ci : enable git lfs for build.yml

This reverts commit 97e40df5d6.

.gitattributes

@@ -0,0 +1 @@
+models/ggml-vocab-*.gguf filter=lfs diff=lfs merge=lfs -text

.github/workflows/build.yml

@@ -33,6 +33,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Dependencies
@@ -91,6 +92,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Dependencies
@@ -153,6 +155,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -188,6 +192,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -211,6 +217,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Dependencies
@@ -285,6 +292,8 @@ jobs:
 #      - name: Clone
 #        id: checkout
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        id: depends
@@ -319,6 +328,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -347,6 +358,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -369,6 +382,8 @@ jobs:
 
     steps:
       - uses: actions/checkout@v2
+        with:
+          lfs: true
 
       - name: add oneAPI to apt
         shell: bash
@@ -393,6 +408,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Build
         id: cmake_build
@@ -410,6 +427,8 @@ jobs:
 
     steps:
       - uses: actions/checkout@v2
+        with:
+          lfs: true
 
       - name: add oneAPI to apt
         shell: bash
@@ -434,6 +453,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Build
         id: cmake_build
@@ -454,6 +475,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -485,6 +508,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -514,6 +539,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -543,6 +570,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -576,6 +605,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -606,6 +637,8 @@ jobs:
     steps:
       - name: Clone
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Setup ${{ matrix.sys }}
         uses: msys2/setup-msys2@v2
@@ -687,6 +720,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Clone Kompute submodule
@@ -833,6 +867,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - uses: Jimver/cuda-toolkit@v0.2.11
@@ -906,6 +941,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Install
@@ -947,6 +983,8 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
@@ -957,6 +995,8 @@ jobs:
     steps:
       - name: Clone
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Set up JDK
         uses: actions/setup-java@v3
@@ -979,7 +1019,9 @@ jobs:
 #    runs-on: macos-12
 #    steps:
 #    - name: Clone
-#      uses: actions/checkout@v4
+#      uses: actions/checkout@#v4
+#        with:
+#          lfs: true
 #
 #    - name: Build
 #      uses: cross-platform-actions/action@v0.19.0
@@ -1012,6 +1054,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Determine tag name
@@ -1077,6 +1120,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |
@@ -1101,6 +1146,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |
@@ -1125,6 +1172,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |
@@ -1155,6 +1204,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1194,6 +1245,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1240,6 +1293,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |

.github/workflows/code-coverage.yml

@@ -13,14 +13,16 @@ jobs:
   run:
     runs-on: ubuntu-20.04
     steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
       - name: Dependencies
         run: |
           sudo apt-get update
           sudo apt-get install build-essential gcc-8 lcov
 
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+
       - name: Build
         run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
 

README.md

@@ -20,7 +20,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- **BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920**
+- **Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021**
+- BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
 - Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
@@ -935,17 +936,25 @@ If your issue is with model generation quality, then please at least scan the fo
 
 ### Android
 
+#### Build on Android using Termux
+[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
+
 #### Building the Project using Android NDK
-You can easily run `llama.cpp` on Android device with [termux](https://termux.dev/).
-
-First, install the essential packages for termux:
-```
-pkg install clang wget git cmake
-```
-Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
-
-You can execute the following commands on your computer to avoid downloading the NDK to your mobile. Of course, you can also do this in Termux.
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
 
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
 ```
 $ mkdir build-android
 $ cd build-android
@@ -953,7 +962,9 @@ $ export NDK=<your_ndk_directory>
 $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
 $ make
 ```
-Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
 Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
 
 (Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
@@ -975,25 +986,10 @@ $cd /data/data/com.termux/files/home/bