||||
for (int i = 0; i < end_ls; ++i) {
|
||||
int src_id = i % n_blocks;
|
||||
int src_offset = (i / n_blocks) * blck_size_interleave;
|
||||
int dst_offset = i * blck_size_interleave;
|
||||
|
||||
uint64_t elem_ls;
|
||||
memcpy(&elem_ls, &in[src_id].ql[src_offset], sizeof(uint64_t));
|
||||
memcpy(&out.ql[dst_offset], &elem_ls, sizeof(uint64_t));
|
||||
memcpy(&elem_ls, &in[src_id].ql[src_offset], blck_size_interleave);
|
||||
memcpy(&out.ql[dst_offset], &elem_ls, blck_size_interleave);
|
||||
}
|
||||
|
||||
// Interleave high bits using same 8-byte pattern as low bits
|
||||
// Interleave high bits using same chunk size as low bits
|
||||
const int end_hs = end_ls / 2;
|
||||
for (int i = 0; i < end_hs; ++i) {
|
||||
int src_id = i % n_blocks;
|
||||
@@ -2116,8 +2131,8 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
|
||||
int dst_offset = i * blck_size_interleave;
|
||||
|
||||
uint64_t elem_hs;
|
||||
memcpy(&elem_hs, &in[src_id].qh[src_offset], sizeof(uint64_t));
|
||||
memcpy(&out.qh[dst_offset], &elem_hs, sizeof(uint64_t));
|
||||
memcpy(&elem_hs, &in[src_id].qh[src_offset], blck_size_interleave);
|
||||
memcpy(&out.qh[dst_offset], &elem_hs, blck_size_interleave);
|
||||
}
|
||||
|
||||
// The below logic is designed so as to unpack and rearrange scales in Q6_K
|
||||
@@ -2262,7 +2277,7 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor * t,
|
||||
|
||||
static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
|
||||
GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
|
||||
GGML_ASSERT(interleave_block == 8);
|
||||
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
|
||||
constexpr int nrows_interleaved = 8;
|
||||
|
||||
block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data;
|
||||
@@ -2511,6 +2526,10 @@ template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * da
|
||||
return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
|
||||
}
|
||||
|
||||
template <> int repack<block_q6_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
|
||||
return repack_q6_K_to_q6_K_8_bl(t, 4, data, data_size);
|
||||
}
|
||||
|
||||
template <> int repack<block_q6_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
|
||||
return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size);
|
||||
}
|
||||
@@ -2575,6 +2594,10 @@ template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
|
||||
ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
template <> void gemv<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
|
||||
ggml_gemv_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
|
||||
ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
@@ -2634,6 +2657,10 @@ template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
|
||||
ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
template <> void gemm<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
|
||||
ggml_gemm_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
|
||||
template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
|
||||
ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
|
||||
}
|
||||
@@ -3043,6 +3070,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
|
||||
static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
|
||||
|
||||
// instance for Q6_K
|
||||
static const ggml::cpu::repack::tensor_traits<block_q6_K, 4, 8, GGML_TYPE_Q8_K> q6_K_8x4_q8_K;
|
||||
static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K;
|
||||
|
||||
// instance for Q2
|
||||
@@ -3107,6 +3135,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
|
||||
return &q6_K_8x8_q8_K;
|
||||
}
|
||||
}
|
||||
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
|
||||
if (cur->ne[1] % 8 == 0) {
|
||||
return &q6_K_8x4_q8_K;
|
||||
}
|
||||
}
|
||||
} else if (cur->type == GGML_TYPE_IQ4_NL) {
|
||||
if (ggml_cpu_has_avx2()) {
|
||||
if (cur->ne[1] % 8 == 0) {
|
||||
|
||||
@@ -112,6 +112,7 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
@@ -122,6 +123,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
@@ -142,6 +144,7 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
|
||||
void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
@@ -152,6 +155,7 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
|
||||
void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
|
||||
@@ -116,6 +116,17 @@ extern "C" {
|
||||
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
|
||||
extern float ggml_table_f32_f16[1 << 16];
|
||||
|
||||
// precomputed f32 table for e8m0 half (1 KB)
|
||||
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
|
||||
extern float ggml_table_f32_e8m0_half[1 << 8];
|
||||
|
||||
// Use lookup table for E8M0 on x86 (faster than bit manipulation)
|
||||
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
||||
#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
|
||||
#else
|
||||
#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
|
||||
#endif
|
||||
|
||||
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
|
||||
// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
|
||||
// This is also true for POWER9.
|
||||
|
||||
@@ -64,7 +64,7 @@ if (CUDAToolkit_FOUND)
|
||||
FetchContent_Declare(
|
||||
CCCL
|
||||
GIT_REPOSITORY https://github.com/nvidia/cccl.git
|
||||
GIT_TAG v3.2.0-rc2
|
||||
GIT_TAG v3.2.0
|
||||
GIT_SHALLOW TRUE
|
||||
)
|
||||
|
||||
|
||||
@@ -39,13 +39,16 @@ static __global__ void k_bin_bcast(const src0_t * src0,
|
||||
const uint3 ne11,
|
||||
const uint3 ne12,
|
||||
const uint3 ne13,
|
||||
/*int s0, */ const int s1,
|
||||
/*const int s0,*/
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
/*int s00,*/ const int s01,
|
||||
const int s00,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
/*int s10,*/ const int s11,
|
||||
const int s10,
|
||||
const int s11,
|
||||
const int s12,
|
||||
const int s13,
|
||||
src1_ptrs... src1s) {
|
||||
@@ -72,11 +75,11 @@ static __global__ void k_bin_bcast(const src0_t * src0,
|
||||
for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
|
||||
const uint32_t i10 = fastmodulo(i0, ne10);
|
||||
|
||||
float result = src0_row ? (float) src0_row[i0] : 0.0f;
|
||||
float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
|
||||
if constexpr (sizeof...(src1_ptrs) > 0) {
|
||||
result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
|
||||
result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
|
||||
} else {
|
||||
result = bin_op(result, (float)src1[i_src1 + i10]);
|
||||
result = bin_op(result, (float)src1[i_src1 + i10*s10]);
|
||||
}
|
||||
|
||||
dst_row[i0] = (dst_t) result;
|
||||
@@ -101,13 +104,16 @@ static __global__ void k_bin_bcast_unravel(const src0_t * src0,
|
||||
const uint3 ne11,
|
||||
const uint3 ne12,
|
||||
const uint3 ne13,
|
||||
/*int s0, */ const int s1,
|
||||
/*const int s0,*/
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
/*int s00,*/ const int s01,
|
||||
const int s00,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
/*int s10,*/ const int s11,
|
||||
const int s10,
|
||||
const int s11,
|
||||
const int s12,
|
||||
const int s13,
|
||||
src1_ptrs... src1s) {
|
||||
@@ -135,11 +141,11 @@ static __global__ void k_bin_bcast_unravel(const src0_t * src0,
|
||||
|
||||
const int i10 = fastmodulo(i0, ne10);
|
||||
|
||||
float result = src0_row ? (float) src0_row[i0] : 0.0f;
|
||||
float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
|
||||
if constexpr (sizeof...(src1_ptrs) > 0) {
|
||||
result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
|
||||
result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
|
||||
} else {
|
||||
result = bin_op(result, (float)src1[i_src1 + i10]);
|
||||
result = bin_op(result, (float)src1[i_src1 + i10*s10]);
|
||||
}
|
||||
|
||||
dst_row[i0] = (dst_t) result;
|
||||
@@ -179,7 +185,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
|
||||
cnb[3] *= cne[3];
|
||||
};
|
||||
|
||||
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
|
||||
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
|
||||
for (int i = 0; i < 4; i++) {
|
||||
if (nr[i] != 1) {
|
||||
break;
|
||||
@@ -221,7 +227,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
|
||||
size_t nb12 = cnb1[2];
|
||||
size_t nb13 = cnb1[3];
|
||||
|
||||
size_t s0 = nb0 / sizeof(dst_t);
|
||||
//size_t s0 = nb0 / sizeof(dst_t);
|
||||
size_t s1 = nb1 / sizeof(dst_t);
|
||||
size_t s2 = nb2 / sizeof(dst_t);
|
||||
size_t s3 = nb3 / sizeof(dst_t);
|
||||
@@ -251,10 +257,6 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
|
||||
GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
|
||||
GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
|
||||
|
||||
GGML_ASSERT(s0 == 1);
|
||||
GGML_ASSERT(s00 == 1);
|
||||
GGML_ASSERT(s10 == 1);
|
||||
|
||||
const int block_size = 128;
|
||||
|
||||
int64_t hne0 = std::max(ne0 / 2LL, 1LL);
|
||||
@@ -284,31 +286,31 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
|
||||
k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t><<<block_num, block_size, 0, stream>>>(
|
||||
src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
|
||||
ne12, ne13,
|
||||
/* s0, */ s1, s2, s3,
|
||||
/* s00,*/ s01, s02, s03,
|
||||
/* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
|
||||
/*s0,*/ s1, s2, s3,
|
||||
s00, s01, s02, s03,
|
||||
s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
|
||||
} else {
|
||||
k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
|
||||
<<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv,
|
||||
ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11, ne12, ne13,
|
||||
/* s0, */ s1, s2, s3,
|
||||
/* s00,*/ s01, s02, s03,
|
||||
/* s10,*/ s11, s12, s13);
|
||||
/*s0,*/ s1, s2, s3,
|
||||
s00, s01, s02, s03,
|
||||
s10, s11, s12, s13);
|
||||
}
|
||||
} else {
|
||||
const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
|
||||
if constexpr (sizeof...(I) > 0) {
|
||||
k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
|
||||
src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
|
||||
/* s0, */ s1, s2, s3,
|
||||
/* s00,*/ s01, s02, s03,
|
||||
/* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
|
||||
/*s0,*/ s1, s2, s3,
|
||||
s00 ,s01, s02, s03,
|
||||
s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
|
||||
} else {
|
||||
k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
|
||||
src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
|
||||
/* s0, */ s1, s2, s3,
|
||||
/* s00,*/ s01, s02, s03,
|
||||
/* s10,*/ s11, s12, s13);
|
||||
/*s0,*/ s1, s2, s3,
|
||||
s00, s01, s02, s03,
|
||||
s10, s11, s12, s13);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -2979,8 +2979,7 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
|
||||
}
|
||||
}
|
||||
|
||||
if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
|
||||
memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
|
||||
if (memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
|
||||
return false;
|
||||
}
|
||||
|
||||
@@ -4835,8 +4834,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
|
||||
case GGML_OP_SUM_ROWS:
|
||||
case GGML_OP_MEAN:
|
||||
case GGML_OP_GROUP_NORM:
|
||||
case GGML_OP_PAD:
|
||||
return ggml_is_contiguous(op->src[0]);
|
||||
case GGML_OP_PAD:
|
||||
return true;
|
||||
case GGML_OP_UPSCALE:
|
||||
case GGML_OP_PAD_REFLECT_1D:
|
||||
case GGML_OP_ARANGE:
|
||||
|
||||
@@ -7,7 +7,7 @@ __device__ __forceinline__ int64_t wrap_around(int64_t coord, int64_t size) {
|
||||
return (coord + size) % size;
|
||||
}
|
||||
|
||||
static __global__ void pad_f32(const float * src, float * dst,
|
||||
static __global__ void pad_f32(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
|
||||
const int lp0, const int rp0, const int lp1, const int rp1,
|
||||
const int lp2, const int rp2, const int lp3, const int rp3,
|
||||
const int ne0, const int ne1, const int ne2, const int ne3,
|
||||
@@ -34,11 +34,8 @@ static __global__ void pad_f32(const float * src, float * dst,
|
||||
const int64_t i01 = i1 - lp1;
|
||||
const int64_t i02 = i2 - lp2;
|
||||
const int64_t i03 = i3 - lp3;
|
||||
const int64_t ne02 = ne2 - lp2 - rp2;
|
||||
const int64_t ne01 = ne1 - lp1 - rp1;
|
||||
const int64_t ne00 = ne0 - lp0 - rp0;
|
||||
|
||||
const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
|
||||
const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
|
||||
|
||||
dst[dst_idx] = src[src_idx];
|
||||
} else {
|
||||
@@ -57,21 +54,21 @@ static __global__ void pad_f32(const float * src, float * dst,
|
||||
const int64_t i02 = wrap_around(i2 - lp2, ne02);
|
||||
const int64_t i03 = wrap_around(i3 - lp3, ne03);
|
||||
|
||||
const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
|
||||
const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
|
||||
|
||||
dst[dst_idx] = src[src_idx];
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
static void pad_f32_cuda(const float * src, float * dst,
|
||||
static void pad_f32_cuda(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
|
||||
const int lp0, const int rp0, const int lp1, const int rp1,
|
||||
const int lp2, const int rp2, const int lp3, const int rp3,
|
||||
const int ne0, const int ne1, const int ne2, const int ne3,
|
||||
const bool circular, cudaStream_t stream) {
|
||||
int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
|
||||
dim3 gridDim(num_blocks, ne1, ne2 * ne3);
|
||||
pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst,
|
||||
pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, s00, s01, s02, s03, dst,
|
||||
lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
|
||||
ne0, ne1, ne2, ne3, circular);
|
||||
}
|
||||
@@ -82,9 +79,10 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
||||
float * dst_d = (float *) dst->data;
|
||||
cudaStream_t stream = ctx.stream();
|
||||
|
||||
GGML_TENSOR_UNARY_OP_LOCALS;
|
||||
|
||||
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
||||
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
||||
GGML_ASSERT(ggml_is_contiguous(src0));
|
||||
|
||||
const int32_t lp0 = ((const int32_t *) (dst->op_params))[0];
|
||||
const int32_t rp0 = ((const int32_t *) (dst->op_params))[1];
|
||||
@@ -96,7 +94,12 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
||||
const int32_t rp3 = ((const int32_t *) (dst->op_params))[7];
|
||||
const int32_t circular = ((const int32_t *) (dst->op_params))[8];
|
||||
|
||||
pad_f32_cuda(src0_d, dst_d,
|
||||
const size_t s00 = nb00 / ggml_type_size(src0->type);
|
||||
const size_t s01 = nb01 / ggml_type_size(src0->type);
|
||||
const size_t s02 = nb02 / ggml_type_size(src0->type);
|
||||
const size_t s03 = nb03 / ggml_type_size(src0->type);
|
||||
|
||||
pad_f32_cuda(src0_d, s00, s01, s02, s03, dst_d,
|
||||
lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
|
||||
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
|
||||
(bool) circular, stream);
|
||||
|
||||
@@ -43,10 +43,15 @@ static __device__ void rope_yarn(
|
||||
template <bool forward, bool has_ff, typename T, typename D>
|
||||
static __global__ void rope_norm(const T * x,
|
||||
D * dst,
|
||||
const int ne0,
|
||||
const int ne1,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int32_t * pos,
|
||||
const float freq_scale,
|
||||
@@ -59,23 +64,23 @@ static __global__ void rope_norm(const T * x,
|
||||
const int set_rows_stride) {
|
||||
const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
|
||||
|
||||
if (i0 >= ne0) {
|
||||
if (i0 >= ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
|
||||
|
||||
const int row_x = row_dst % ne1;
|
||||
const int channel_x = row_dst / ne1;
|
||||
|
||||
int idst = row_dst * ne0 + i0;
|
||||
const int ix = channel_x*s2 + row_x*s1 + i0;
|
||||
const uint32_t i3 = row_dst / (ne01 * ne02);
|
||||
const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
|
||||
const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
|
||||
|
||||
int idst = i0 + i1 * s1 + i2 * s2 + i3 * s3;
|
||||
const int ix = i0 + i1 * s01 + i2 * s02 + i3 * s03;
|
||||
// Fusion optimization: ROPE + VIEW + SET_ROWS.
|
||||
// The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
|
||||
if (set_rows_stride != 0) {
|
||||
idst = row_x * ne0 + i0;
|
||||
idst += row_indices[channel_x] * set_rows_stride;
|
||||
idst = i1 * s1 + i0;
|
||||
idst += row_indices[i2] * set_rows_stride;
|
||||
}
|
||||
|
||||
const auto & store_coaelsced = [&](float x0, float x1) {
|
||||
@@ -92,7 +97,7 @@ static __global__ void rope_norm(const T * x,
|
||||
return;
|
||||
}
|
||||
|
||||
const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
|
||||
const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
|
||||
|
||||
const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
|
||||
|
||||
@@ -110,10 +115,15 @@ static __global__ void rope_norm(const T * x,
|
||||
template <bool forward, bool has_ff, typename T, typename D>
|
||||
static __global__ void rope_neox(const T * x,
|
||||
D * dst,
|
||||
const int ne0,
|
||||
const int ne1,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int32_t * pos,
|
||||
const float freq_scale,
|
||||
@@ -126,23 +136,24 @@ static __global__ void rope_neox(const T * x,
|
||||
const int set_rows_stride) {
|
||||
const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
|
||||
|
||||
if (i0 >= ne0) {
|
||||
if (i0 >= ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
|
||||
|
||||
const int row_x = row_dst % ne1;
|
||||
const int channel_x = row_dst / ne1;
|
||||
const uint32_t i3 = row_dst / (ne01 * ne02);
|
||||
const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
|
||||
const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
|
||||
|
||||
int idst = row_dst * ne0 + i0 / 2;
|
||||
const int ix = channel_x*s2 + row_x*s1 + i0/2;
|
||||
int idst = i0 / 2 + i1 * s1 + i2 * s2 + i3 * s3;
|
||||
const int ix = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
|
||||
|
||||
// Fusion optimization: ROPE + VIEW + SET_ROWS.
|
||||
// The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
|
||||
if (set_rows_stride != 0) {
|
||||
idst = row_x * ne0 + i0 / 2;
|
||||
idst += row_indices[channel_x] * set_rows_stride;
|
||||
idst = i1 * s1 + i0 / 2;
|
||||
idst += row_indices[i2] * set_rows_stride;
|
||||
}
|
||||
|
||||
if (i0 >= n_dims) {
|
||||
@@ -152,7 +163,7 @@ static __global__ void rope_neox(const T * x,
|
||||
return;
|
||||
}
|
||||
|
||||
const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
|
||||
const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
|
||||
|
||||
const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
|
||||
|
||||
@@ -168,24 +179,42 @@ static __global__ void rope_neox(const T * x,
|
||||
dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
|
||||
}
|
||||
|
||||
template<bool forward, bool has_ff, typename T>
|
||||
static __global__ void rope_multi(
|
||||
const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
|
||||
const int n_dims, const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
|
||||
const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections, const bool is_imrope) {
|
||||
const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
|
||||
template <bool forward, bool has_ff, typename T>
|
||||
static __global__ void rope_multi(const T * x,
|
||||
T * dst,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int32_t * pos,
|
||||
const float freq_scale,
|
||||
const float ext_factor,
|
||||
const float attn_factor,
|
||||
const rope_corr_dims corr_dims,
|
||||
const float theta_scale,
|
||||
const float * freq_factors,
|
||||
const mrope_sections sections,
|
||||
const bool is_imrope) {
|
||||
const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);
|
||||
|
||||
if (i0 >= ne0) {
|
||||
if (i0 >= ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
|
||||
|
||||
const int row_x = row_dst % ne1;
|
||||
const int channel_x = row_dst / ne1;
|
||||
const uint32_t i3 = row_dst / (ne01 * ne02);
|
||||
const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
|
||||
const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
|
||||
|
||||
const int idst = row_dst*ne0 + i0/2;
|
||||
const int ix = channel_x*s2 + row_x*s1 + i0/2;
|
||||
int idst = i0 / 2 + i1 * s1 + i2 * s2 + i3 * s3;
|
||||
const int ix = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
|
||||
|
||||
if (i0 >= n_dims) {
|
||||
dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
|
||||
@@ -200,27 +229,24 @@ static __global__ void rope_multi(
|
||||
|
||||
float theta_base = 0.0;
|
||||
if (is_imrope) {
|
||||
if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
|
||||
theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
|
||||
} else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
|
||||
theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
|
||||
} else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
|
||||
theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
|
||||
if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
|
||||
theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
|
||||
} else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
|
||||
theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
|
||||
} else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
|
||||
theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
|
||||
} else {
|
||||
theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
|
||||
theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
|
||||
}
|
||||
} else {
|
||||
if (sector < sections.v[0]) {
|
||||
theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
|
||||
}
|
||||
else if (sector >= sections.v[0] && sector < sec_w) {
|
||||
theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
|
||||
}
|
||||
else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
|
||||
theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
|
||||
}
|
||||
else if (sector >= sec_w + sections.v[2]) {
|
||||
theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
|
||||
theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
|
||||
} else if (sector >= sections.v[0] && sector < sec_w) {
|
||||
theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
|
||||
} else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
|
||||
theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
|
||||
} else if (sector >= sec_w + sections.v[2]) {
|
||||
theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -238,37 +264,53 @@ static __global__ void rope_multi(
|
||||
dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
|
||||
}
|
||||
|
||||
template<bool forward, bool has_ff, typename T>
|
||||
static __global__ void rope_vision(
|
||||
const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims,
|
||||
const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
|
||||
const float theta_scale, const float * freq_factors, const mrope_sections sections) {
|
||||
template <bool forward, bool has_ff, typename T>
|
||||
static __global__ void rope_vision(const T * x,
|
||||
T * dst,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int32_t * pos,
|
||||
const float freq_scale,
|
||||
const float ext_factor,
|
||||
const float attn_factor,
|
||||
const rope_corr_dims corr_dims,
|
||||
const float theta_scale,
|
||||
const float * freq_factors,
|
||||
const mrope_sections sections) {
|
||||
const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
|
||||
|
||||
if (i0 >= ne0) {
|
||||
if (i0 >= ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
|
||||
|
||||
const int row_x = row_dst % ne1;
|
||||
const int channel_x = row_dst / ne1;
|
||||
const uint32_t i3 = row_dst / (ne01 * ne02);
|
||||
const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
|
||||
const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
|
||||
|
||||
const int idst = row_dst*ne0 + i0/2;
|
||||
const int ix = channel_x*s2 + row_x*s1 + i0/2;
|
||||
int idst = i0 / 2 + i1 * s1 + i2 * s2 + i3 * s3;
|
||||
const int ix = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
|
||||
|
||||
const int sect_dims = sections.v[0] + sections.v[1];
|
||||
const int sec_w = sections.v[1] + sections.v[0];
|
||||
const int sector = (i0 / 2) % sect_dims;
|
||||
const int sec_w = sections.v[1] + sections.v[0];
|
||||
const int sector = (i0 / 2) % sect_dims;
|
||||
|
||||
float theta_base = 0.0;
|
||||
if (sector < sections.v[0]) {
|
||||
const int p = sector;
|
||||
theta_base = pos[channel_x]*powf(theta_scale, p);
|
||||
}
|
||||
else if (sector >= sections.v[0] && sector < sec_w) {
|
||||
theta_base = pos[i2] * powf(theta_scale, p);
|
||||
} else if (sector >= sections.v[0] && sector < sec_w) {
|
||||
const int p = sector - sections.v[0];
|
||||
theta_base = pos[channel_x + ne2]*powf(theta_scale, p);
|
||||
theta_base = pos[i2 + ne02] * powf(theta_scale, p);
|
||||
}
|
||||
|
||||
const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
|
||||
@@ -288,10 +330,15 @@ static __global__ void rope_vision(
|
||||
template <bool forward, typename T, typename D>
|
||||
static void rope_norm_cuda(const T * x,
|
||||
D * dst,
|
||||
const int ne0,
|
||||
const int ne1,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int nr,
|
||||
const int32_t * pos,
|
||||
@@ -304,31 +351,36 @@ static void rope_norm_cuda(const T * x,
|
||||
const int64_t * row_indices,
|
||||
const int set_rows_stride,
|
||||
cudaStream_t stream) {
|
||||
GGML_ASSERT(ne0 % 2 == 0);
|
||||
GGML_ASSERT(ne00 % 2 == 0);
|
||||
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
|
||||
const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
|
||||
const int n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
|
||||
const dim3 block_nums(nr, n_blocks_x, 1);
|
||||
|
||||
const float theta_scale = powf(freq_base, -2.0f/n_dims);
|
||||
const float theta_scale = powf(freq_base, -2.0f / n_dims);
|
||||
|
||||
if (freq_factors == nullptr) {
|
||||
rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
|
||||
freq_factors, row_indices, set_rows_stride);
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
|
||||
} else {
|
||||
rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
|
||||
freq_factors, row_indices, set_rows_stride);
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
|
||||
}
|
||||
}
|
||||
|
||||
template <bool forward, typename T, typename D>
|
||||
static void rope_neox_cuda(const T * x,
|
||||
D * dst,
|
||||
const int ne0,
|
||||
const int ne1,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int nr,
|
||||
const int32_t * pos,
|
||||
@@ -341,55 +393,92 @@ static void rope_neox_cuda(const T * x,
|
||||
const int64_t * row_indices,
|
||||
const int set_rows_stride,
|
||||
cudaStream_t stream) {
|
||||
GGML_ASSERT(ne0 % 2 == 0);
|
||||
GGML_ASSERT(ne00 % 2 == 0);
|
||||
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
|
||||
const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
|
||||
const int n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
|
||||
const dim3 block_nums(nr, n_blocks_x, 1);
|
||||
|
||||
const float theta_scale = powf(freq_base, -2.0f/n_dims);
|
||||
const float theta_scale = powf(freq_base, -2.0f / n_dims);
|
||||
|
||||
if (freq_factors == nullptr) {
|
||||
rope_neox<forward, false><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
|
||||
freq_factors, row_indices, set_rows_stride);
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
|
||||
} else {
|
||||
rope_neox<forward, true><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
|
||||
freq_factors, row_indices, set_rows_stride);
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
|
||||
}
|
||||
}
|
||||
|
||||
template<bool forward, typename T>
|
||||
static void rope_multi_cuda(
|
||||
const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
|
||||
const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
|
||||
const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, const bool is_imrope, cudaStream_t stream) {
|
||||
GGML_ASSERT(ne0 % 2 == 0);
|
||||
template <bool forward, typename T>
|
||||
static void rope_multi_cuda(const T * x,
|
||||
T * dst,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int nr,
|
||||
const int32_t * pos,
|
||||
const float freq_scale,
|
||||
const float freq_base,
|
||||
const float ext_factor,
|
||||
const float attn_factor,
|
||||
const rope_corr_dims corr_dims,
|
||||
const float * freq_factors,
|
||||
const mrope_sections sections,
|
||||
const bool is_imrope,
|
||||
cudaStream_t stream) {
|
||||
GGML_ASSERT(ne00 % 2 == 0);
|
||||
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
|
||||
const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
|
||||
const int n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
|
||||
const dim3 block_nums(nr, n_blocks_x, 1);
|
||||
|
||||
const float theta_scale = powf(freq_base, -2.0f/n_dims);
|
||||
const float theta_scale = powf(freq_base, -2.0f / n_dims);
|
||||
|
||||
if (freq_factors == nullptr) {
|
||||
rope_multi<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
|
||||
} else {
|
||||
rope_multi<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
|
||||
}
|
||||
}
|
||||
|
||||
template<bool forward, typename T>
|
||||
static void rope_vision_cuda(
|
||||
const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
|
||||
const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
|
||||
const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
|
||||
GGML_ASSERT(ne0 % 2 == 0);
|
||||
template <bool forward, typename T>
|
||||
static void rope_vision_cuda(const T * x,
|
||||
T * dst,
|
||||
const int ne00,
|
||||
const int ne01,
|
||||
const int ne02,
|
||||
const int s01,
|
||||
const int s02,
|
||||
const int s03,
|
||||
const int s1,
|
||||
const int s2,
|
||||
const int s3,
|
||||
const int n_dims,
|
||||
const int nr,
|
||||
const int32_t * pos,
|
||||
const float freq_scale,
|
||||
const float freq_base,
|
||||
const float ext_factor,
|
||||
const float attn_factor,
|
||||
const rope_corr_dims corr_dims,
|
||||
const float * freq_factors,
|
||||
const mrope_sections sections,
|
||||
cudaStream_t stream) {
|
||||
GGML_ASSERT(ne00 % 2 == 0);
|
||||
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
|
||||
const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
|
||||
const int n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
|
||||
const dim3 block_nums(nr, n_blocks_x, 1);
|
||||
// break down (head_dim, heads, seq) into (CUDA_ROPE_BLOCK_SIZE, x, heads * seq)
|
||||
// where x ~= ceil(head_dim / CUDA_ROPE_BLOCK_SIZE);
|
||||
@@ -398,11 +487,11 @@ static void rope_vision_cuda(
|
||||
|
||||
if (freq_factors == nullptr) {
|
||||
rope_vision<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, sections);
|
||||
} else {
|
||||
rope_vision<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
|
||||
x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
|
||||
x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
|
||||
attn_factor, corr_dims, theta_scale, freq_factors, sections);
|
||||
}
|
||||
}
|
||||
@@ -445,6 +534,11 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
|
||||
|
||||
const size_t s01 = src0->nb[1] / ggml_type_size(src0->type);
|
||||
const size_t s02 = src0->nb[2] / ggml_type_size(src0->type);
|
||||
const size_t s03 = src0->nb[3] / ggml_type_size(src0->type);
|
||||
|
||||
const size_t s1 = dst->nb[1] / ggml_type_size(dst->type);
|
||||
const size_t s2 = dst->nb[2] / ggml_type_size(dst->type);
|
||||
const size_t s3 = dst->nb[3] / ggml_type_size(dst->type);
|
||||
|
||||
//const int n_past = ((int32_t *) dst->op_params)[0];
|
||||
const int n_dims = ((int32_t *) dst->op_params)[1];
|
||||
@@ -495,57 +589,63 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
|
||||
// compute
|
||||
if (is_neox) {
|
||||
if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
|
||||
rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
|
||||
nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
|
||||
freq_factors, row_indices, set_rows_stride, stream);
|
||||
rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
|
||||
s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
|
||||
ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
|
||||
set_rows_stride, stream);
|
||||
} else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
|
||||
rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
|
||||
nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
|
||||
freq_factors, row_indices, set_rows_stride, stream);
|
||||
rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
|
||||
s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
|
||||
ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
|
||||
set_rows_stride, stream);
|
||||
} else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
|
||||
rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
|
||||
pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
|
||||
freq_factors, row_indices, set_rows_stride, stream);
|
||||
rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
|
||||
s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
|
||||
ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
|
||||
set_rows_stride, stream);
|
||||
} else {
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
} else if (is_mrope && !is_vision) {
|
||||
if (src0->type == GGML_TYPE_F32) {
|
||||
rope_multi_cuda<forward>(
|
||||
(const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
|
||||
freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
|
||||
rope_multi_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
|
||||
s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
|
||||
corr_dims, freq_factors, sections, is_imrope, stream);
|
||||
} else if (src0->type == GGML_TYPE_F16) {
|
||||
rope_multi_cuda<forward>(
|
||||
(const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
|
||||
freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
|
||||
rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
|
||||
s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
|
||||
corr_dims, freq_factors, sections, is_imrope, stream);
|
||||
} else {
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
} else if (is_vision) {
|
||||
if (src0->type == GGML_TYPE_F32) {
|
||||
rope_vision_cuda<forward>(
|
||||
(const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
|
||||
freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
|
||||
rope_vision_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
|
||||
s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
|
||||
corr_dims, freq_factors, sections, stream);
|
||||
} else if (src0->type == GGML_TYPE_F16) {
|
||||
rope_vision_cuda<forward>(
|
||||
(const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
|
||||
freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
|
||||
rope_vision_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
|
||||
s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
|
||||
corr_dims, freq_factors, sections, stream);
|
||||
} else {
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
} else {
|
||||
if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
|
||||
rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
|
||||
nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
|
||||
freq_factors, row_indices, set_rows_stride, stream);
|
||||
rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
|
||||
s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
|
||||
ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
|
||||
set_rows_stride, stream);
|
||||
} else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
|
||||
rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
|
||||
nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
|
||||
freq_factors, row_indices, set_rows_stride, stream);
|
||||
rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
|
||||
s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
|
||||
ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
|
||||
set_rows_stride, stream);
|
||||
} else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
|
||||
rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
|
||||
pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
|
||||
freq_factors, row_indices, set_rows_stride, stream);
|
||||
rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
|
||||
s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
|
||||
ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
|
||||
set_rows_stride, stream);
|
||||
} else {
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
|
||||
@@ -1935,11 +1935,6 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
|
||||
return false;
|
||||
}
|
||||
|
||||
// TODO: add support for non-contigiuos tensors
|
||||
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
@@ -1991,6 +1986,25 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool ggml_hexagon_supported_sum_rows(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
|
||||
const struct ggml_tensor * src0 = op->src[0];
|
||||
const struct ggml_tensor * dst = op;
|
||||
|
||||
if (!hex_supported_src0_type(src0->type)) {
|
||||
return false;
|
||||
}
|
||||
if (!hex_supported_dst_type(dst->type)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// TODO: add support for non-contigiuos tensors
|
||||
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(dst)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session * sess,
|
||||
const struct ggml_tensor * op) {
|
||||
const struct ggml_tensor * src0 = op->src[0];
|
||||
@@ -2111,6 +2125,26 @@ static bool ggml_hexagon_supported_get_rows(const struct ggml_hexagon_session *
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
|
||||
const struct ggml_tensor * src0 = op->src[0]; // values
|
||||
const struct ggml_tensor * dst = op; // indices
|
||||
|
||||
if (src0->type != GGML_TYPE_F32) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (dst->type != GGML_TYPE_I32) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (src0->ne[0] > (16*1024)) {
|
||||
// reject tensors with huge rows for now
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
|
||||
const int32_t * op_params = &op->op_params[0];
|
||||
|
||||
@@ -2278,6 +2312,9 @@ static inline size_t init_binary_req(htp_general_req * req, dspqueue_buffer * bu
|
||||
case GGML_OP_SUB:
|
||||
req->op = HTP_OP_SUB;
|
||||
break;
|
||||
case GGML_OP_DIV:
|
||||
req->op = HTP_OP_DIV;
|
||||
break;
|
||||
default:
|
||||
GGML_ABORT("ggml-hex: binary : unsupported op: %d\n", t->op);
|
||||
break;
|
||||
@@ -2316,6 +2353,17 @@ static inline size_t init_get_rows_req(htp_general_req * req, dspqueue_buffer *
|
||||
return n_bufs;
|
||||
}
|
||||
|
||||
static inline size_t init_argsort_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
|
||||
req->op = HTP_OP_ARGSORT;
|
||||
memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
|
||||
|
||||
size_t n_bufs = 0;
|
||||
n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
|
||||
n_bufs += htp_req_buff_init(&req->dst, &bufs[n_bufs], t, DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
|
||||
|
||||
return n_bufs;
|
||||
}
|
||||
|
||||
template <bool _is_src0_constant>
|
||||
static inline size_t init_binary_id_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
|
||||
switch (t->op) {
|
||||
@@ -2370,6 +2418,16 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
|
||||
supported = true;
|
||||
break;
|
||||
|
||||
case GGML_OP_SQR:
|
||||
req->op = HTP_OP_SQR;
|
||||
supported = true;
|
||||
break;
|
||||
|
||||
case GGML_OP_SQRT:
|
||||
req->op = HTP_OP_SQRT;
|
||||
supported = true;
|
||||
break;
|
||||
|
||||
case GGML_OP_UNARY:
|
||||
if (ggml_get_unary_op(t) == GGML_UNARY_OP_SILU) {
|
||||
req->op = HTP_OP_UNARY_SILU;
|
||||
@@ -2387,6 +2445,9 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
|
||||
} else if (ggml_get_glu_op(t) == GGML_GLU_OP_SWIGLU_OAI) {
|
||||
req->op = HTP_OP_GLU_SWIGLU_OAI;
|
||||
supported = true;
|
||||
} else if (ggml_get_glu_op(t) == GGML_GLU_OP_GEGLU) {
|
||||
req->op = HTP_OP_GLU_GEGLU;
|
||||
supported = true;
|
||||
}
|
||||
break;
|
||||
|
||||
@@ -2411,6 +2472,17 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
|
||||
return n_bufs;
|
||||
}
|
||||
|
||||
static inline size_t init_sum_rows_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
|
||||
memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
|
||||
req->op = HTP_OP_SUM_ROWS;
|
||||
|
||||
size_t n_bufs = 0;
|
||||
n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
|
||||
n_bufs += htp_req_buff_init(&req->dst, &bufs[n_bufs], t, DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
|
||||
|
||||
return n_bufs;
|
||||
}
|
||||
|
||||
static inline size_t init_rope_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
|
||||
memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
|
||||
req->op = HTP_OP_ROPE;
|
||||
@@ -2519,6 +2591,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
|
||||
case GGML_OP_MUL:
|
||||
case GGML_OP_ADD:
|
||||
case GGML_OP_SUB:
|
||||
case GGML_OP_DIV:
|
||||
ggml_hexagon_dispatch_op<init_binary_req<false>>(sess, node, flags);
|
||||
break;
|
||||
case GGML_OP_ADD_ID:
|
||||
@@ -2528,6 +2601,13 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
|
||||
case GGML_OP_SCALE:
|
||||
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
|
||||
break;
|
||||
case GGML_OP_SQR:
|
||||
case GGML_OP_SQRT:
|
||||
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
|
||||
break;
|
||||
case GGML_OP_SUM_ROWS:
|
||||
ggml_hexagon_dispatch_op<init_sum_rows_req>(sess, node, flags);
|
||||
break;
|
||||
case GGML_OP_UNARY:
|
||||
if ((ggml_get_unary_op(node) == GGML_UNARY_OP_SILU) ||
|
||||
(ggml_get_unary_op(node) == GGML_UNARY_OP_GELU)) {
|
||||
@@ -2536,7 +2616,8 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
|
||||
break;
|
||||
case GGML_OP_GLU:
|
||||
if ((ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU) ||
|
||||
(ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI)) {
|
||||
(ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI) ||
|
||||
(ggml_get_glu_op(node) == GGML_GLU_OP_GEGLU)) {
|
||||
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
|
||||
}
|
||||
break;
|
||||
@@ -2564,6 +2645,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
|
||||
ggml_hexagon_dispatch_op<init_cpy_req>(sess, node, flags);
|
||||
break;
|
||||
|
||||
case GGML_OP_ARGSORT:
|
||||
ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
|
||||
break;
|
||||
|
||||
default:
|
||||
GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
|
||||
}
|
||||
@@ -2916,6 +3001,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
|
||||
case GGML_OP_MUL:
|
||||
case GGML_OP_ADD:
|
||||
case GGML_OP_SUB:
|
||||
case GGML_OP_DIV:
|
||||
supp = ggml_hexagon_supported_binary(sess, op);
|
||||
break;
|
||||
|
||||
@@ -2928,6 +3014,15 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
|
||||
supp = ggml_hexagon_supported_unary(sess, op);
|
||||
break;
|
||||
|
||||
case GGML_OP_SQR:
|
||||
case GGML_OP_SQRT:
|
||||
supp = ggml_hexagon_supported_unary(sess, op);
|
||||
break;
|
||||
|
||||
case GGML_OP_SUM_ROWS:
|
||||
supp = ggml_hexagon_supported_sum_rows(sess, op);
|
||||
break;
|
||||
|
||||
case GGML_OP_SOFT_MAX:
|
||||
supp = ggml_hexagon_supported_softmax(sess, op);
|
||||
break;
|
||||
@@ -2943,7 +3038,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
|
||||
case GGML_OP_GLU:
|
||||
{
|
||||
const auto glu_op = ggml_get_glu_op(op);
|
||||
if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI)) {
|
||||
if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI) || (glu_op == GGML_GLU_OP_GEGLU)) {
|
||||
supp = ggml_hexagon_supported_activations(sess, op);
|
||||
}
|
||||
break;
|
||||
@@ -2968,6 +3063,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
|
||||
supp = ggml_hexagon_supported_cpy(sess, op);
|
||||
break;
|
||||
|
||||
case GGML_OP_ARGSORT:
|
||||
supp = ggml_hexagon_supported_argsort(sess, op);
|
||||
break;
|
||||
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
@@ -6,6 +6,7 @@ include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
|
||||
include_directories(
|
||||
${HEXAGON_SDK_ROOT}/incs
|
||||
${HEXAGON_SDK_ROOT}/incs/stddef
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/../../../include
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/../..
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/..
|
||||
${CMAKE_CURRENT_SOURCE_DIR}
|
||||
@@ -21,6 +22,7 @@ add_library(${HTP_LIB} SHARED
|
||||
matmul-ops.c
|
||||
binary-ops.c
|
||||
unary-ops.c
|
||||
sum-rows-ops.c
|
||||
softmax-ops.c
|
||||
act-ops.c
|
||||
rope-ops.c
|
||||
@@ -28,6 +30,7 @@ add_library(${HTP_LIB} SHARED
|
||||
set-rows-ops.c
|
||||
get-rows-ops.c
|
||||
cpy-ops.c
|
||||
argsort-ops.c
|
||||
)
|
||||
|
||||
target_compile_definitions(${HTP_LIB} PRIVATE
|
||||
|
||||
@@ -410,7 +410,7 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
|
||||
// gelu = x * sigmoid(1.702 * x) // current implementation
|
||||
hvx_mul_scalar_f32((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (float) 1.702, ne0);
|
||||
hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
|
||||
hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
|
||||
hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
|
||||
}
|
||||
|
||||
dma_queue_push_vtcm_to_ddr(dma_queue,
|
||||
@@ -516,7 +516,7 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
|
||||
|
||||
// silu = x * sigmoid(x)
|
||||
hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, ne0);
|
||||
hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
|
||||
hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
|
||||
}
|
||||
|
||||
dma_queue_push_vtcm_to_ddr(dma_queue,
|
||||
@@ -541,6 +541,143 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
|
||||
ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
|
||||
}
|
||||
|
||||
static const float GELU_COEF_A = 0.044715f;
|
||||
static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
|
||||
|
||||
static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
|
||||
const struct htp_tensor * src1,
|
||||
struct htp_tensor * dst,
|
||||
const int32_t * op_params,
|
||||
struct htp_spad * src0_spad,
|
||||
struct htp_spad * src1_spad,
|
||||
struct htp_spad * dst_spad,
|
||||
uint32_t nth,
|
||||
uint32_t ith,
|
||||
uint32_t src0_nrows_per_thread,
|
||||
dma_queue * dma_queue) {
|
||||
htp_act_preamble3;
|
||||
|
||||
size_t src0_row_size = nb01;
|
||||
size_t src1_row_size = nb11;
|
||||
size_t dst_row_size = nb1;
|
||||
|
||||
uint64_t t1, t2;
|
||||
t1 = HAP_perf_get_qtimer_count();
|
||||
|
||||
const uint32_t src0_nrows = ne01 * ne02 * ne03; // src0 rows
|
||||
|
||||
const uint32_t src0_start_row = src0_nrows_per_thread * ith;
|
||||
const uint32_t src0_end_row = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
|
||||
|
||||
// no work for this thread
|
||||
if (src0_start_row >= src0_end_row) {
|
||||
return;
|
||||
}
|
||||
|
||||
const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
|
||||
const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
|
||||
uint8_t * restrict data_dst = (uint8_t *) dst->data;
|
||||
|
||||
const bool src1_valid = src1->ne[0];
|
||||
const int nc = (src1_valid) ? ne00 : ne00 / 2;
|
||||
if (!src1_valid) {
|
||||
const int32_t swapped = op_params[1];
|
||||
data_src1 = data_src0;
|
||||
src1_row_size = src0_row_size;
|
||||
|
||||
const size_t nc_in_bytes = nc * SIZEOF_FP32;
|
||||
data_src0 += swapped ? nc_in_bytes : 0;
|
||||
data_src1 += swapped ? 0 : nc_in_bytes;
|
||||
}
|
||||
|
||||
const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
|
||||
const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
|
||||
const size_t dst_row_size_aligned = hex_round_up(dst_row_size, VLEN);
|
||||
|
||||
uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
|
||||
uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
|
||||
uint8_t * restrict dst_spad_data = dst_spad->data + (ith * dst_spad->size_per_thread);
|
||||
|
||||
// While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
|
||||
size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
|
||||
size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
|
||||
size_t dst_spad_half_size = dst_spad->size_per_thread / 2;
|
||||
|
||||
const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block
|
||||
if (BLOCK == 0) {
|
||||
FARF(ERROR,
|
||||
"geglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
|
||||
src0_spad->size_per_thread, src0_row_size_aligned);
|
||||
return;
|
||||
}
|
||||
|
||||
// See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
|
||||
for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
|
||||
const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
|
||||
|
||||
// Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
|
||||
dma_queue_push_vtcm_to_ddr(dma_queue,
|
||||
dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
|
||||
dst_row_size, dst_row_size_aligned, 0);
|
||||
|
||||
dma_queue_push_ddr_to_vtcm(dma_queue,
|
||||
dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)),
|
||||
src0_row_size_aligned, src0_row_size, block_size);
|
||||
dma_queue_push_ddr_to_vtcm(dma_queue,
|
||||
dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_size)),
|
||||
src1_row_size_aligned, src1_row_size, block_size);
|
||||
}
|
||||
|
||||
for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
|
||||
const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
|
||||
|
||||
float * dst_spad = (float *) dma_queue_pop(dma_queue).src;
|
||||
float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;
|
||||
float * src1_spad = (float *) dma_queue_pop(dma_queue).dst;
|
||||
|
||||
for (uint32_t ib = 0; ib < block_size; ib++) {
|
||||
const uint8_t * src0_spad_ptr = (const uint8_t *)(src0_spad + ib * (src0_row_size_aligned / sizeof(float)));
|
||||
const uint8_t * src1_spad_ptr = (const uint8_t *)(src1_spad + ib * (src1_row_size_aligned / sizeof(float)));
|
||||
uint8_t * dst_spad_ptr = (uint8_t *)(dst_spad + ib * (dst_row_size_aligned / sizeof(float)));
|
||||
|
||||
// geglu tanh implementation
|
||||
// geglu(x, g) = gelu(x) * g
|
||||
// gelu(x) = 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)))
|
||||
hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, src0_spad_ptr, nc); // res = x*x
|
||||
hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, GELU_COEF_A, nc); // res = res * GELU_COEF_A
|
||||
hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 1.0f, nc); // res = res + 1.0f
|
||||
hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc); // res = res * x
|
||||
hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, SQRT_2_OVER_PI, nc); // res = result * SQRT_2_OVER_PI
|
||||
hvx_tanh_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, nc); // res = tanh(res)
|
||||
hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, 1.0f, nc); // res = res + 1.0f
|
||||
hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc); // res = res * x
|
||||
hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 0.5f, nc); // res = res + 0.5f
|
||||
hvx_mul_f32_aaa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, src1_spad_ptr, nc); // res = res * g
|
||||
}
|
||||
|
||||
dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), dst_row_size,
|
||||
dst_row_size_aligned, block_size);
|
||||
|
||||
// prefetch N+2 loop iteration if any
|
||||
const uint32_t pref_block = (ir + BLOCK * 2);
|
||||
if (pref_block < src0_end_row) {
|
||||
const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
|
||||
dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)),
|
||||
src0_row_size_aligned, src0_row_size, pref_block_size);
|
||||
dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_size)),
|
||||
src1_row_size_aligned, src1_row_size, pref_block_size);
|
||||
}
|
||||
}
|
||||
|
||||
dma_queue_flush(dma_queue);
|
||||
|
||||
t2 = HAP_perf_get_qtimer_count();
|
||||
|
||||
FARF(HIGH, "geglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
|
||||
ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3,
|
||||
(unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
|
||||
}
|
||||
|
||||
static void unary_silu_f32(unsigned int n, unsigned int i, void * data) {
|
||||
struct htp_ops_context * octx = (struct htp_ops_context *) data;
|
||||
unary_silu_f32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
|
||||
@@ -559,6 +696,12 @@ static void glu_swiglu_oai_f32(unsigned int n, unsigned int i, void * data) {
|
||||
&octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
|
||||
}
|
||||
|
||||
static void glu_geglu_f32(unsigned int n, unsigned int i, void * data) {
|
||||
struct htp_ops_context * octx = (struct htp_ops_context *) data;
|
||||
glu_geglu_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
|
||||
&octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
|
||||
}
|
||||
|
||||
static int execute_op_activations_f32(struct htp_ops_context * octx) {
|
||||
int err = HTP_STATUS_OK;
|
||||
|
||||
@@ -593,6 +736,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
|
||||
act_op_func = unary_gelu_f32;
|
||||
op_type = "gelu-f32";
|
||||
break;
|
||||
|
||||
case HTP_OP_GLU_GEGLU:
|
||||
act_op_func = glu_geglu_f32;
|
||||
op_type = "geglu-f32";
|
||||
break;
|
||||
default:
|
||||
FARF(ERROR, "Unsupported activations Op %u\n", octx->op);
|
||||
return HTP_STATUS_NO_SUPPORT;
|
||||
|
||||
281
ggml/src/ggml-hexagon/htp/argsort-ops.c
Normal file
281
ggml/src/ggml-hexagon/htp/argsort-ops.c
Normal file
@@ -0,0 +1,281 @@
|
||||
#include <string.h>
|
||||
#include <stdlib.h>
|
||||
#include <math.h>
|
||||
#include <HAP_farf.h>
|
||||
#include <HAP_perf.h>
|
||||
|
||||
#define GGML_COMMON_DECL_C
|
||||
#include "ggml-common.h"
|
||||
#include "ggml.h"
|
||||
|
||||
#include "hvx-utils.h"
|
||||
#include "hex-dma.h"
|
||||
|
||||
#include "htp-ctx.h"
|
||||
#include "htp-msg.h"
|
||||
#include "htp-ops.h"
|
||||
|
||||
#ifndef MIN
|
||||
#define MIN(a, b) ((a) < (b) ? (a) : (b))
|
||||
#endif
|
||||
|
||||
struct htp_argsort_context {
|
||||
struct htp_ops_context * octx;
|
||||
uint32_t nrows_per_thread;
|
||||
};
|
||||
|
||||
static inline bool all_greater_f32(HVX_Vector x, HVX_Vector y)
|
||||
{
|
||||
const HVX_Vector one = Q6_V_vsplat_R(1);
|
||||
const HVX_Vector zero = Q6_V_vzero();
|
||||
|
||||
HVX_VectorPred pred = Q6_Q_vcmp_gt_VsfVsf(x, y);
|
||||
HVX_Vector matches = Q6_V_vmux_QVV(pred, one, zero);
|
||||
HVX_Vector sum = hvx_vec_reduce_sum_i32(matches);
|
||||
return hvx_vec_get_i32(sum) == 32;
|
||||
}
|
||||
|
||||
// Sorts values and mirrors swaps to indices.
|
||||
static void quicksort_values_indices_asc(float * values, int32_t * indices, int left, int right) {
|
||||
if (left >= right) return;
|
||||
|
||||
int pivot_idx = (left + right) / 2;
|
||||
float pivot = values[pivot_idx];
|
||||
int i = left;
|
||||
int j = right;
|
||||
|
||||
HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
|
||||
while (i <= j) {
|
||||
// Vectorized scan for i
|
||||
while (i <= j) {
|
||||
// Check if we have at least one full vector
|
||||
if (i + 32 <= j) {
|
||||
HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
|
||||
if (all_greater_f32(pivot_vec, vals_vec)) {
|
||||
// If all elements are < pivot, we can skip this whole block
|
||||
i += 32;
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
// Scalar fallback / cleanup
|
||||
if (values[i] < pivot) {
|
||||
i++;
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// Vectorized scan for j
|
||||
while (i <= j) {
|
||||
if (j - 32 >= i) {
|
||||
// Load 32 elements ending at j.
|
||||
// Since we want `values[j] > pivot`, let's load from j-31 to j.
|
||||
HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
|
||||
if (all_greater_f32(vals_vec, pivot_vec)) {
|
||||
j -= 32;
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
if (values[j] > pivot) {
|
||||
j--;
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (i <= j) {
|
||||
float tmp_val = values[i];
|
||||
values[i] = values[j];
|
||||
values[j] = tmp_val;
|
||||
|
||||
int32_t tmp_idx = indices[i];
|
||||
indices[i] = indices[j];
|
||||
indices[j] = tmp_idx;
|
||||
i++;
|
||||
j--;
|
||||
}
|
||||
}
|
||||
|
||||
if (left < j) quicksort_values_indices_asc(values, indices, left, j);
|
||||
if (i < right) quicksort_values_indices_asc(values, indices, i, right);
|
||||
}
|
||||
|
||||
static void quicksort_values_indices_desc(float * values, int32_t * indices, int left, int right) {
|
||||
if (left >= right) return;
|
||||
|
||||
int pivot_idx = (left + right) / 2;
|
||||
float pivot = values[pivot_idx];
|
||||
int i = left;
|
||||
int j = right;
|
||||
|
||||
HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
|
||||
|
||||
while (i <= j) {
|
||||
// Vectorized scan for i (values[i] > pivot)
|
||||
while (i <= j) {
|
||||
if (i + 32 <= j) {
|
||||
HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
|
||||
if (all_greater_f32(vals_vec, pivot_vec)) {
|
||||
i += 32;
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
if (values[i] > pivot) {
|
||||
i++;
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// Vectorized scan for j (values[j] < pivot)
|
||||
while (i <= j) {
|
||||
if (j - 32 >= i) {
|
||||
HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
|
||||
if (all_greater_f32(pivot_vec, vals_vec)) {
|
||||
j -= 32;
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
if (values[j] < pivot) {
|
||||
j--;
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (i <= j) {
|
||||
float tmp_val = values[i];
|
||||
values[i] = values[j];
|
||||
values[j] = tmp_val;
|
||||
|
||||
int32_t tmp_idx = indices[i];
|
||||
indices[i] = indices[j];
|
||||
indices[j] = tmp_idx;
|
||||
i++;
|
||||
j--;
|
||||
}
|
||||
}
|
||||
|
||||
if (left < j) quicksort_values_indices_desc(values, indices, left, j);
|
||||
if (i < right) quicksort_values_indices_desc(values, indices, i, right);
|
||||
}
|
||||
|
||||
static void htp_argsort_f32(unsigned int n, unsigned int i, void * data) {
|
||||
struct htp_argsort_context * actx = (struct htp_argsort_context *)data;
|
||||
struct htp_ops_context * octx = actx->octx;
|
||||
|
||||
// Unpack context
|
||||
const struct htp_tensor * src0 = &octx->src0;
|
||||
const struct htp_tensor * dst = &octx->dst;
|
||||
|
||||
// Scratchpad memory
|
||||
uint8_t * spad = octx->src0_spad.data + octx->src0_spad.size_per_thread * i;
|
||||
|
||||
// Dimensions
|
||||
uint32_t ne00 = src0->ne[0];
|
||||
uint32_t ne01 = src0->ne[1];
|
||||
uint32_t ne02 = src0->ne[2];
|
||||
uint32_t ne03 = src0->ne[3];
|
||||
|
||||
uint32_t nb01 = src0->nb[1];
|
||||
//uint32_t nb02 = src0->nb[2];
|
||||
//uint32_t nb03 = src0->nb[3];
|
||||
|
||||
uint32_t nb1 = dst->nb[1];
|
||||
//uint32_t nb2 = dst->nb[2];
|
||||
//uint32_t nb3 = dst->nb[3];
|
||||
|
||||
// Sort order
|
||||
enum ggml_sort_order order = (enum ggml_sort_order) octx->op_params[0];
|
||||
|
||||
// Rows to process
|
||||
uint32_t total_rows = ne01 * ne02 * ne03;
|
||||
uint32_t rows_per_thread = actx->nrows_per_thread;
|
||||
uint32_t start_row = rows_per_thread * i;
|
||||
uint32_t end_row = MIN(start_row + rows_per_thread, total_rows);
|
||||
|
||||
// Scratchpad layout:
|
||||
// We need space for one row of float data (values) and one row of int32 indices.
|
||||
// values: ne00 * sizeof(float)
|
||||
// indices: ne00 * sizeof(int32_t)
|
||||
// Padded to 128 bytes.
|
||||
|
||||
size_t values_size = hex_round_up(ne00 * sizeof(float), 128);
|
||||
float * values_buf = (float *) spad;
|
||||
int32_t * indices_buf = (int32_t *) (spad + values_size);
|
||||
|
||||
for (uint32_t r = start_row; r < end_row; r++) {
|
||||
uint32_t src_offset = r * nb01;
|
||||
uint32_t dst_offset = r * nb1;
|
||||
|
||||
uint8_t * src_ptr = (uint8_t *) src0->data + src_offset;
|
||||
uint8_t * dst_ptr = (uint8_t *) dst->data + dst_offset;
|
||||
|
||||
hex_l2fetch(src_ptr, ne00 * sizeof(float), ne00 * sizeof(float), 1);
|
||||
hvx_copy_f32_au((uint8_t*)values_buf, src_ptr, ne00);
|
||||
|
||||
// Initialize indices
|
||||
for (uint32_t j = 0; j < ne00; j++) {
|
||||
indices_buf[j] = j;
|
||||
}
|
||||
|
||||
// Sort values and mirror swaps to indices
|
||||
if (order == GGML_SORT_ORDER_ASC) {
|
||||
quicksort_values_indices_asc(values_buf, indices_buf, 0, ne00 - 1);
|
||||
} else {
|
||||
quicksort_values_indices_desc(values_buf, indices_buf, 0, ne00 - 1);
|
||||
}
|
||||
|
||||
// Copy indices back to DDR
|
||||
hvx_copy_f32_ua(dst_ptr, (const uint8_t *) indices_buf, ne00);
|
||||
}
|
||||
}
|
||||
|
||||
int op_argsort(struct htp_ops_context * octx) {
|
||||
// Check supported types
|
||||
if (octx->src0.type != HTP_TYPE_F32) {
|
||||
return HTP_STATUS_NO_SUPPORT;
|
||||
}
|
||||
|
||||
// Allocate scratchpad
|
||||
// We need 1 row of float + 1 row of int32 per thread.
|
||||
uint32_t ne00 = octx->src0.ne[0];
|
||||
size_t values_size = hex_round_up(ne00 * sizeof(float), 128);
|
||||
size_t indices_size = hex_round_up(ne00 * sizeof(int32_t), 128);
|
||||
size_t spad_per_thread = values_size + indices_size;
|
||||
|
||||
// Make sure we round up to 256 for alignment requirements
|
||||
spad_per_thread = hex_round_up(spad_per_thread, 256);
|
||||
|
||||
size_t total_spad_size = spad_per_thread * octx->n_threads;
|
||||
|
||||
if (octx->ctx->vtcm_size < total_spad_size) {
|
||||
FARF(ERROR, "argsort: VTCM size too small. Needed %zu, have %zu", total_spad_size, octx->ctx->vtcm_size);
|
||||
return HTP_STATUS_VTCM_TOO_SMALL;
|
||||
}
|
||||
|
||||
octx->src0_spad.data = octx->ctx->vtcm_base;
|
||||
octx->src0_spad.size = total_spad_size;
|
||||
octx->src0_spad.size_per_thread = spad_per_thread;
|
||||
|
||||
FARF(HIGH, "argsort: %ux%ux%ux%u -> %ux%ux%ux%u (0x%x, 0x%x)",
|
||||
octx->src0.ne[0], octx->src0.ne[1], octx->src0.ne[2], octx->src0.ne[3],
|
||||
octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
|
||||
octx->src0.data, octx->dst.data);
|
||||
|
||||
uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
|
||||
uint32_t n_jobs = MIN(total_rows, octx->n_threads);
|
||||
|
||||
struct htp_argsort_context actx;
|
||||
actx.octx = octx;
|
||||
actx.nrows_per_thread = (total_rows + n_jobs - 1) / n_jobs;
|
||||
|
||||
// Run jobs
|
||||
worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_jobs);
|
||||
|
||||
return HTP_STATUS_OK;
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -42,32 +42,36 @@ enum htp_data_type {
|
||||
HTP_TYPE_COUNT
|
||||
};
|
||||
|
||||
// These values are manually translated over to HTP
|
||||
// !!!! DO NOT ALTER THE ORDER OF THE FIRST FOUR ENUMS !!!!
|
||||
// Do not reorder first 4 (used as an index)
|
||||
enum htp_op {
|
||||
HTP_OP_MUL = 0,
|
||||
HTP_OP_ADD = 1,
|
||||
HTP_OP_SUB = 2,
|
||||
HTP_OP_DIV = 3,
|
||||
HTP_OP_MUL_MAT = 4,
|
||||
HTP_OP_MUL_MAT_ID = 5,
|
||||
HTP_OP_RMS_NORM = 6,
|
||||
HTP_OP_UNARY_SILU = 7,
|
||||
HTP_OP_UNARY_GELU = 8,
|
||||
HTP_OP_GLU_SWIGLU = 9,
|
||||
HTP_OP_GLU_SWIGLU_OAI = 10,
|
||||
HTP_OP_SOFTMAX = 11,
|
||||
HTP_OP_ADD_ID = 12,
|
||||
HTP_OP_ROPE = 13,
|
||||
HTP_OP_FLASH_ATTN_EXT = 14,
|
||||
HTP_OP_SET_ROWS = 15,
|
||||
HTP_OP_SCALE = 16,
|
||||
HTP_OP_GET_ROWS = 17,
|
||||
HTP_OP_CPY = 18,
|
||||
HTP_OP_MUL = 0,
|
||||
HTP_OP_ADD = 1,
|
||||
HTP_OP_SUB = 2,
|
||||
HTP_OP_DIV = 3,
|
||||
HTP_OP_MUL_MAT,
|
||||
HTP_OP_MUL_MAT_ID,
|
||||
HTP_OP_RMS_NORM,
|
||||
HTP_OP_UNARY_SILU,
|
||||
HTP_OP_UNARY_GELU,
|
||||
HTP_OP_GLU_SWIGLU,
|
||||
HTP_OP_GLU_SWIGLU_OAI,
|
||||
HTP_OP_GLU_GEGLU,
|
||||
HTP_OP_SOFTMAX,
|
||||
HTP_OP_ADD_ID,
|
||||
HTP_OP_ROPE,
|
||||
HTP_OP_FLASH_ATTN_EXT,
|
||||
HTP_OP_SET_ROWS,
|
||||
HTP_OP_GET_ROWS,
|
||||
HTP_OP_SCALE,
|
||||
HTP_OP_CPY,
|
||||
HTP_OP_ARGSORT,
|
||||
HTP_OP_SQR,
|
||||
HTP_OP_SQRT,
|
||||
HTP_OP_SUM_ROWS,
|
||||
INVALID
|
||||
};
|
||||
|
||||
static inline size_t htp_type_block_size(uint32_t t) {
|
||||
static inline size_t htp_t_block_size(uint32_t t) {
|
||||
switch (t) {
|
||||
case HTP_TYPE_F32:
|
||||
return 1;
|
||||
@@ -103,22 +107,6 @@ static inline size_t htp_type_nbytes(uint32_t t) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
static const char * htp_type_name(uint32_t t) {
|
||||
switch (t) {
|
||||
case HTP_TYPE_F32:
|
||||
return "fp32";
|
||||
case HTP_TYPE_F16:
|
||||
return "fp16";
|
||||
case HTP_TYPE_Q4_0:
|
||||
return "q4_0";
|
||||
case HTP_TYPE_Q8_0:
|
||||
return "q8_0";
|
||||
case HTP_TYPE_MXFP4:
|
||||
return "mxfp4";
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
|
||||
// Internal types
|
||||
#define QK_Q4_0x4x2 256 // 4x Q4_0 blocks packed with next 4x Q4_0 blocks (size in bytes 128)
|
||||
#define QK_Q8_0x4x2 256 // 4x Q8_0 blocks concat with next 4x Q8_0 blocks
|
||||
|
||||
@@ -90,6 +90,7 @@ int op_matmul(struct htp_ops_context * octx);
|
||||
int op_matmul_id(struct htp_ops_context * octx);
|
||||
int op_binary(struct htp_ops_context * octx);
|
||||
int op_unary(struct htp_ops_context * octx);
|
||||
int op_sum_rows(struct htp_ops_context * octx);
|
||||
int op_activations(struct htp_ops_context * octx);
|
||||
int op_softmax(struct htp_ops_context * octx);
|
||||
int op_add_id(struct htp_ops_context * octx);
|
||||
@@ -98,5 +99,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx);
|
||||
int op_set_rows(struct htp_ops_context * octx);
|
||||
int op_get_rows(struct htp_ops_context * octx);
|
||||
int op_cpy(struct htp_ops_context * octx);
|
||||
int op_argsort(struct htp_ops_context * octx);
|
||||
|
||||
#endif /* HTP_OPS_H */
|
||||
|
||||
@@ -46,127 +46,76 @@
|
||||
#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
|
||||
#endif
|
||||
|
||||
// ADD variants
|
||||
// Generic macro to define alignment permutations for an op
|
||||
#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO) \
|
||||
static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) dst % 128 == 0); \
|
||||
assert((uintptr_t) src0 % 128 == 0); \
|
||||
assert((uintptr_t) src1 % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) dst % 128 == 0); \
|
||||
assert((uintptr_t) src0 % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) dst % 128 == 0); \
|
||||
assert((uintptr_t) src1 % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) dst % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) src0 % 128 == 0); \
|
||||
assert((uintptr_t) src1 % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) src0 % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
assert((uintptr_t) src1 % 128 == 0); \
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
|
||||
} \
|
||||
static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
|
||||
} \
|
||||
|
||||
static inline void hvx_add_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
assert((unsigned long) src1 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_ADD);
|
||||
DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD)
|
||||
DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB)
|
||||
DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL)
|
||||
|
||||
// Dispatcher logic
|
||||
#define HVX_BINARY_DISPATCHER(OP_NAME) \
|
||||
static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
|
||||
if (hex_is_aligned((void *) dst, 128)) { \
|
||||
if (hex_is_aligned((void *) src0, 128)) { \
|
||||
if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
|
||||
else OP_NAME##_aau(dst, src0, src1, num_elems); \
|
||||
} else { \
|
||||
if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
|
||||
else OP_NAME##_auu(dst, src0, src1, num_elems); \
|
||||
} \
|
||||
} else { \
|
||||
if (hex_is_aligned((void *) src0, 128)) { \
|
||||
if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
|
||||
else OP_NAME##_uau(dst, src0, src1, num_elems); \
|
||||
} else { \
|
||||
if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
|
||||
else OP_NAME##_uuu(dst, src0, src1, num_elems); \
|
||||
} \
|
||||
} \
|
||||
}
|
||||
|
||||
static inline void hvx_add_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_ADD);
|
||||
}
|
||||
|
||||
static inline void hvx_add_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
assert((unsigned long) src1 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_ADD);
|
||||
}
|
||||
|
||||
static inline void hvx_add_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_ADD);
|
||||
}
|
||||
|
||||
// SUB variants
|
||||
|
||||
static inline void hvx_sub_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
assert((unsigned long) src1 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_SUB);
|
||||
}
|
||||
|
||||
static inline void hvx_sub_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_SUB);
|
||||
}
|
||||
|
||||
static inline void hvx_sub_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
assert((unsigned long) src1 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_SUB);
|
||||
}
|
||||
|
||||
static inline void hvx_sub_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_SUB);
|
||||
}
|
||||
|
||||
// MUL variants
|
||||
|
||||
static inline void hvx_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
assert((unsigned long) src1 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MUL);
|
||||
}
|
||||
|
||||
static inline void hvx_mul_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MUL);
|
||||
}
|
||||
|
||||
static inline void hvx_mul_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
assert((unsigned long) src1 % 128 == 0);
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_MUL);
|
||||
}
|
||||
|
||||
static inline void hvx_mul_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MUL);
|
||||
}
|
||||
|
||||
// Dispatchers
|
||||
|
||||
static inline void hvx_add_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
|
||||
if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
|
||||
if (hex_is_aligned((void *) src1, 128)) {
|
||||
hvx_add_f32_aa(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
hvx_add_f32_au(dst, src0, src1, num_elems);
|
||||
}
|
||||
} else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
|
||||
hvx_add_f32_ua(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
hvx_add_f32_uu(dst, src0, src1, num_elems);
|
||||
}
|
||||
}
|
||||
|
||||
static inline void hvx_sub_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
|
||||
if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
|
||||
if (hex_is_aligned((void *) src1, 128)) {
|
||||
hvx_sub_f32_aa(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
hvx_sub_f32_au(dst, src0, src1, num_elems);
|
||||
}
|
||||
} else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
|
||||
hvx_sub_f32_ua(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
hvx_sub_f32_uu(dst, src0, src1, num_elems);
|
||||
}
|
||||
}
|
||||
|
||||
static inline void hvx_mul_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
|
||||
if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
|
||||
if (hex_is_aligned((void *) src1, 128)) {
|
||||
hvx_mul_f32_aa(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
hvx_mul_f32_au(dst, src0, src1, num_elems);
|
||||
}
|
||||
} else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
|
||||
hvx_mul_f32_ua(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
hvx_mul_f32_uu(dst, src0, src1, num_elems);
|
||||
}
|
||||
}
|
||||
HVX_BINARY_DISPATCHER(hvx_add_f32)
|
||||
HVX_BINARY_DISPATCHER(hvx_sub_f32)
|
||||
HVX_BINARY_DISPATCHER(hvx_mul_f32)
|
||||
|
||||
// Mul-Mul Optimized
|
||||
|
||||
static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint8_t * restrict src2, const uint32_t num_elems) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src0 % 128 == 0);
|
||||
@@ -443,6 +392,68 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
|
||||
}
|
||||
}
|
||||
|
||||
//
|
||||
// Square
|
||||
//
|
||||
|
||||
#define hvx_sqr_loop_body(dst_type, src_type, vec_store) \
|
||||
do { \
|
||||
dst_type * restrict vdst = (dst_type *) dst; \
|
||||
src_type * restrict vsrc = (src_type *) src; \
|
||||
\
|
||||
const uint32_t elem_size = sizeof(float); \
|
||||
const uint32_t epv = 128 / elem_size; \
|
||||
const uint32_t nvec = n / epv; \
|
||||
const uint32_t nloe = n % epv; \
|
||||
\
|
||||
uint32_t i = 0; \
|
||||
\
|
||||
_Pragma("unroll(4)") \
|
||||
for (; i < nvec; i++) { \
|
||||
vdst[i] = HVX_OP_MUL(vsrc[i], vsrc[i]); \
|
||||
} \
|
||||
if (nloe) { \
|
||||
HVX_Vector v = HVX_OP_MUL(vsrc[i], vsrc[i]); \
|
||||
vec_store((void *) &vdst[i], nloe * elem_size, v); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
static inline void hvx_sqr_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_sqr_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_sqr_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
hvx_sqr_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_sqr_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
hvx_sqr_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
|
||||
if (hex_is_aligned((void *) dst, 128)) {
|
||||
if (hex_is_aligned((void *) src, 128)) {
|
||||
hvx_sqr_f32_aa(dst, src, num_elems);
|
||||
} else {
|
||||
hvx_sqr_f32_au(dst, src, num_elems);
|
||||
}
|
||||
} else {
|
||||
if (hex_is_aligned((void *) src, 128)) {
|
||||
hvx_sqr_f32_ua(dst, src, num_elems);
|
||||
} else {
|
||||
hvx_sqr_f32_uu(dst, src, num_elems);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#undef HVX_OP_ADD
|
||||
#undef HVX_OP_SUB
|
||||
#undef HVX_OP_MUL
|
||||
@@ -453,5 +464,7 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
|
||||
#undef hvx_scalar_loop_body
|
||||
#undef HVX_OP_MIN_SCALAR
|
||||
#undef HVX_OP_CLAMP_SCALAR
|
||||
#undef DEFINE_HVX_BINARY_OP_VARIANTS
|
||||
#undef HVX_BINARY_DISPATCHER
|
||||
|
||||
#endif // HVX_ARITH_H
|
||||
|
||||
@@ -66,6 +66,12 @@ static inline float hvx_vec_get_f32(HVX_Vector v) {
|
||||
return x;
|
||||
}
|
||||
|
||||
static inline int32_t hvx_vec_get_i32(HVX_Vector v) {
|
||||
int32_t __attribute__((aligned(128))) x;
|
||||
hvx_vec_store_a(&x, 4, v);
|
||||
return x;
|
||||
}
|
||||
|
||||
static inline HVX_Vector hvx_vec_abs_f16(HVX_Vector v) {
|
||||
// abs by clearing the fp16 sign bit
|
||||
HVX_Vector mask = Q6_Vh_vsplat_R(0x7fff);
|
||||
|
||||
@@ -136,8 +136,6 @@ static inline void hvx_copy_f32_uu(uint8_t * restrict dst, const uint8_t * restr
|
||||
dst_type * restrict vdst = (dst_type *) dst; \
|
||||
src_type * restrict vsrc = (src_type *) src; \
|
||||
\
|
||||
const HVX_Vector zero = Q6_V_vsplat_R(0); \
|
||||
\
|
||||
const uint32_t elem_size = sizeof(__fp16); \
|
||||
const uint32_t epv = 128 / elem_size; \
|
||||
const uint32_t nvec = n / epv; \
|
||||
|
||||
116
ggml/src/ggml-hexagon/htp/hvx-div.h
Normal file
116
ggml/src/ggml-hexagon/htp/hvx-div.h
Normal file
@@ -0,0 +1,116 @@
|
||||
#ifndef HVX_DIV_H
|
||||
#define HVX_DIV_H
|
||||
|
||||
#include <HAP_farf.h>
|
||||
|
||||
#include <math.h>
|
||||
#include <string.h>
|
||||
#include <assert.h>
|
||||
#include <stddef.h>
|
||||
#include <stdint.h>
|
||||
|
||||
#include "hvx-base.h"
|
||||
#include "hex-utils.h"
|
||||
#include "hvx-inverse.h"
|
||||
#include "hvx-arith.h"
|
||||
|
||||
#if __HVX_ARCH__ < 79
|
||||
#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
|
||||
#else
|
||||
#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
|
||||
#endif
|
||||
|
||||
#define hvx_div_f32_loop_body(dst_type, src0_type, src1_type, vec_store) \
|
||||
do { \
|
||||
dst_type * restrict vdst = (dst_type *) dst; \
|
||||
src0_type * restrict vsrc0 = (src0_type *) src0; \
|
||||
src1_type * restrict vsrc1 = (src1_type *) src1; \
|
||||
\
|
||||
const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000); \
|
||||
\
|
||||
const uint32_t nvec = n / VLEN_FP32; \
|
||||
const uint32_t nloe = n % VLEN_FP32; \
|
||||
\
|
||||
uint32_t i = 0; \
|
||||
\
|
||||
_Pragma("unroll(4)") \
|
||||
for (; i < nvec; i++) { \
|
||||
HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
|
||||
HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1); \
|
||||
vdst[i] = res; \
|
||||
} \
|
||||
if (nloe) { \
|
||||
HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
|
||||
HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1); \
|
||||
vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, res); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
// 3-letter suffix variants
|
||||
static inline void hvx_div_f32_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) dst % 128 == 0);
|
||||
assert((uintptr_t) src0 % 128 == 0);
|
||||
assert((uintptr_t) src1 % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) dst % 128 == 0);
|
||||
assert((uintptr_t) src0 % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) dst % 128 == 0);
|
||||
assert((uintptr_t) src1 % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) dst % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) src0 % 128 == 0);
|
||||
assert((uintptr_t) src1 % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) src0 % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
assert((uintptr_t) src1 % 128 == 0);
|
||||
hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
|
||||
hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_div_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
|
||||
if (hex_is_aligned((void *) dst, 128)) {
|
||||
if (hex_is_aligned((void *) src0, 128)) {
|
||||
if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aaa(dst, src0, src1, num_elems);
|
||||
else hvx_div_f32_aau(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aua(dst, src0, src1, num_elems);
|
||||
else hvx_div_f32_auu(dst, src0, src1, num_elems);
|
||||
}
|
||||
} else {
|
||||
if (hex_is_aligned((void *) src0, 128)) {
|
||||
if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uaa(dst, src0, src1, num_elems);
|
||||
else hvx_div_f32_uau(dst, src0, src1, num_elems);
|
||||
} else {
|
||||
if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uua(dst, src0, src1, num_elems);
|
||||
else hvx_div_f32_uuu(dst, src0, src1, num_elems);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#undef HVX_OP_MUL
|
||||
|
||||
#endif // HVX_DIV_H
|
||||
@@ -91,6 +91,27 @@ static inline HVX_Vector hvx_vec_tanh_f32(HVX_Vector x) {
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define hvx_tanh_loop_body(dst_type, src_type, vec_store) \
|
||||
do { \
|
||||
dst_type * restrict vdst = (dst_type *) dst; \
|
||||
src_type * restrict vsrc = (src_type *) src; \
|
||||
\
|
||||
const uint32_t epv = 128 / sizeof(float); \
|
||||
const uint32_t nvec = n / epv; \
|
||||
const uint32_t nloe = n % epv; \
|
||||
\
|
||||
uint32_t i = 0; \
|
||||
\
|
||||
_Pragma("unroll(4)") \
|
||||
for (; i < nvec; i++) { \
|
||||
vdst[i] = hvx_vec_tanh_f32(vsrc[i]); \
|
||||
} \
|
||||
if (nloe) { \
|
||||
HVX_Vector tmp = hvx_vec_tanh_f32(vsrc[i]); \
|
||||
vec_store((void *) &vdst[i], nloe * sizeof(float), tmp); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
static inline void hvx_sigmoid_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
@@ -111,4 +132,10 @@ static inline void hvx_sigmoid_f32_uu(uint8_t * restrict dst, const uint8_t * re
|
||||
hvx_sigmoid_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_tanh_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
hvx_tanh_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
#endif /* HVX_SIGMOID_H */
|
||||
|
||||
@@ -12,11 +12,17 @@
|
||||
#define RSQRT_ONE_HALF 0x3f000000 // 0.5
|
||||
#define RSQRT_THREE_HALVES 0x3fc00000 // 1.5
|
||||
|
||||
#if __HVX_ARCH__ < 79
|
||||
#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
|
||||
#else
|
||||
#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
|
||||
#endif
|
||||
|
||||
static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
|
||||
//Algorithm :
|
||||
// x2 = input*0.5
|
||||
// y = * (long *) &input
|
||||
// y = 0x5f3759df - (y>>2)
|
||||
// y = 0x5f3759df - (y>>1)
|
||||
// y = y*(threehalfs - x2*y*y)
|
||||
|
||||
HVX_Vector rsqrtconst = Q6_V_vsplat_R(RSQRT_CONST);
|
||||
@@ -57,4 +63,64 @@ static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
|
||||
return Q6_Vsf_equals_Vqf32(temp);
|
||||
}
|
||||
|
||||
// Compute sqrt(x) as x*inv_sqrt(x)
|
||||
#define hvx_sqrt_f32_loop_body(dst_type, src_type, vec_store) \
|
||||
do { \
|
||||
dst_type * restrict vdst = (dst_type *) dst; \
|
||||
src_type * restrict vsrc = (src_type *) src; \
|
||||
\
|
||||
const uint32_t nvec = n / VLEN_FP32; \
|
||||
const uint32_t nloe = n % VLEN_FP32; \
|
||||
\
|
||||
uint32_t i = 0; \
|
||||
\
|
||||
_Pragma("unroll(4)") \
|
||||
for (; i < nvec; i++) { \
|
||||
HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]); \
|
||||
HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]); \
|
||||
vdst[i] = sqrt_res; \
|
||||
} \
|
||||
if (nloe) { \
|
||||
HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]); \
|
||||
HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]); \
|
||||
vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, sqrt_res); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
static inline void hvx_sqrt_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
hvx_sqrt_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_sqrt_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
hvx_sqrt_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
|
||||
}
|
||||
|
||||
static inline void hvx_sqrt_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
hvx_sqrt_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_sqrt_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
hvx_sqrt_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
|
||||
}
|
||||
|
||||
static inline void hvx_sqrt_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems) {
|
||||
if ((unsigned long) dst % 128 == 0) {
|
||||
if ((unsigned long) src % 128 == 0) {
|
||||
hvx_sqrt_f32_aa(dst, src, num_elems);
|
||||
} else {
|
||||
hvx_sqrt_f32_au(dst, src, num_elems);
|
||||
}
|
||||
} else {
|
||||
if ((unsigned long) src % 128 == 0) {
|
||||
hvx_sqrt_f32_ua(dst, src, num_elems);
|
||||
} else {
|
||||
hvx_sqrt_f32_uu(dst, src, num_elems);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#endif /* HVX_SQRT_H */
|
||||
|
||||
@@ -12,6 +12,7 @@
|
||||
#include "hvx-sigmoid.h"
|
||||
#include "hvx-sqrt.h"
|
||||
#include "hvx-arith.h"
|
||||
#include "hvx-div.h"
|
||||
#include "hvx-base.h"
|
||||
|
||||
#endif /* HVX_UTILS_H */
|
||||
|
||||
@@ -440,6 +440,45 @@ static void proc_matmul_req(struct htp_context * ctx,
|
||||
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
|
||||
}
|
||||
|
||||
static void proc_argsort_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
|
||||
struct dspqueue_buffer rsp_bufs[1];
|
||||
|
||||
// We had written to the output buffer, we'd also need to flush it
|
||||
rsp_bufs[0].fd = bufs[1].fd;
|
||||
rsp_bufs[0].ptr = bufs[1].ptr;
|
||||
rsp_bufs[0].offset = bufs[1].offset;
|
||||
rsp_bufs[0].size = bufs[1].size;
|
||||
rsp_bufs[0].flags = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | // Flush HTP
|
||||
DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT); // Invalidate CPU
|
||||
|
||||
// Setup Op context
|
||||
struct htp_ops_context octx = { 0 };
|
||||
octx.ctx = ctx;
|
||||
octx.src0 = req->src0;
|
||||
octx.dst = req->dst;
|
||||
octx.flags = req->flags;
|
||||
octx.op = req->op;
|
||||
|
||||
memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
|
||||
|
||||
// Update data pointers
|
||||
octx.src0.data = (uint32_t) bufs[0].ptr;
|
||||
octx.dst.data = (uint32_t) bufs[1].ptr;
|
||||
octx.n_threads = ctx->n_threads;
|
||||
|
||||
struct profile_data prof;
|
||||
profile_start(&prof);
|
||||
|
||||
uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
|
||||
if (vtcm_acquire(ctx) == AEE_SUCCESS) {
|
||||
rsp_status = op_argsort(&octx);
|
||||
vtcm_release(ctx);
|
||||
}
|
||||
|
||||
profile_stop(&prof);
|
||||
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
|
||||
}
|
||||
|
||||
static void proc_cpy_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
|
||||
struct dspqueue_buffer rsp_bufs[1];
|
||||
|
||||
@@ -679,6 +718,45 @@ static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * re
|
||||
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
|
||||
}
|
||||
|
||||
static void proc_sum_rows_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
|
||||
struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
|
||||
|
||||
// We had written to the output buffer, we'd also need to flush it
|
||||
rsp_bufs[0].fd = bufs[1].fd;
|
||||
rsp_bufs[0].ptr = bufs[1].ptr;
|
||||
rsp_bufs[0].offset = bufs[1].offset;
|
||||
rsp_bufs[0].size = bufs[1].size;
|
||||
rsp_bufs[0].flags = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | // Flush HTP
|
||||
DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT); // Invalidate CPU
|
||||
|
||||
// Setup Op context
|
||||
struct htp_ops_context octx = { 0 };
|
||||
octx.ctx = ctx;
|
||||
octx.src0 = req->src0;
|
||||
octx.dst = req->dst;
|
||||
octx.flags = req->flags;
|
||||
octx.op = req->op;
|
||||
|
||||
memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
|
||||
|
||||
// Update data pointers
|
||||
octx.src0.data = (uint32_t) bufs[0].ptr;
|
||||
octx.dst.data = (uint32_t) bufs[1].ptr;
|
||||
octx.n_threads = ctx->n_threads;
|
||||
|
||||
struct profile_data prof;
|
||||
profile_start(&prof);
|
||||
|
||||
uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
|
||||
if (vtcm_acquire(ctx) == AEE_SUCCESS) {
|
||||
rsp_status = op_sum_rows(&octx);
|
||||
vtcm_release(ctx);
|
||||
}
|
||||
|
||||
profile_stop(&prof);
|
||||
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
|
||||
}
|
||||
|
||||
static void proc_activations_req(struct htp_context * ctx,
|
||||
struct htp_general_req * req,
|
||||
struct dspqueue_buffer * bufs,
|
||||
@@ -951,6 +1029,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
|
||||
case HTP_OP_MUL:
|
||||
case HTP_OP_ADD:
|
||||
case HTP_OP_SUB:
|
||||
case HTP_OP_DIV:
|
||||
if (n_bufs != 3) {
|
||||
FARF(ERROR, "Bad binary-req buffer list");
|
||||
continue;
|
||||
@@ -968,6 +1047,25 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
|
||||
proc_unary_req(ctx, &req, bufs);
|
||||
break;
|
||||
|
||||
case HTP_OP_SQR:
|
||||
case HTP_OP_SQRT:
|
||||
if (n_bufs != 2) {
|
||||
FARF(ERROR, "Bad unary-req buffer list");
|
||||
continue;
|
||||
}
|
||||
|
||||
proc_unary_req(ctx, &req, bufs);
|
||||
break;
|
||||
|
||||
case HTP_OP_SUM_ROWS:
|
||||
if (n_bufs != 2) {
|
||||
FARF(ERROR, "Bad unary-req buffer list");
|
||||
continue;
|
||||
}
|
||||
|
||||
proc_sum_rows_req(ctx, &req, bufs);
|
||||
break;
|
||||
|
||||
case HTP_OP_UNARY_SILU:
|
||||
case HTP_OP_UNARY_GELU:
|
||||
if (n_bufs != 2) {
|
||||
@@ -980,6 +1078,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
|
||||
case HTP_OP_GLU_SWIGLU:
|
||||
case HTP_OP_GLU_SWIGLU_OAI:
|
||||
case HTP_OP_SOFTMAX:
|
||||
case HTP_OP_GLU_GEGLU:
|
||||
if ((n_bufs != 2) && (n_bufs != 3)) {
|
||||
FARF(ERROR, "Bad act-req buffer list");
|
||||
continue;
|
||||
@@ -1035,6 +1134,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
|
||||
proc_cpy_req(ctx, &req, bufs);
|
||||
break;
|
||||
|
||||
case HTP_OP_ARGSORT:
|
||||
if (n_bufs != 2) {
|
||||
FARF(ERROR, "Bad argsort-req buffer list");
|
||||
continue;
|
||||
}
|
||||
proc_argsort_req(ctx, &req, bufs);
|
||||
break;
|
||||
|
||||
default:
|
||||
FARF(ERROR, "Unknown Op %u", req.op);
|
||||
break;
|
||||
|
||||
115
ggml/src/ggml-hexagon/htp/sum-rows-ops.c
Normal file
115
ggml/src/ggml-hexagon/htp/sum-rows-ops.c
Normal file
@@ -0,0 +1,115 @@
|
||||
#pragma clang diagnostic ignored "-Wunused-variable"
|
||||
#pragma clang diagnostic ignored "-Wunused-function"
|
||||
#pragma clang diagnostic ignored "-Wunused-but-set-variable"
|
||||
|
||||
#include <HAP_farf.h>
|
||||
#include <HAP_perf.h>
|
||||
|
||||
#include <string.h>
|
||||
#include <math.h>
|
||||
|
||||
#include "hex-dma.h"
|
||||
#include "hvx-utils.h"
|
||||
|
||||
#define GGML_COMMON_DECL_C
|
||||
#include "ggml-common.h"
|
||||
#include "htp-ctx.h"
|
||||
#include "htp-msg.h"
|
||||
#include "htp-ops.h"
|
||||
|
||||
|
||||
#define sum_rows_preamble \
|
||||
struct htp_tensor *src0 = &octx->src0;\
|
||||
struct htp_tensor *dst = &octx->dst; \
|
||||
\
|
||||
const uint32_t ne00 = src0->ne[0]; \
|
||||
const uint32_t ne01 = src0->ne[1]; \
|
||||
const uint32_t ne02 = src0->ne[2]; \
|
||||
const uint32_t ne03 = src0->ne[3]; \
|
||||
\
|
||||
const uint32_t nb00 = src0->nb[0]; \
|
||||
const uint32_t nb01 = src0->nb[1]; \
|
||||
const uint32_t nb02 = src0->nb[2]; \
|
||||
const uint32_t nb03 = src0->nb[3]; \
|
||||
\
|
||||
const uint32_t ne0 = dst->ne[0]; \
|
||||
const uint32_t ne1 = dst->ne[1]; \
|
||||
const uint32_t ne2 = dst->ne[2]; \
|
||||
const uint32_t ne3 = dst->ne[3]; \
|
||||
\
|
||||
const uint32_t nb0 = dst->nb[0]; \
|
||||
const uint32_t nb1 = dst->nb[1]; \
|
||||
const uint32_t nb2 = dst->nb[2]; \
|
||||
const uint32_t nb3 = dst->nb[3]; \
|
||||
|
||||
static int sum_rows_thread_f32(struct htp_ops_context * octx, const int nth, const int ith) {
|
||||
sum_rows_preamble;
|
||||
|
||||
const uint32_t src0_nrows_per_thread = octx->src0_nrows_per_thread;
|
||||
const size_t src0_row_size = nb01;
|
||||
const size_t dst_row_size = nb1;
|
||||
|
||||
const uint32_t src0_nrows = ne01 * ne02 * ne03; // src0 rows
|
||||
|
||||
const uint32_t src0_start_row = src0_nrows_per_thread * ith;
|
||||
const uint32_t src0_end_row = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
|
||||
|
||||
// no work for this thread
|
||||
if (src0_start_row >= src0_end_row) {
|
||||
return HTP_STATUS_OK;
|
||||
}
|
||||
|
||||
int opt_path = 0;
|
||||
if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
|
||||
opt_path = 1;
|
||||
}
|
||||
|
||||
const uint8_t * restrict data_src = (const uint8_t *) src0->data;
|
||||
uint8_t * restrict data_dst = (uint8_t *) dst->data;
|
||||
|
||||
const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
|
||||
float * restrict dst_th = (float *) (data_dst + (src0_start_row * dst_row_size));
|
||||
|
||||
for (uint32_t ir = 0; ir < src0_nrows_per_thread; ir++) {
|
||||
const float * restrict src_local = src_th + (ir * ne00);
|
||||
|
||||
if (ir + 1 < src0_nrows_per_thread) {
|
||||
hex_l2fetch(src_local + ne00, src0_row_size, src0_row_size, 1);
|
||||
}
|
||||
|
||||
if (1 == opt_path) {
|
||||
dst_th[ir] = hvx_reduce_sum_f32_a((const uint8_t *) src_local, ne00);
|
||||
} else {
|
||||
dst_th[ir] = hvx_reduce_sum_f32((const uint8_t *) src_local, ne00);
|
||||
}
|
||||
}
|
||||
|
||||
return HTP_STATUS_OK;
|
||||
}
|
||||
|
||||
static void sum_rows_work_f32(unsigned int n, unsigned int i, void *data) {
|
||||
sum_rows_thread_f32((struct htp_ops_context *) data, n, i);
|
||||
}
|
||||
|
||||
int op_sum_rows(struct htp_ops_context * octx) {
|
||||
sum_rows_preamble;
|
||||
|
||||
if (octx->src0.type != HTP_TYPE_F32) {
|
||||
return HTP_STATUS_NO_SUPPORT;
|
||||
}
|
||||
|
||||
if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
|
||||
return HTP_STATUS_OK;
|
||||
}
|
||||
|
||||
const int n_threads = octx->n_threads;
|
||||
const uint32_t src0_nrows = ne01 * ne02 * ne03;
|
||||
|
||||
uint32_t n_jobs = MIN(n_threads, src0_nrows);
|
||||
octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
|
||||
|
||||
worker_pool_run_func(octx->ctx->worker_pool, sum_rows_work_f32, octx, n_jobs);
|
||||
|
||||
return HTP_STATUS_OK;
|
||||
}
|
||||
|
||||
@@ -132,6 +132,56 @@ static void rms_norm_htp_f32(const float * restrict src,
|
||||
}
|
||||
}
|
||||
|
||||
static void sqr_htp_f32(const float * restrict src,
|
||||
float * restrict dst,
|
||||
uint8_t * restrict spad,
|
||||
const uint32_t num_rows,
|
||||
const uint32_t row_elems,
|
||||
const size_t row_size,
|
||||
int32_t * op_params,
|
||||
int opt_path) {
|
||||
|
||||
for (uint32_t ir = 0; ir < num_rows; ir++) {
|
||||
const float * restrict src_local = src + (ir * row_elems);
|
||||
float * restrict dst_local = dst + (ir * row_elems);
|
||||
|
||||
if (ir + 1 < num_rows) {
|
||||
hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
|
||||
}
|
||||
|
||||
if (1 == opt_path) {
|
||||
hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
|
||||
} else {
|
||||
hvx_sqr_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static void sqrt_htp_f32(const float * restrict src,
|
||||
float * restrict dst,
|
||||
uint8_t * restrict spad,
|
||||
const uint32_t num_rows,
|
||||
const uint32_t row_elems,
|
||||
const size_t row_size,
|
||||
int32_t * op_params,
|
||||
int opt_path) {
|
||||
|
||||
for (uint32_t ir = 0; ir < num_rows; ir++) {
|
||||
const float * restrict src_local = src + (ir * row_elems);
|
||||
float * restrict dst_local = dst + (ir * row_elems);
|
||||
|
||||
if (ir + 1 < num_rows) {
|
||||
hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
|
||||
}
|
||||
|
||||
if (1 == opt_path) {
|
||||
hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
|
||||
} else {
|
||||
hvx_sqrt_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static void unary_job_f32_per_thread(const struct htp_tensor * src,
|
||||
struct htp_tensor * dst,
|
||||
uint8_t * spad,
|
||||
@@ -181,6 +231,12 @@ static void unary_job_f32_per_thread(const struct htp_tensor * src,
|
||||
case HTP_OP_SCALE:
|
||||
scale_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
|
||||
break;
|
||||
case HTP_OP_SQR:
|
||||
sqr_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
|
||||
break;
|
||||
case HTP_OP_SQRT:
|
||||
sqrt_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
|
||||
break;
|
||||
|
||||
default:
|
||||
break;
|
||||
@@ -218,6 +274,14 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
|
||||
unary_op_func = unary_job_dispatcher_f32;
|
||||
op_type = "scale-f32";
|
||||
break;
|
||||
case HTP_OP_SQR:
|
||||
unary_op_func = unary_job_dispatcher_f32;
|
||||
op_type = "sqr-f32";
|
||||
break;
|
||||
case HTP_OP_SQRT:
|
||||
unary_op_func = unary_job_dispatcher_f32;
|
||||
op_type = "sqrt-f32";
|
||||
break;
|
||||
|
||||
default:
|
||||
FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
|
||||
|
||||
@@ -394,7 +394,7 @@ bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, con
|
||||
[encoder endEncoding];
|
||||
|
||||
ggml_metal_event_t ev_cpy = ggml_metal_get_ev_cpy(ctx_src);
|
||||
ggml_metal_event_record(ctx_src, ev_cpy);
|
||||
ggml_metal_event_encode_signal(ev_cpy, cmd_buf);
|
||||
|
||||
[cmd_buf commit];
|
||||
|
||||
@@ -415,7 +415,7 @@ bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, con
|
||||
|
||||
enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf) {
|
||||
// number of nodes encoded by the main thread (empirically determined)
|
||||
const int n_main = 64;
|
||||
const int n_main = MAX(64, 0.1*gf->n_nodes);
|
||||
|
||||
// number of threads in addition to the main thread
|
||||
const int n_cb = ctx->n_cb;
|
||||
|
||||
@@ -176,6 +176,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_me
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag(ggml_metal_library_t lib, const ggml_tensor * op) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
const int n = op->src[0]->ne[0];
|
||||
|
||||
snprintf(base, 256, "kernel_diag_%s", ggml_type_name(op->src[0]->type));
|
||||
snprintf(name, 256, "%s_n=%d", base, n);
|
||||
|
||||
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
|
||||
if (!res.pipeline) {
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
|
||||
}
|
||||
|
||||
res.nsg = 1;
|
||||
res.smem = 0;
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat(ggml_metal_library_t lib, ggml_type tsrc) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
@@ -192,61 +212,69 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat(ggml_meta
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary(ggml_metal_library_t lib, const ggml_tensor * op) {
|
||||
GGML_ASSERT(ggml_is_contiguous(op->src[0]));
|
||||
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
const int64_t n = ggml_nelements(op);
|
||||
int op_num = -1;
|
||||
|
||||
const char * op_str = "undefined";
|
||||
switch (op->op) {
|
||||
case GGML_OP_SCALE: op_str = "scale"; break;
|
||||
case GGML_OP_FILL: op_str = "fill"; break;
|
||||
case GGML_OP_CLAMP: op_str = "clamp"; break;
|
||||
case GGML_OP_SQR: op_str = "sqr"; break;
|
||||
case GGML_OP_SQRT: op_str = "sqrt"; break;
|
||||
case GGML_OP_SIN: op_str = "sin"; break;
|
||||
case GGML_OP_COS: op_str = "cos"; break;
|
||||
case GGML_OP_LOG: op_str = "log"; break;
|
||||
case GGML_OP_LEAKY_RELU: op_str = "leaky_relu"; break;
|
||||
case GGML_OP_SCALE: op_num = OP_UNARY_NUM_SCALE; break;
|
||||
case GGML_OP_FILL: op_num = OP_UNARY_NUM_FILL; break;
|
||||
case GGML_OP_CLAMP: op_num = OP_UNARY_NUM_CLAMP; break;
|
||||
case GGML_OP_SQR: op_num = OP_UNARY_NUM_SQR; break;
|
||||
case GGML_OP_SQRT: op_num = OP_UNARY_NUM_SQRT; break;
|
||||
case GGML_OP_SIN: op_num = OP_UNARY_NUM_SIN; break;
|
||||
case GGML_OP_COS: op_num = OP_UNARY_NUM_COS; break;
|
||||
case GGML_OP_LOG: op_num = OP_UNARY_NUM_LOG; break;
|
||||
case GGML_OP_LEAKY_RELU: op_num = OP_UNARY_NUM_LEAKY_RELU; break;
|
||||
case GGML_OP_UNARY:
|
||||
switch (ggml_get_unary_op(op)) {
|
||||
case GGML_UNARY_OP_TANH: op_str = "tanh"; break;
|
||||
case GGML_UNARY_OP_RELU: op_str = "relu"; break;
|
||||
case GGML_UNARY_OP_SIGMOID: op_str = "sigmoid"; break;
|
||||
case GGML_UNARY_OP_GELU: op_str = "gelu"; break;
|
||||
case GGML_UNARY_OP_GELU_ERF: op_str = "gelu_erf"; break;
|
||||
case GGML_UNARY_OP_GELU_QUICK: op_str = "gelu_quick"; break;
|
||||
case GGML_UNARY_OP_SILU: op_str = "silu"; break;
|
||||
case GGML_UNARY_OP_ELU: op_str = "elu"; break;
|
||||
case GGML_UNARY_OP_NEG: op_str = "neg"; break;
|
||||
case GGML_UNARY_OP_ABS: op_str = "abs"; break;
|
||||
case GGML_UNARY_OP_SGN: op_str = "sgn"; break;
|
||||
case GGML_UNARY_OP_STEP: op_str = "step"; break;
|
||||
case GGML_UNARY_OP_HARDSWISH: op_str = "hardswish"; break;
|
||||
case GGML_UNARY_OP_HARDSIGMOID: op_str = "hardsigmoid"; break;
|
||||
case GGML_UNARY_OP_EXP: op_str = "exp"; break;
|
||||
case GGML_UNARY_OP_SOFTPLUS: op_str = "softplus"; break;
|
||||
case GGML_UNARY_OP_EXPM1: op_str = "expm1"; break;
|
||||
case GGML_UNARY_OP_TANH: op_num = OP_UNARY_NUM_TANH; break;
|
||||
case GGML_UNARY_OP_RELU: op_num = OP_UNARY_NUM_RELU; break;
|
||||
case GGML_UNARY_OP_SIGMOID: op_num = OP_UNARY_NUM_SIGMOID; break;
|
||||
case GGML_UNARY_OP_GELU: op_num = OP_UNARY_NUM_GELU; break;
|
||||
case GGML_UNARY_OP_GELU_ERF: op_num = OP_UNARY_NUM_GELU_ERF; break;
|
||||
case GGML_UNARY_OP_GELU_QUICK: op_num = OP_UNARY_NUM_GELU_QUICK; break;
|
||||
case GGML_UNARY_OP_SILU: op_num = OP_UNARY_NUM_SILU; break;
|
||||
case GGML_UNARY_OP_ELU: op_num = OP_UNARY_NUM_ELU; break;
|
||||
case GGML_UNARY_OP_NEG: op_num = OP_UNARY_NUM_NEG; break;
|
||||
case GGML_UNARY_OP_ABS: op_num = OP_UNARY_NUM_ABS; break;
|
||||
case GGML_UNARY_OP_SGN: op_num = OP_UNARY_NUM_SGN; break;
|
||||
case GGML_UNARY_OP_STEP: op_num = OP_UNARY_NUM_STEP; break;
|
||||
case GGML_UNARY_OP_HARDSWISH: op_num = OP_UNARY_NUM_HARDSWISH; break;
|
||||
case GGML_UNARY_OP_HARDSIGMOID: op_num = OP_UNARY_NUM_HARDSIGMOID; break;
|
||||
case GGML_UNARY_OP_EXP: op_num = OP_UNARY_NUM_EXP; break;
|
||||
case GGML_UNARY_OP_SOFTPLUS: op_num = OP_UNARY_NUM_SOFTPLUS; break;
|
||||
case GGML_UNARY_OP_EXPM1: op_num = OP_UNARY_NUM_EXPM1; break;
|
||||
default: GGML_ABORT("fatal error");
|
||||
} break;
|
||||
default: GGML_ABORT("fatal error");
|
||||
};
|
||||
|
||||
const char * suffix = "";
|
||||
if (n % 4 == 0) {
|
||||
suffix = "_4";
|
||||
}
|
||||
const char * t0_str = ggml_type_name(op->src[0]->type);
|
||||
const char * t_str = ggml_type_name(op->type);
|
||||
|
||||
snprintf(base, 256, "kernel_%s_%s%s", op_str, ggml_type_name(op->src[0]->type), suffix);
|
||||
snprintf(name, 256, "%s", base);
|
||||
const bool is_c4 = op->src[0]->ne[0] % 4 == 0;
|
||||
const bool is_cnt = ggml_is_contiguous(op->src[0]) && ggml_nelements(op) < 32768;
|
||||
|
||||
snprintf(base, 256, "kernel_unary_%s_%s%s", t0_str, t_str, is_c4 ? "_4" : "");
|
||||
snprintf(name, 256, "%s_op=%d_cnt=%d", base, op_num, is_cnt);
|
||||
|
||||
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
|
||||
if (!res.pipeline) {
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
|
||||
ggml_metal_cv_t cv = ggml_metal_cv_init();
|
||||
|
||||
ggml_metal_cv_set_int16(cv, op_num, FC_UNARY + 0);
|
||||
ggml_metal_cv_set_bool (cv, is_cnt, FC_UNARY + 1);
|
||||
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
|
||||
|
||||
ggml_metal_cv_free(cv);
|
||||
}
|
||||
|
||||
res.c4 = is_c4;
|
||||
res.cnt = is_cnt;
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
@@ -1372,34 +1400,78 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
|
||||
GGML_UNUSED(op);
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(
|
||||
ggml_metal_library_t lib,
|
||||
ggml_op op,
|
||||
int32_t n_fuse,
|
||||
bool row) {
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_library_t lib, const ggml_tensor * op, int32_t n_fuse) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
const char * op_str = "undefined";
|
||||
switch (op) {
|
||||
case GGML_OP_ADD: op_str = "add"; break;
|
||||
case GGML_OP_SUB: op_str = "sub"; break;
|
||||
case GGML_OP_MUL: op_str = "mul"; break;
|
||||
case GGML_OP_DIV: op_str = "div"; break;
|
||||
int op_num = -1;
|
||||
|
||||
switch (op->op) {
|
||||
case GGML_OP_ADD: op_num = 0; break;
|
||||
case GGML_OP_SUB: op_num = 1; break;
|
||||
case GGML_OP_MUL: op_num = 2; break;
|
||||
case GGML_OP_DIV: op_num = 3; break;
|
||||
default: GGML_ABORT("fatal error");
|
||||
};
|
||||
|
||||
if (row) {
|
||||
snprintf(base, 256, "kernel_%s_row_c4_fuse_%d", op_str, n_fuse);
|
||||
} else {
|
||||
snprintf(base, 256, "kernel_%s_fuse_%d", op_str, n_fuse);
|
||||
}
|
||||
const char * t0_str = ggml_type_name(op->src[0]->type);
|
||||
const char * t1_str = ggml_type_name(op->src[1]->type);
|
||||
const char * t_str = ggml_type_name(op->type);
|
||||
|
||||
snprintf(name, 256, "%s", base);
|
||||
const bool is_c4 = (op->src[0]->ne[0] % 4 == 0) && (op->src[1]->ne[0] % 4 == 0);
|
||||
|
||||
const bool is_rb = ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && (ggml_nrows(op->src[1]) == 1) && ggml_nelements(op) < 65536;
|
||||
|
||||
snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s%s", t0_str, t1_str, t_str, is_c4 ? "_4" : "");
|
||||
snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d", base, op_num, n_fuse, is_rb);
|
||||
|
||||
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
|
||||
if (!res.pipeline) {
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
|
||||
ggml_metal_cv_t cv = ggml_metal_cv_init();
|
||||
|
||||
ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
|
||||
ggml_metal_cv_set_int16(cv, n_fuse, FC_BIN + 1);
|
||||
ggml_metal_cv_set_bool (cv, is_rb, FC_BIN + 2);
|
||||
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
|
||||
|
||||
ggml_metal_cv_free(cv);
|
||||
}
|
||||
|
||||
res.c4 = is_c4;
|
||||
res.cnt = is_rb;
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one(ggml_metal_library_t lib, ggml_op op) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
int op_num = -1;
|
||||
|
||||
switch (op) {
|
||||
case GGML_OP_ADD: op_num = 0; break;
|
||||
case GGML_OP_SUB: op_num = 1; break;
|
||||
case GGML_OP_MUL: op_num = 2; break;
|
||||
case GGML_OP_DIV: op_num = 3; break;
|
||||
default: GGML_ABORT("fatal error");
|
||||
};
|
||||
|
||||
snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s", "f32", "f32", "f32");
|
||||
snprintf(name, 256, "%s_op=%d_nf=%d", base, op_num, 1);
|
||||
|
||||
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
|
||||
if (!res.pipeline) {
|
||||
ggml_metal_cv_t cv = ggml_metal_cv_init();
|
||||
|
||||
ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
|
||||
ggml_metal_cv_set_int16(cv, 1, FC_BIN + 1);
|
||||
ggml_metal_cv_set_bool (cv, false, FC_BIN + 2);
|
||||
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
|
||||
|
||||
ggml_metal_cv_free(cv);
|
||||
}
|
||||
|
||||
return res;
|
||||
|
||||
@@ -53,6 +53,9 @@ struct ggml_metal_pipeline_with_params {
|
||||
int nr1;
|
||||
|
||||
size_t smem;
|
||||
|
||||
bool c4;
|
||||
bool cnt;
|
||||
};
|
||||
|
||||
int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_with_params pipeline);
|
||||
@@ -108,6 +111,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_1d
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_2d (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows (ggml_metal_library_t lib, enum ggml_type tsrc);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat (ggml_metal_library_t lib, enum ggml_type tsrc);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
@@ -133,7 +137,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse );
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one (ggml_metal_library_t lib, enum ggml_op op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
|
||||
|
||||
@@ -346,10 +346,12 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_meta
|
||||
|
||||
struct ggml_metal_pipeline_with_params res = {
|
||||
/*.pipeline =*/ nil,
|
||||
/*.nsg =*/ 0,
|
||||
/*.nr0 =*/ 0,
|
||||
/*.nr1 =*/ 0,
|
||||
/*.nsg =*/ 0,
|
||||
/*.smem =*/ 0,
|
||||
/*.c4 =*/ false,
|
||||
/*.cnt =*/ false,
|
||||
};
|
||||
|
||||
res.pipeline = ggml_metal_pipelines_get(lib->pipelines, name);
|
||||
@@ -362,10 +364,12 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_meta
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv) {
|
||||
struct ggml_metal_pipeline_with_params res = {
|
||||
/*.pipeline =*/ nil,
|
||||
/*.nsg =*/ 0,
|
||||
/*.nr0 =*/ 0,
|
||||
/*.nr1 =*/ 0,
|
||||
/*.nsg =*/ 0,
|
||||
/*.smem =*/ 0,
|
||||
/*.c4 =*/ false,
|
||||
/*.cnt =*/ false,
|
||||
};
|
||||
|
||||
[lib->lock lock];
|
||||
@@ -1007,6 +1011,15 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
}
|
||||
|
||||
switch (op->op) {
|
||||
case GGML_OP_SCALE:
|
||||
case GGML_OP_FILL:
|
||||
case GGML_OP_CLAMP:
|
||||
case GGML_OP_SQR:
|
||||
case GGML_OP_SQRT:
|
||||
case GGML_OP_SIN:
|
||||
case GGML_OP_COS:
|
||||
case GGML_OP_LOG:
|
||||
return ggml_is_contiguous_rows(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_UNARY:
|
||||
switch (ggml_get_unary_op(op)) {
|
||||
case GGML_UNARY_OP_TANH:
|
||||
@@ -1026,7 +1039,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
case GGML_UNARY_OP_EXP:
|
||||
case GGML_UNARY_OP_SOFTPLUS:
|
||||
case GGML_UNARY_OP_EXPM1:
|
||||
return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
|
||||
return ggml_is_contiguous_rows(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
|
||||
default:
|
||||
return false;
|
||||
}
|
||||
@@ -1054,11 +1067,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
case GGML_OP_MUL:
|
||||
case GGML_OP_DIV:
|
||||
case GGML_OP_ADD_ID:
|
||||
return op->src[0]->type == GGML_TYPE_F32;
|
||||
return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_ACC:
|
||||
case GGML_OP_REPEAT:
|
||||
case GGML_OP_SCALE:
|
||||
case GGML_OP_FILL:
|
||||
case GGML_OP_CONV_TRANSPOSE_1D:
|
||||
return true;
|
||||
case GGML_OP_CONV_TRANSPOSE_2D:
|
||||
@@ -1066,14 +1077,6 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32) &&
|
||||
op->src[1]->type == GGML_TYPE_F32 &&
|
||||
op->type == GGML_TYPE_F32;
|
||||
case GGML_OP_CLAMP:
|
||||
return op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_SQR:
|
||||
case GGML_OP_SQRT:
|
||||
case GGML_OP_SIN:
|
||||
case GGML_OP_COS:
|
||||
case GGML_OP_LOG:
|
||||
return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_SUM:
|
||||
return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
|
||||
case GGML_OP_TRI:
|
||||
@@ -1152,8 +1155,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
return has_simdgroup_reduction;
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
case GGML_OP_RWKV_WKV7:
|
||||
case GGML_OP_SOLVE_TRI:
|
||||
return true;
|
||||
case GGML_OP_SOLVE_TRI:
|
||||
case GGML_OP_MUL_MAT:
|
||||
case GGML_OP_MUL_MAT_ID:
|
||||
return has_simdgroup_reduction;
|
||||
@@ -1235,6 +1238,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
return false;
|
||||
};
|
||||
}
|
||||
case GGML_OP_DIAG:
|
||||
return true;
|
||||
case GGML_OP_OPT_STEP_ADAMW:
|
||||
case GGML_OP_OPT_STEP_SGD:
|
||||
return has_simdgroup_reduction;
|
||||
|
||||
@@ -80,6 +80,8 @@
|
||||
#define FC_SSM_CONV 900
|
||||
#define FC_SOLVE_TRI 1000
|
||||
#define FC_COUNT_EQUAL 1100
|
||||
#define FC_UNARY 1200
|
||||
#define FC_BIN 1300
|
||||
|
||||
// op-specific constants
|
||||
#define OP_FLASH_ATTN_EXT_NQPSG 8
|
||||
@@ -88,6 +90,35 @@
|
||||
#define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
|
||||
#define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
|
||||
|
||||
#define OP_UNARY_NUM_SCALE 10
|
||||
#define OP_UNARY_NUM_FILL 11
|
||||
#define OP_UNARY_NUM_CLAMP 12
|
||||
#define OP_UNARY_NUM_SQR 13
|
||||
#define OP_UNARY_NUM_SQRT 14
|
||||
#define OP_UNARY_NUM_SIN 15
|
||||
#define OP_UNARY_NUM_COS 16
|
||||
#define OP_UNARY_NUM_LOG 17
|
||||
#define OP_UNARY_NUM_LEAKY_RELU 18
|
||||
|
||||
#define OP_UNARY_NUM_TANH 100
|
||||
#define OP_UNARY_NUM_RELU 101
|
||||
#define OP_UNARY_NUM_SIGMOID 102
|
||||
#define OP_UNARY_NUM_GELU 103
|
||||
#define OP_UNARY_NUM_GELU_ERF 104
|
||||
#define OP_UNARY_NUM_GELU_QUICK 105
|
||||
#define OP_UNARY_NUM_SILU 106
|
||||
#define OP_UNARY_NUM_ELU 107
|
||||
#define OP_UNARY_NUM_NEG 108
|
||||
#define OP_UNARY_NUM_ABS 109
|
||||
#define OP_UNARY_NUM_SGN 110
|
||||
#define OP_UNARY_NUM_STEP 111
|
||||
#define OP_UNARY_NUM_HARDSWISH 112
|
||||
#define OP_UNARY_NUM_HARDSIGMOID 113
|
||||
#define OP_UNARY_NUM_EXP 114
|
||||
#define OP_UNARY_NUM_SOFTPLUS 115
|
||||
#define OP_UNARY_NUM_EXPM1 116
|
||||
|
||||
|
||||
// kernel argument structs
|
||||
//
|
||||
// - element counters (e.g. ne00) typically use int32_t to reduce register usage
|
||||
@@ -123,6 +154,31 @@ typedef struct {
|
||||
int32_t dim;
|
||||
} ggml_metal_kargs_concat;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00;
|
||||
int32_t ne01;
|
||||
int32_t ne02;
|
||||
int32_t ne03;
|
||||
uint64_t nb00;
|
||||
uint64_t nb01;
|
||||
uint64_t nb02;
|
||||
uint64_t nb03;
|
||||
int32_t ne0;
|
||||
int32_t ne1;
|
||||
int32_t ne2;
|
||||
int32_t ne3;
|
||||
uint64_t nb0;
|
||||
uint64_t nb1;
|
||||
uint64_t nb2;
|
||||
uint64_t nb3;
|
||||
float slope;
|
||||
float scale;
|
||||
float bias;
|
||||
float val;
|
||||
float min;
|
||||
float max;
|
||||
} ggml_metal_kargs_unary;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00;
|
||||
int32_t ne01;
|
||||
@@ -180,20 +236,6 @@ typedef struct {
|
||||
uint64_t nb3;
|
||||
} ggml_metal_kargs_repeat;
|
||||
|
||||
typedef struct {
|
||||
float scale;
|
||||
float bias;
|
||||
} ggml_metal_kargs_scale;
|
||||
|
||||
typedef struct {
|
||||
float val;
|
||||
} ggml_metal_kargs_fill;
|
||||
|
||||
typedef struct {
|
||||
float min;
|
||||
float max;
|
||||
} ggml_metal_kargs_clamp;
|
||||
|
||||
typedef struct {
|
||||
int64_t nk0;
|
||||
int64_t ne00;
|
||||
@@ -792,6 +834,25 @@ typedef struct {
|
||||
uint64_t nb3;
|
||||
} ggml_metal_kargs_set_rows;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00;
|
||||
int32_t ne01;
|
||||
int32_t ne02;
|
||||
int32_t ne03;
|
||||
uint64_t nb00;
|
||||
uint64_t nb01;
|
||||
uint64_t nb02;
|
||||
uint64_t nb03;
|
||||
int32_t ne0;
|
||||
int32_t ne1;
|
||||
int32_t ne2;
|
||||
int32_t ne3;
|
||||
uint64_t nb0;
|
||||
uint64_t nb1;
|
||||
uint64_t nb2;
|
||||
uint64_t nb3;
|
||||
} ggml_metal_kargs_diag;
|
||||
|
||||
typedef struct {
|
||||
int64_t ne00;
|
||||
int64_t ne01;
|
||||
@@ -861,10 +922,6 @@ typedef struct {
|
||||
int max_period;
|
||||
} ggml_metal_kargs_timestep_embedding;
|
||||
|
||||
typedef struct {
|
||||
float slope;
|
||||
} ggml_metal_kargs_leaky_relu;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00;
|
||||
int32_t ne01;
|
||||
|
||||
@@ -287,17 +287,9 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
|
||||
n_fuse = ggml_metal_op_acc(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_SCALE:
|
||||
{
|
||||
n_fuse = ggml_metal_op_scale(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_FILL:
|
||||
{
|
||||
n_fuse = ggml_metal_op_fill(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_CLAMP:
|
||||
{
|
||||
n_fuse = ggml_metal_op_clamp(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_LEAKY_RELU:
|
||||
case GGML_OP_SQR:
|
||||
case GGML_OP_SQRT:
|
||||
case GGML_OP_SIN:
|
||||
@@ -361,6 +353,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
|
||||
{
|
||||
n_fuse = ggml_metal_op_set_rows(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_DIAG:
|
||||
{
|
||||
n_fuse = ggml_metal_op_diag(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_L2_NORM:
|
||||
{
|
||||
n_fuse = ggml_metal_op_l2_norm(ctx, idx);
|
||||
@@ -422,10 +418,6 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
|
||||
{
|
||||
n_fuse = ggml_metal_op_top_k(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_LEAKY_RELU:
|
||||
{
|
||||
n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_TRI:
|
||||
{
|
||||
n_fuse = ggml_metal_op_tri(ctx, idx);
|
||||
@@ -703,7 +695,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
|
||||
/*.o1 =*/ { 0 },
|
||||
};
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_bin(lib, GGML_OP_ADD, 1, false);
|
||||
auto pipeline = ggml_metal_library_get_pipeline_bin_one(lib, GGML_OP_ADD);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
@@ -718,119 +710,6 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_scale(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
float scale;
|
||||
float bias;
|
||||
memcpy(&scale, ((const int32_t *) op->op_params) + 0, sizeof(float));
|
||||
memcpy(&bias, ((const int32_t *) op->op_params) + 1, sizeof(float));
|
||||
|
||||
ggml_metal_kargs_scale args = {
|
||||
/*.scale =*/ scale,
|
||||
/*.bias =*/ bias,
|
||||
};
|
||||
|
||||
int64_t n = ggml_nelements(op);
|
||||
|
||||
if (n % 4 == 0) {
|
||||
n /= 4;
|
||||
}
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_fill(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
const float val = ggml_get_op_params_f32(op, 0);
|
||||
|
||||
ggml_metal_kargs_fill args = {
|
||||
/*.val =*/ val
|
||||
};
|
||||
|
||||
int64_t n = ggml_nelements(op);
|
||||
|
||||
if (n % 4 == 0) {
|
||||
n /= 4;
|
||||
}
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_clamp(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
float min;
|
||||
float max;
|
||||
memcpy(&min, ((const int32_t *) op->op_params) + 0, sizeof(float));
|
||||
memcpy(&max, ((const int32_t *) op->op_params) + 1, sizeof(float));
|
||||
|
||||
ggml_metal_kargs_clamp args = {
|
||||
/*.min =*/ min,
|
||||
/*.max =*/ max,
|
||||
};
|
||||
|
||||
int64_t n = ggml_nelements(op);
|
||||
|
||||
if (n % 4 == 0) {
|
||||
n /= 4;
|
||||
}
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
@@ -842,19 +721,79 @@ int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
|
||||
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
int64_t n = ggml_nelements(op);
|
||||
GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
|
||||
|
||||
if (n % 4 == 0) {
|
||||
n /= 4;
|
||||
ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
|
||||
ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
|
||||
|
||||
ggml_metal_kargs_unary args = {
|
||||
/*.ne00 =*/ ne00,
|
||||
/*.ne01 =*/ ne01,
|
||||
/*.ne02 =*/ ne02,
|
||||
/*.ne03 =*/ ne03,
|
||||
/*.nb00 =*/ nb00,
|
||||
/*.nb01 =*/ nb01,
|
||||
/*.nb02 =*/ nb02,
|
||||
/*.nb03 =*/ nb03,
|
||||
/*.ne0 =*/ ne0,
|
||||
/*.ne1 =*/ ne1,
|
||||
/*.ne2 =*/ ne2,
|
||||
/*.ne3 =*/ ne3,
|
||||
/*.nb0 =*/ nb0,
|
||||
/*.nb1 =*/ nb1,
|
||||
/*.nb2 =*/ nb2,
|
||||
/*.nb3 =*/ nb3,
|
||||
/*.slope =*/ 0.0,
|
||||
/*.scale =*/ 0.0,
|
||||
/*.bias =*/ 0.0,
|
||||
/*.val =*/ 0.0,
|
||||
/*.min =*/ 0.0,
|
||||
/*.max =*/ 0.0,
|
||||
};
|
||||
|
||||
if (op->op == GGML_OP_LEAKY_RELU) {
|
||||
args.slope = ggml_get_op_params_f32(op, 0);
|
||||
}
|
||||
|
||||
if (op->op == GGML_OP_SCALE) {
|
||||
args.scale = ggml_get_op_params_f32(op, 0);
|
||||
args.bias = ggml_get_op_params_f32(op, 1);
|
||||
}
|
||||
|
||||
if (op->op == GGML_OP_FILL) {
|
||||
args.val = ggml_get_op_params_f32(op, 0);
|
||||
}
|
||||
|
||||
if (op->op == GGML_OP_CLAMP) {
|
||||
args.min = ggml_get_op_params_f32(op, 0);
|
||||
args.max = ggml_get_op_params_f32(op, 1);
|
||||
}
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 1);
|
||||
if (pipeline.c4) {
|
||||
args.ne00 = ne00/4;
|
||||
args.ne0 = ne0/4;
|
||||
}
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, bid_src0, 1);
|
||||
ggml_metal_encoder_set_buffer (enc, bid_dst, 2);
|
||||
|
||||
if (pipeline.cnt) {
|
||||
const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
} else {
|
||||
const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
|
||||
|
||||
const int nth = MIN(args.ne00, nth_max);
|
||||
|
||||
const int nk0 = (args.ne00 + nth - 1)/nth;
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, nk0*ne01, ne02, ne03, nth, 1, 1);
|
||||
}
|
||||
|
||||
return 1;
|
||||
}
|
||||
@@ -1259,6 +1198,48 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS(int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS(int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
ggml_metal_kargs_diag args = {
|
||||
/*.ne00 =*/ne00,
|
||||
/*.ne01 =*/ne01,
|
||||
/*.ne02 =*/ne02,
|
||||
/*.ne03 =*/ne03,
|
||||
/*.nb00 =*/nb00,
|
||||
/*.nb01 =*/nb01,
|
||||
/*.nb02 =*/nb02,
|
||||
/*.nb03 =*/nb03,
|
||||
/*.ne0 =*/ne0,
|
||||
/*.ne1 =*/ne1,
|
||||
/*.ne2 =*/ne2,
|
||||
/*.ne3 =*/ne3,
|
||||
/*.nb0 =*/nb0,
|
||||
/*.nb1 =*/nb1,
|
||||
/*.nb2 =*/nb2,
|
||||
/*.nb3 =*/nb3,
|
||||
};
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_diag(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 2);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, 32, 1, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
@@ -2849,8 +2830,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
|
||||
GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
|
||||
GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
|
||||
|
||||
bool bcast_row = false;
|
||||
|
||||
ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
|
||||
ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
|
||||
ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
|
||||
@@ -2944,18 +2923,7 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
|
||||
|
||||
struct ggml_metal_pipeline_with_params pipeline;
|
||||
|
||||
if (ggml_nelements(op->src[1]) == ne10 && ggml_is_contiguous(op->src[1]) && ne00 % 4 == 0 && ne10 % 4 == 0) {
|
||||
GGML_ASSERT(ggml_is_contiguous(op->src[0]));
|
||||
|
||||
// src1 is a row
|
||||
GGML_ASSERT(ne11 == 1);
|
||||
|
||||
pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, true);
|
||||
|
||||
bcast_row = true;
|
||||
} else {
|
||||
pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, false);
|
||||
}
|
||||
pipeline = ggml_metal_library_get_pipeline_bin(lib, op, n_fuse);
|
||||
|
||||
if (n_fuse > 1) {
|
||||
bid_dst = ggml_metal_get_buffer_id(ctx->node(idx + n_fuse - 1));
|
||||
@@ -2969,20 +2937,28 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
|
||||
}
|
||||
}
|
||||
|
||||
if (pipeline.c4) {
|
||||
args.ne00 = ne00/4;
|
||||
args.ne10 = ne10/4;
|
||||
args.ne0 = ne0/4;
|
||||
}
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, bid_src0, 1);
|
||||
ggml_metal_encoder_set_buffer (enc, bid_src1, 2);
|
||||
ggml_metal_encoder_set_buffer (enc, bid_dst, 3);
|
||||
|
||||
if (bcast_row) {
|
||||
const int64_t n = ggml_nelements(op)/4;
|
||||
if (pipeline.cnt) {
|
||||
const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
} else {
|
||||
int nth = 32;
|
||||
const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
|
||||
|
||||
while (16*nth < ne0 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
|
||||
int nth = 1;
|
||||
|
||||
while (2*nth < args.ne0 && nth < nth_max) {
|
||||
nth *= 2;
|
||||
}
|
||||
|
||||
@@ -4043,42 +4019,6 @@ int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
float slope;
|
||||
memcpy(&slope, op->op_params, sizeof(float));
|
||||
|
||||
ggml_metal_kargs_leaky_relu args = {
|
||||
/*.slope =*/ slope
|
||||
};
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
|
||||
|
||||
int64_t n = ggml_nelements(op);
|
||||
|
||||
if (n % 4 == 0) {
|
||||
n /= 4;
|
||||
}
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_tri(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
|
||||
@@ -46,9 +46,6 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
|
||||
int ggml_metal_op_concat (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_repeat (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_acc (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_scale (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_fill (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_clamp (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_unary (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_glu (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_sum (ggml_metal_op_t ctx, int idx);
|
||||
@@ -56,6 +53,7 @@ int ggml_metal_op_sum_rows (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_cumsum (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_get_rows (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_set_rows (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_diag (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_soft_max (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_ssm_conv (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_ssm_scan (ggml_metal_op_t ctx, int idx);
|
||||
@@ -85,7 +83,6 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_argmax (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_argsort (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_top_k (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_leaky_relu (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_tri (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_opt_step_adamw (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_opt_step_sgd (ggml_metal_op_t ctx, int idx);
|
||||
|
||||
@@ -7,6 +7,9 @@
|
||||
#include "ggml-metal-context.h"
|
||||
#include "ggml-metal-ops.h"
|
||||
|
||||
#include <mutex>
|
||||
#include <string>
|
||||
|
||||
#define GGML_METAL_NAME "MTL"
|
||||
#define GGML_METAL_MAX_DEVICES 16
|
||||
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -836,16 +836,9 @@ static inline void ggml_sycl_op_floor(ggml_backend_sycl_context & ctx, ggml_tens
|
||||
}
|
||||
|
||||
static inline void ggml_sycl_op_ceil(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
|
||||
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
|
||||
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
|
||||
const int num_blocks = ceil_div(k_elements, 256);
|
||||
stream->parallel_for(
|
||||
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
|
||||
sycl::range<1>(256)),
|
||||
[=](sycl::nd_item<1> item_ct1) {
|
||||
unary_op_ceil_kernel(src, dst_ptr, k_elements, item_ct1);
|
||||
});
|
||||
});
|
||||
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
|
||||
return op_ceil(x);
|
||||
});
|
||||
}
|
||||
|
||||
static inline void ggml_sycl_op_round(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
|
||||
|
||||
@@ -4591,9 +4591,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
|
||||
case GGML_UNARY_OP_EXP:
|
||||
case GGML_UNARY_OP_SOFTPLUS:
|
||||
case GGML_UNARY_OP_ELU:
|
||||
case GGML_UNARY_OP_CEIL:
|
||||
return true;
|
||||
case GGML_UNARY_OP_FLOOR:
|
||||
case GGML_UNARY_OP_CEIL:
|
||||
case GGML_UNARY_OP_ROUND:
|
||||
case GGML_UNARY_OP_TRUNC:
|
||||
#if defined (GGML_SYCL_F16)
|
||||
|
||||
@@ -36,7 +36,7 @@ apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor) {
|
||||
result.data = reinterpret_cast<uint64_t>(tensor->data);
|
||||
if (tensor->data) {
|
||||
if (!tensor->buffer) {
|
||||
GGML_ABORT("tensor has data but not buffer");
|
||||
GGML_ABORT("%s: tensor has data but not buffer", __func__);
|
||||
}
|
||||
// tensor->data is serialized as an offset to the buffer base address
|
||||
result.data -= reinterpret_cast<uint64_t>(BUFFER_TO_GGML_CONTEXT(tensor->buffer)->base);
|
||||
|
||||
@@ -27,7 +27,7 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
|
||||
|
||||
const void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
if (!shmem_data) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
return 1;
|
||||
}
|
||||
@@ -45,7 +45,7 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
|
||||
if (dev->iface.supports_op(dev, op)) {
|
||||
continue;
|
||||
}
|
||||
GGML_LOG_ERROR("Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
|
||||
|
||||
status = GGML_STATUS_ABORTED;
|
||||
apir_encode_ggml_status(enc, &status);
|
||||
|
||||
@@ -36,18 +36,22 @@ uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec
|
||||
ggml_backend_buffer_type_t buft;
|
||||
buft = apir_decode_ggml_buffer_type(dec);
|
||||
|
||||
size_t value = buft->iface.get_max_size(buft);
|
||||
size_t value = SIZE_MAX;
|
||||
if (buft->iface.get_max_size) {
|
||||
value = buft->iface.get_max_size(buft);
|
||||
}
|
||||
|
||||
apir_encode_size_t(enc, &value);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
|
||||
uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
|
||||
GGML_UNUSED(ctx);
|
||||
ggml_backend_buffer_type_t buft;
|
||||
buft = apir_decode_ggml_buffer_type(dec);
|
||||
GGML_UNUSED(dec);
|
||||
const bool is_host = false;
|
||||
|
||||
bool is_host = buft->iface.is_host(buft);
|
||||
apir_encode_bool_t(enc, &is_host);
|
||||
|
||||
return 0;
|
||||
|
||||
@@ -40,7 +40,7 @@ uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl
|
||||
void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
|
||||
if (!shmem_data) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -71,7 +71,7 @@ uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl
|
||||
|
||||
void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
if (!shmem_data) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -121,7 +121,7 @@ uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virg
|
||||
buffer = apir_decode_ggml_buffer(dec);
|
||||
|
||||
if (!apir_untrack_backend_buffer(buffer)) {
|
||||
GGML_LOG_WARN("%s: unknown buffer %p\n", __func__, (void *) buffer);
|
||||
GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: unknown buffer %p\n", __func__, (void *) buffer);
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
||||
@@ -124,7 +124,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
|
||||
|
||||
void * shmem_ptr = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
if (!shmem_ptr) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -17,26 +17,26 @@ uint64_t timer_count = 0;
|
||||
|
||||
uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) {
|
||||
if (reg != NULL) {
|
||||
GGML_LOG_WARN("%s: already initialized\n", __func__);
|
||||
GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: already initialized\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_ALREADY_INITED;
|
||||
}
|
||||
ggml_backend_reg_t (*ggml_backend_reg_fct)(void) = (ggml_backend_reg_t (*)()) ggml_backend_reg_fct_p;
|
||||
|
||||
reg = ggml_backend_reg_fct();
|
||||
if (reg == NULL) {
|
||||
GGML_LOG_ERROR("%s: backend registration failed\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend registration failed\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED;
|
||||
}
|
||||
|
||||
if (!reg->iface.get_device_count(reg)) {
|
||||
GGML_LOG_ERROR("%s: backend initialization failed: no device found\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device found\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_NO_DEVICE;
|
||||
}
|
||||
|
||||
dev = reg->iface.get_device(reg, 0);
|
||||
|
||||
if (!dev) {
|
||||
GGML_LOG_ERROR("%s: backend initialization failed: no device received\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device received\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_NO_DEVICE;
|
||||
}
|
||||
|
||||
|
||||
@@ -16,6 +16,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
|
||||
uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
|
||||
uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
@@ -62,7 +63,7 @@ static inline const char * backend_dispatch_command_name(ApirBackendCommandType
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
|
||||
return "backend_buffer_type_get_max_size";
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
|
||||
return "backend_buffer_type_is_host";
|
||||
return "backend_buffer_type_is_host (DEPRECATED)";
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
|
||||
return "backend_buffer_type_alloc_buffer";
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
|
||||
@@ -110,7 +111,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME = */ backend_buffer_type_get_name,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT = */ backend_buffer_type_get_alignment,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE = */ backend_buffer_type_get_max_size,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host /* DEPRECATED */,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER = */ backend_buffer_type_alloc_buffer,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = */ backend_buffer_type_get_alloc_size,
|
||||
|
||||
|
||||
@@ -11,6 +11,8 @@
|
||||
#include "shared/apir_cs.h"
|
||||
#include "shared/apir_cs_ggml.h"
|
||||
|
||||
#define GGML_VIRTGPU_BCK "ggml-virtgpu-backend: "
|
||||
|
||||
struct virgl_apir_context {
|
||||
uint32_t ctx_id;
|
||||
virgl_apir_callbacks * iface;
|
||||
|
||||
@@ -35,14 +35,8 @@ void apir_backend_deinit(uint32_t virgl_ctx_id) {
|
||||
buffer->iface.free_buffer(buffer);
|
||||
}
|
||||
|
||||
if (dev) {
|
||||
size_t free, total;
|
||||
dev->iface.get_memory(dev, &free, &total);
|
||||
GGML_LOG_INFO("%s: free memory: %ld MB\n", __func__, (size_t) free / 1024 / 1024);
|
||||
}
|
||||
|
||||
if (backend_library_handle) {
|
||||
GGML_LOG_INFO("%s: The GGML backend library was loaded. Unloading it.\n", __func__);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU_BCK "The GGML backend library was loaded. Unloading it.\n");
|
||||
dlclose(backend_library_handle);
|
||||
backend_library_handle = NULL;
|
||||
}
|
||||
@@ -65,7 +59,7 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
if (apir_logfile) {
|
||||
ggml_log_set(log_to_file_callback, apir_logfile);
|
||||
} else {
|
||||
GGML_LOG_INFO("Could not open the log file at '%s'\n", apir_log_to_file);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU_BCK "Could not open the log file at '%s'\n", apir_log_to_file);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -74,7 +68,10 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
|
||||
|
||||
if (!library_name) {
|
||||
GGML_LOG_ERROR("cannot open the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot open the GGML library: env var '%s' not defined\n",
|
||||
__func__, APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
|
||||
|
||||
|
||||
return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
|
||||
}
|
||||
@@ -82,13 +79,16 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
backend_library_handle = dlopen(library_name, RTLD_LAZY);
|
||||
|
||||
if (!backend_library_handle) {
|
||||
GGML_LOG_ERROR("cannot open the GGML library: %s\n", dlerror());
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot open the GGML library: %s\n", __func__, dlerror());
|
||||
|
||||
return APIR_LOAD_LIBRARY_CANNOT_OPEN;
|
||||
}
|
||||
|
||||
if (!library_reg) {
|
||||
GGML_LOG_ERROR("cannot register the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot register the GGML library: env var '%s' not defined\n",
|
||||
__func__, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
|
||||
|
||||
return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
|
||||
}
|
||||
@@ -96,8 +96,10 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg);
|
||||
dlsym_error = dlerror();
|
||||
if (dlsym_error) {
|
||||
GGML_LOG_ERROR("cannot find the GGML backend registration symbol '%s' (from %s): %s\n", library_reg,
|
||||
APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot find the GGML backend registration symbol '%s' (from %s): %s\n",
|
||||
__func__, library_reg, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
|
||||
|
||||
|
||||
return APIR_LOAD_LIBRARY_SYMBOL_MISSING;
|
||||
}
|
||||
@@ -134,7 +136,9 @@ uint32_t apir_backend_dispatcher(uint32_t virgl_ctx_id,
|
||||
};
|
||||
|
||||
if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) {
|
||||
GGML_LOG_ERROR("Received an invalid dispatch index (%d >= %d)\n", cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: Received an invalid dispatch index (%d >= %d)\n",
|
||||
__func__, cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
|
||||
return APIR_BACKEND_FORWARD_INDEX_INVALID;
|
||||
}
|
||||
|
||||
|
||||
@@ -86,7 +86,7 @@ static inline bool apir_decoder_peek_internal(apir_decoder * dec,
|
||||
assert(val_size <= size);
|
||||
|
||||
if (unlikely(size > (size_t) (dec->end - dec->cur))) {
|
||||
GGML_LOG_ERROR("reading too much from the decoder ...\n");
|
||||
GGML_LOG_ERROR("%s: reading too much from the decoder ...\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
memset(val, 0, val_size);
|
||||
return false;
|
||||
@@ -103,7 +103,7 @@ static inline void apir_decoder_peek(apir_decoder * dec, size_t size, void * val
|
||||
|
||||
static inline const void * apir_decoder_use_inplace(apir_decoder * dec, size_t size) {
|
||||
if (unlikely(size > (size_t) (dec->end - dec->cur))) {
|
||||
GGML_LOG_ERROR("reading too much from the decoder ...\n");
|
||||
GGML_LOG_ERROR("%s: reading too much from the decoder ...\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
return NULL;
|
||||
}
|
||||
@@ -221,7 +221,7 @@ static inline uint64_t apir_decode_array_size(apir_decoder * dec, uint64_t expec
|
||||
uint64_t size;
|
||||
apir_decode_uint64_t(dec, &size);
|
||||
if (size != expected_size) {
|
||||
GGML_LOG_ERROR("Couldn't decode array from the decoder\n");
|
||||
GGML_LOG_ERROR("%s: Couldn't decode array from the decoder\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
size = 0;
|
||||
}
|
||||
@@ -322,7 +322,7 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
|
||||
if (size) {
|
||||
val[size - 1] = '\0';
|
||||
} else {
|
||||
GGML_LOG_ERROR("Couldn't decode the blog array\n");
|
||||
GGML_LOG_ERROR("%s: Couldn't decode the blog array\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
}
|
||||
}
|
||||
@@ -332,7 +332,8 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
|
||||
static inline void * apir_decoder_alloc_array(size_t size, size_t count) {
|
||||
size_t alloc_size;
|
||||
if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) {
|
||||
GGML_LOG_ERROR("overflow in array allocation of %zu * %zu bytes\n", size, count);
|
||||
GGML_LOG_ERROR("%s: overflow in array allocation of %zu * %zu bytes\n",
|
||||
__func__, size, count);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
|
||||
@@ -39,11 +39,17 @@ static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor
|
||||
|
||||
static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
|
||||
const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec);
|
||||
|
||||
if (!apir_rpc_tensor) {
|
||||
return NULL;
|
||||
}
|
||||
|
||||
ggml_init_params params{
|
||||
/*.mem_size =*/ ggml_tensor_overhead(),
|
||||
/*.mem_buffer =*/ NULL,
|
||||
/*.no_alloc =*/ true,
|
||||
};
|
||||
|
||||
ggml_context * ctx = ggml_init(params);
|
||||
|
||||
const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor);
|
||||
@@ -71,6 +77,10 @@ static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decod
|
||||
return (ggml_backend_buffer_type_t) handle;
|
||||
}
|
||||
|
||||
static inline void apir_encode_apir_buffer_type_host_handle(apir_encoder * enc, apir_buffer_type_host_handle_t handle) {
|
||||
apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
|
||||
}
|
||||
|
||||
static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) {
|
||||
apir_buffer_type_host_handle_t handle;
|
||||
|
||||
@@ -154,13 +164,13 @@ static inline void apir_encode_ggml_tensor_inline(apir_encoder * enc, const ggml
|
||||
size_t tensor_size = sizeof(*tensor);
|
||||
|
||||
if (tensor->extra) {
|
||||
GGML_ABORT("Cannot pass tensors with extra");
|
||||
GGML_ABORT("%s: Cannot pass tensors with extra", __func__);
|
||||
}
|
||||
|
||||
if (tensor->src[0] && tensor->buffer) {
|
||||
static int first = 1;
|
||||
if (first) {
|
||||
GGML_LOG_WARN("Cannot pass tensors with src and buffer\n");
|
||||
GGML_LOG_WARN("%s: Cannot pass tensors with src and buffer\n", __func__);
|
||||
first = 0;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -6,7 +6,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
|
||||
|
||||
ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
|
||||
if (!context) {
|
||||
GGML_ABORT("Couldn't allocate the buffer context ...");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
|
||||
}
|
||||
|
||||
context->gpu = gpu;
|
||||
@@ -20,7 +20,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
|
||||
context->base = context->apir_context.shmem.mmap_ptr;
|
||||
context->is_from_ptr = true;
|
||||
} else {
|
||||
context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size);
|
||||
context->apir_context = apir_buffer_type_alloc_buffer(gpu, gpu->cached_buffer_type.host_handle, size);
|
||||
context->is_from_ptr = false;
|
||||
context->base = NULL;
|
||||
}
|
||||
@@ -34,36 +34,19 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
|
||||
static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
return apir_buffer_type_get_name(gpu, buft);
|
||||
return gpu->cached_buffer_type.name;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
static size_t align = 0;
|
||||
|
||||
if (align == 0) {
|
||||
align = apir_buffer_type_get_alignment(gpu, buft);
|
||||
}
|
||||
|
||||
return align;
|
||||
return gpu->cached_buffer_type.alignment;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
static size_t max_size = 0;
|
||||
if (max_size == 0) {
|
||||
max_size = apir_buffer_type_get_max_size(gpu, buft);
|
||||
}
|
||||
|
||||
return max_size;
|
||||
}
|
||||
|
||||
static bool ggml_backend_remoting_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
return apir_buffer_type_is_host(gpu, buft);
|
||||
return gpu->cached_buffer_type.max_size;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
|
||||
@@ -76,7 +59,7 @@ static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buff
|
||||
return ggml_nbytes(tensor);
|
||||
}
|
||||
|
||||
return apir_buffer_type_get_alloc_size(gpu, buft, tensor);
|
||||
return apir_buffer_type_get_alloc_size(gpu, gpu->cached_buffer_type.host_handle, tensor);
|
||||
}
|
||||
|
||||
const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = {
|
||||
|
||||
@@ -3,32 +3,27 @@
|
||||
static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
return apir_device_get_name(gpu);
|
||||
return gpu->cached_device_info.name;
|
||||
}
|
||||
|
||||
static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
return apir_device_get_description(gpu);
|
||||
// Return the pre-cached description from the virtgpu structure
|
||||
return gpu->cached_device_info.description;
|
||||
}
|
||||
|
||||
static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
static enum ggml_backend_dev_type type;
|
||||
static bool has_type = false;
|
||||
if (!has_type) {
|
||||
has_type = true;
|
||||
type = (enum ggml_backend_dev_type) apir_device_get_type(gpu);
|
||||
}
|
||||
|
||||
return type;
|
||||
return (enum ggml_backend_dev_type) gpu->cached_device_info.type;
|
||||
}
|
||||
|
||||
static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
return apir_device_get_memory(gpu, free, total);
|
||||
*free = gpu->cached_device_info.memory_free;
|
||||
*total = gpu->cached_device_info.memory_total;
|
||||
}
|
||||
|
||||
static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
|
||||
@@ -77,13 +72,22 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
|
||||
ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
|
||||
static std::atomic<bool> initialized = false;
|
||||
static ggml_backend_buffer_type buft;
|
||||
|
||||
static ggml_backend_buffer_type buft{
|
||||
/* .iface = */ ggml_backend_remoting_buffer_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) ctx,
|
||||
};
|
||||
if (!initialized) {
|
||||
static std::mutex mutex;
|
||||
std::lock_guard<std::mutex> lock(mutex);
|
||||
|
||||
if (!initialized) {
|
||||
buft = {
|
||||
/* .iface = */ ggml_backend_remoting_buffer_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) gpu->cached_buffer_type.host_handle,
|
||||
};
|
||||
initialized = true;
|
||||
}
|
||||
}
|
||||
|
||||
return &buft;
|
||||
}
|
||||
@@ -91,13 +95,22 @@ ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_bac
|
||||
static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
|
||||
static std::atomic<bool> initialized = false;
|
||||
static ggml_backend_buffer_type buft;
|
||||
|
||||
static ggml_backend_buffer_type buft{
|
||||
/* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) ctx,
|
||||
};
|
||||
if (!initialized) {
|
||||
static std::mutex mutex;
|
||||
std::lock_guard<std::mutex> lock(mutex);
|
||||
|
||||
if (!initialized) {
|
||||
buft = {
|
||||
/* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) gpu->cached_buffer_type.host_handle,
|
||||
};
|
||||
initialized = true;
|
||||
}
|
||||
}
|
||||
|
||||
return &buft;
|
||||
}
|
||||
@@ -110,7 +123,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_device_buffer_from_ptr(ggml_b
|
||||
|
||||
ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
|
||||
if (!context) {
|
||||
GGML_ABORT("Couldn't allocate the buffer context ...");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
|
||||
}
|
||||
|
||||
context->gpu = gpu;
|
||||
|
||||
@@ -4,37 +4,70 @@
|
||||
#include <iostream>
|
||||
#include <mutex>
|
||||
|
||||
void ggml_virtgpu_cleanup(virtgpu * gpu);
|
||||
|
||||
static virtgpu * apir_initialize() {
|
||||
static virtgpu * apir_gpu_instance = NULL;
|
||||
static bool apir_initialized = false;
|
||||
static virtgpu * gpu = NULL;
|
||||
static std::atomic<bool> initialized = false;
|
||||
|
||||
if (initialized) {
|
||||
// fast track
|
||||
return gpu;
|
||||
}
|
||||
|
||||
{
|
||||
static std::mutex mutex;
|
||||
std::lock_guard<std::mutex> lock(mutex);
|
||||
|
||||
if (apir_initialized) {
|
||||
return apir_gpu_instance;
|
||||
if (initialized) {
|
||||
// thread safe
|
||||
return gpu;
|
||||
}
|
||||
|
||||
apir_gpu_instance = create_virtgpu();
|
||||
if (!apir_gpu_instance) {
|
||||
GGML_ABORT("failed to initialize the virtgpu");
|
||||
gpu = create_virtgpu();
|
||||
if (!gpu) {
|
||||
initialized = true;
|
||||
return NULL;
|
||||
}
|
||||
|
||||
apir_initialized = true;
|
||||
// Pre-fetch and cache all device information, it will not change
|
||||
gpu->cached_device_info.description = apir_device_get_description(gpu);
|
||||
if (!gpu->cached_device_info.description) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device description", __func__);
|
||||
}
|
||||
gpu->cached_device_info.name = apir_device_get_name(gpu);
|
||||
if (!gpu->cached_device_info.name) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device name", __func__);
|
||||
}
|
||||
gpu->cached_device_info.device_count = apir_device_get_count(gpu);
|
||||
gpu->cached_device_info.type = apir_device_get_type(gpu);
|
||||
|
||||
apir_device_get_memory(gpu,
|
||||
&gpu->cached_device_info.memory_free,
|
||||
&gpu->cached_device_info.memory_total);
|
||||
|
||||
apir_buffer_type_host_handle_t buft_host_handle = apir_device_get_buffer_type(gpu);
|
||||
gpu->cached_buffer_type.host_handle = buft_host_handle;
|
||||
gpu->cached_buffer_type.name = apir_buffer_type_get_name(gpu, buft_host_handle);
|
||||
if (!gpu->cached_buffer_type.name) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu buffer type name", __func__);
|
||||
}
|
||||
gpu->cached_buffer_type.alignment = apir_buffer_type_get_alignment(gpu, buft_host_handle);
|
||||
gpu->cached_buffer_type.max_size = apir_buffer_type_get_max_size(gpu, buft_host_handle);
|
||||
|
||||
initialized = true;
|
||||
}
|
||||
|
||||
return apir_gpu_instance;
|
||||
return gpu;
|
||||
}
|
||||
|
||||
static int ggml_backend_remoting_get_device_count() {
|
||||
virtgpu * gpu = apir_initialize();
|
||||
if (!gpu) {
|
||||
GGML_LOG_WARN("apir_initialize failed\n");
|
||||
return 0;
|
||||
}
|
||||
|
||||
return apir_device_get_count(gpu);
|
||||
return gpu->cached_device_info.device_count;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) {
|
||||
@@ -52,17 +85,21 @@ ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device) {
|
||||
|
||||
static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
|
||||
if (devices.size() > 0) {
|
||||
GGML_LOG_INFO("%s: already initialized\n", __func__);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "%s: already initialized\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
virtgpu * gpu = apir_initialize();
|
||||
if (!gpu) {
|
||||
GGML_LOG_ERROR("apir_initialize failed\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: apir_initialize failed\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
static bool initialized = false;
|
||||
static std::atomic<bool> initialized = false;
|
||||
|
||||
if (initialized) {
|
||||
return; // fast track
|
||||
}
|
||||
|
||||
{
|
||||
static std::mutex mutex;
|
||||
@@ -70,10 +107,10 @@ static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
|
||||
if (!initialized) {
|
||||
for (int i = 0; i < ggml_backend_remoting_get_device_count(); i++) {
|
||||
ggml_backend_remoting_device_context * ctx = new ggml_backend_remoting_device_context;
|
||||
char desc[256] = "API Remoting device";
|
||||
char desc[256] = "ggml-virtgpu API Remoting device";
|
||||
|
||||
ctx->device = i;
|
||||
ctx->name = GGML_REMOTING_FRONTEND_NAME + std::to_string(i);
|
||||
ctx->name = GGML_VIRTGPU_NAME + std::to_string(i);
|
||||
ctx->description = desc;
|
||||
ctx->gpu = gpu;
|
||||
|
||||
@@ -98,7 +135,7 @@ static ggml_backend_dev_t ggml_backend_remoting_reg_get_device(ggml_backend_reg_
|
||||
static const char * ggml_backend_remoting_reg_get_name(ggml_backend_reg_t reg) {
|
||||
UNUSED(reg);
|
||||
|
||||
return GGML_REMOTING_FRONTEND_NAME;
|
||||
return GGML_VIRTGPU_NAME;
|
||||
}
|
||||
|
||||
static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
|
||||
@@ -111,8 +148,7 @@ static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
|
||||
ggml_backend_reg_t ggml_backend_virtgpu_reg() {
|
||||
virtgpu * gpu = apir_initialize();
|
||||
if (!gpu) {
|
||||
GGML_LOG_ERROR("virtgpu_apir_initialize failed\n");
|
||||
return NULL;
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: virtgpu_apir_initialize failed\n", __func__);
|
||||
}
|
||||
|
||||
static ggml_backend_reg reg = {
|
||||
@@ -129,9 +165,25 @@ ggml_backend_reg_t ggml_backend_virtgpu_reg() {
|
||||
|
||||
ggml_backend_remoting_reg_init_devices(®);
|
||||
|
||||
GGML_LOG_INFO("%s: initialized\n", __func__);
|
||||
|
||||
return ®
|
||||
}
|
||||
|
||||
// public function, not exposed in the GGML interface at the moment
|
||||
void ggml_virtgpu_cleanup(virtgpu * gpu) {
|
||||
if (gpu->cached_device_info.name) {
|
||||
free(gpu->cached_device_info.name);
|
||||
gpu->cached_device_info.name = NULL;
|
||||
}
|
||||
if (gpu->cached_device_info.description) {
|
||||
free(gpu->cached_device_info.description);
|
||||
gpu->cached_device_info.description = NULL;
|
||||
}
|
||||
if (gpu->cached_buffer_type.name) {
|
||||
free(gpu->cached_buffer_type.name);
|
||||
gpu->cached_buffer_type.name = NULL;
|
||||
}
|
||||
|
||||
mtx_destroy(&gpu->data_shmem_mutex);
|
||||
}
|
||||
|
||||
GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg)
|
||||
|
||||
@@ -8,6 +8,9 @@
|
||||
#include <memory>
|
||||
#include <string>
|
||||
|
||||
#define GGML_VIRTGPU_NAME "ggml-virtgpu"
|
||||
#define GGML_VIRTGPU "ggml-virtgpu: "
|
||||
|
||||
// USE_ALWAYS_TRUE_SUPPORTS_OP: 1 is fast, 0 avoid micro-benchmark crashes
|
||||
|
||||
#define USE_ALWAYS_TRUE_SUPPORTS_OP 1
|
||||
@@ -62,7 +65,7 @@ static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggm
|
||||
|
||||
static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
|
||||
if (!buffer->context) {
|
||||
GGML_ABORT("%s: no context available :/", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: no context available :/", __func__);
|
||||
}
|
||||
return BUFFER_TO_HOST_HANDLE(buffer);
|
||||
}
|
||||
|
||||
@@ -24,10 +24,10 @@ functions:
|
||||
frontend_return: "int"
|
||||
|
||||
get_name:
|
||||
frontend_return: "const char *"
|
||||
frontend_return: "char *"
|
||||
|
||||
get_description:
|
||||
frontend_return: "const char *"
|
||||
frontend_return: "char *"
|
||||
|
||||
get_type:
|
||||
frontend_return: "uint32_t"
|
||||
@@ -64,35 +64,33 @@ functions:
|
||||
group_description: "buffer-type"
|
||||
functions:
|
||||
get_name:
|
||||
frontend_return: "const char *"
|
||||
frontend_return: "char *"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
|
||||
get_alignment:
|
||||
frontend_return: "size_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
|
||||
get_max_size:
|
||||
frontend_return: "size_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
|
||||
is_host:
|
||||
frontend_return: "bool"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
deprecated: true
|
||||
|
||||
alloc_buffer:
|
||||
frontend_return: "apir_buffer_context_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buffer_buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
- "size_t size"
|
||||
|
||||
get_alloc_size:
|
||||
frontend_return: "size_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
- "const ggml_tensor *op"
|
||||
|
||||
buffer:
|
||||
|
||||
@@ -116,7 +116,7 @@ class RemotingCodebaseGenerator:
|
||||
'frontend_return': func_metadata.get('frontend_return', 'void'),
|
||||
'frontend_extra_params': func_metadata.get('frontend_extra_params', []),
|
||||
'group_description': group_description,
|
||||
'newly_added': func_metadata.get('newly_added', False)
|
||||
'deprecated': func_metadata.get('deprecated', False),
|
||||
})
|
||||
enum_value += 1
|
||||
|
||||
@@ -165,6 +165,9 @@ class RemotingCodebaseGenerator:
|
||||
|
||||
signature = "uint32_t"
|
||||
params = "apir_encoder *enc, apir_decoder *dec, virgl_apir_context *ctx"
|
||||
if func['deprecated']:
|
||||
decl_lines.append(f"/* {func['enum_name']} is deprecated. Keeping the handler for backward compatibility. */")
|
||||
|
||||
decl_lines.append(f"{signature} {func['backend_function']}({params});")
|
||||
|
||||
# Switch cases
|
||||
@@ -176,7 +179,9 @@ class RemotingCodebaseGenerator:
|
||||
switch_lines.append(f" /* {func['group_description']} */")
|
||||
current_group = func['group_name']
|
||||
|
||||
switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}\";")
|
||||
deprecated = " (DEPRECATED)" if func['deprecated'] else ""
|
||||
|
||||
switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}{deprecated}\";")
|
||||
|
||||
# Dispatch table
|
||||
table_lines = []
|
||||
@@ -188,7 +193,8 @@ class RemotingCodebaseGenerator:
|
||||
table_lines.append("")
|
||||
current_group = func['group_name']
|
||||
|
||||
table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']},")
|
||||
deprecated = " /* DEPRECATED */" if func['deprecated'] else ""
|
||||
table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']}{deprecated},")
|
||||
|
||||
header_content = f'''\
|
||||
#pragma once
|
||||
@@ -225,6 +231,10 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
|
||||
decl_lines.append(f"/* {func['group_description']} */")
|
||||
current_group = func['group_name']
|
||||
|
||||
if func['deprecated']:
|
||||
decl_lines.append(f"/* {func['frontend_function']} is deprecated. */")
|
||||
continue
|
||||
|
||||
# Build parameter list
|
||||
params = [self.naming_patterns['frontend_base_param']]
|
||||
params.extend(func['frontend_extra_params'])
|
||||
@@ -287,7 +297,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
|
||||
generated_files = [apir_backend_path, backend_dispatched_path, virtgpu_forward_path]
|
||||
|
||||
if not self.clang_format_available:
|
||||
logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted."
|
||||
logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted.\n"
|
||||
" Install clang-format to enable automatic code formatting.")
|
||||
else:
|
||||
logging.info("\n🎨 Formatting files with clang-format...")
|
||||
|
||||
@@ -18,12 +18,17 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
|
||||
|
||||
virtgpu_shmem temp_shmem; // Local storage for large buffers
|
||||
virtgpu_shmem * shmem = &temp_shmem;
|
||||
bool using_shared_shmem = false;
|
||||
|
||||
if (cgraph_size <= gpu->data_shmem.mmap_size) {
|
||||
// prefer the init-time allocated page, if large enough
|
||||
// Lock mutex before using shared data_shmem buffer
|
||||
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
|
||||
}
|
||||
using_shared_shmem = true;
|
||||
shmem = &gpu->data_shmem;
|
||||
} else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
|
||||
}
|
||||
|
||||
apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
|
||||
@@ -42,7 +47,10 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (shmem != &gpu->data_shmem) {
|
||||
// Unlock mutex before cleanup
|
||||
if (using_shared_shmem) {
|
||||
mtx_unlock(&gpu->data_shmem_mutex);
|
||||
} else {
|
||||
virtgpu_shmem_destroy(gpu, shmem);
|
||||
}
|
||||
|
||||
|
||||
@@ -1,20 +1,20 @@
|
||||
#include "virtgpu-forward-impl.h"
|
||||
|
||||
const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
char * apir_buffer_type_get_name(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
const size_t string_size = apir_decode_array_size_unchecked(decoder);
|
||||
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
if (!string) {
|
||||
GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
|
||||
apir_decoder_set_fatal(decoder);
|
||||
}
|
||||
apir_decode_char_array(decoder, string, string_size);
|
||||
@@ -24,14 +24,14 @@ const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t
|
||||
return string;
|
||||
}
|
||||
|
||||
size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
size_t apir_buffer_type_get_alignment(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
@@ -43,14 +43,14 @@ size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t
|
||||
return alignment;
|
||||
}
|
||||
|
||||
size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
size_t apir_buffer_type_get_max_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
@@ -62,26 +62,7 @@ size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t b
|
||||
return max_size;
|
||||
}
|
||||
|
||||
bool apir_buffer_type_is_host(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
bool is_host;
|
||||
apir_decode_bool_t(decoder, &is_host);
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
return is_host;
|
||||
}
|
||||
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_buffer_type_t buft, size_t size) {
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, size_t size) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -90,7 +71,7 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
apir_encode_size_t(encoder, &size);
|
||||
|
||||
@@ -103,14 +84,14 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
|
||||
return buffer_context;
|
||||
}
|
||||
|
||||
size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op) {
|
||||
size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, const ggml_tensor * op) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
apir_encode_ggml_tensor_inline(encoder, op);
|
||||
|
||||
|
||||
@@ -36,13 +36,18 @@ void apir_buffer_set_tensor(virtgpu * gpu,
|
||||
|
||||
virtgpu_shmem temp_shmem; // Local storage for large buffers
|
||||
virtgpu_shmem * shmem = &temp_shmem;
|
||||
bool using_shared_shmem = false;
|
||||
|
||||
if (size <= gpu->data_shmem.mmap_size) {
|
||||
// prefer the init-time allocated page, if large enough
|
||||
// Lock mutex before using shared data_shmem buffer
|
||||
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
|
||||
}
|
||||
using_shared_shmem = true;
|
||||
shmem = &gpu->data_shmem;
|
||||
|
||||
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
|
||||
}
|
||||
|
||||
memcpy(shmem->mmap_ptr, data, size);
|
||||
@@ -55,7 +60,10 @@ void apir_buffer_set_tensor(virtgpu * gpu,
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (shmem != &gpu->data_shmem) {
|
||||
// Unlock mutex before cleanup
|
||||
if (using_shared_shmem) {
|
||||
mtx_unlock(&gpu->data_shmem_mutex);
|
||||
} else {
|
||||
virtgpu_shmem_destroy(gpu, shmem);
|
||||
}
|
||||
|
||||
@@ -79,13 +87,18 @@ void apir_buffer_get_tensor(virtgpu * gpu,
|
||||
|
||||
virtgpu_shmem temp_shmem; // Local storage for large buffers
|
||||
virtgpu_shmem * shmem = &temp_shmem;
|
||||
bool using_shared_shmem = false;
|
||||
|
||||
if (size <= gpu->data_shmem.mmap_size) {
|
||||
// prefer the init-time allocated page, if large enough
|
||||
// Lock mutex before using shared data_shmem buffer
|
||||
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
|
||||
}
|
||||
using_shared_shmem = true;
|
||||
shmem = &gpu->data_shmem;
|
||||
|
||||
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
|
||||
}
|
||||
|
||||
apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
|
||||
@@ -98,7 +111,10 @@ void apir_buffer_get_tensor(virtgpu * gpu,
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (shmem != &gpu->data_shmem) {
|
||||
// Unlock mutex before cleanup
|
||||
if (using_shared_shmem) {
|
||||
mtx_unlock(&gpu->data_shmem_mutex);
|
||||
} else {
|
||||
virtgpu_shmem_destroy(gpu, shmem);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -2,11 +2,6 @@
|
||||
#include "virtgpu-shm.h"
|
||||
|
||||
int apir_device_get_count(virtgpu * gpu) {
|
||||
static int32_t dev_count = -1;
|
||||
if (dev_count != -1) {
|
||||
return dev_count;
|
||||
}
|
||||
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -14,6 +9,7 @@ int apir_device_get_count(virtgpu * gpu) {
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_COUNT);
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
int32_t dev_count = -1;
|
||||
apir_decode_int32_t(decoder, &dev_count);
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
@@ -21,11 +17,7 @@ int apir_device_get_count(virtgpu * gpu) {
|
||||
return dev_count;
|
||||
}
|
||||
|
||||
const char * apir_device_get_name(virtgpu * gpu) {
|
||||
static char * string = nullptr;
|
||||
if (string) {
|
||||
return string;
|
||||
}
|
||||
char * apir_device_get_name(virtgpu * gpu) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -34,9 +26,9 @@ const char * apir_device_get_name(virtgpu * gpu) {
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
const size_t string_size = apir_decode_array_size_unchecked(decoder);
|
||||
string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
if (!string) {
|
||||
GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
apir_decode_char_array(decoder, string, string_size);
|
||||
@@ -46,7 +38,7 @@ const char * apir_device_get_name(virtgpu * gpu) {
|
||||
return string;
|
||||
}
|
||||
|
||||
const char * apir_device_get_description(virtgpu * gpu) {
|
||||
char * apir_device_get_description(virtgpu * gpu) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -58,7 +50,7 @@ const char * apir_device_get_description(virtgpu * gpu) {
|
||||
const size_t string_size = apir_decode_array_size_unchecked(decoder);
|
||||
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
if (!string) {
|
||||
GGML_LOG_ERROR("%s: Could not allocate the device description buffer\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device description buffer\n", __func__);
|
||||
|
||||
return NULL;
|
||||
}
|
||||
@@ -181,7 +173,7 @@ apir_buffer_context_t apir_device_buffer_from_ptr(virtgpu * gpu, size_t size, si
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR);
|
||||
|
||||
if (virtgpu_shmem_create(gpu, size, &buffer_context.shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "Couldn't allocate the guest-host shared buffer");
|
||||
}
|
||||
|
||||
apir_encode_virtgpu_shmem_res_id(encoder, buffer_context.shmem.res_id);
|
||||
|
||||
@@ -11,7 +11,7 @@
|
||||
int32_t forward_flag = (int32_t) apir_command_type__; \
|
||||
encoder_name = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, forward_flag); \
|
||||
if (!encoder_name) { \
|
||||
GGML_ABORT("%s: failed to prepare the remote call encoder", __func__); \
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__); \
|
||||
} \
|
||||
} while (0)
|
||||
|
||||
@@ -19,10 +19,10 @@
|
||||
do { \
|
||||
ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL); \
|
||||
if (!decoder_name) { \
|
||||
GGML_ABORT("%s: failed to kick the remote call", __func__); \
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to kick the remote call", __func__); \
|
||||
} \
|
||||
if (ret_name < APIR_FORWARD_BASE_INDEX) { \
|
||||
GGML_ABORT("%s: failed to forward the API call: %s: code %d", __func__, \
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to forward the API call: %s: code %d", __func__, \
|
||||
apir_forward_error(ret_name), ret_name); \
|
||||
} \
|
||||
ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX); \
|
||||
|
||||
@@ -3,8 +3,8 @@
|
||||
/* device */
|
||||
void apir_device_get_device_count(struct virtgpu * gpu);
|
||||
int apir_device_get_count(struct virtgpu * gpu);
|
||||
const char * apir_device_get_name(struct virtgpu * gpu);
|
||||
const char * apir_device_get_description(struct virtgpu * gpu);
|
||||
char * apir_device_get_name(struct virtgpu * gpu);
|
||||
char * apir_device_get_description(struct virtgpu * gpu);
|
||||
uint32_t apir_device_get_type(struct virtgpu * gpu);
|
||||
void apir_device_get_memory(struct virtgpu * gpu, size_t * free, size_t * total);
|
||||
bool apir_device_supports_op(struct virtgpu * gpu, const ggml_tensor * op);
|
||||
@@ -17,14 +17,15 @@ void apir_device_get_props(struct virtgpu * gpu,
|
||||
apir_buffer_context_t apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
|
||||
|
||||
/* buffer-type */
|
||||
const char * apir_buffer_type_get_name(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
bool apir_buffer_type_is_host(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu,
|
||||
ggml_backend_buffer_type_t buffer_buft,
|
||||
size_t size);
|
||||
size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op);
|
||||
char * apir_buffer_type_get_name(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
|
||||
size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
|
||||
size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu,
|
||||
apir_buffer_type_host_handle_t host_handle,
|
||||
size_t size);
|
||||
size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu,
|
||||
apir_buffer_type_host_handle_t host_handle,
|
||||
const ggml_tensor * op);
|
||||
|
||||
/* buffer */
|
||||
void * apir_buffer_get_base(struct virtgpu * gpu, apir_buffer_context_t * buffer_context);
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user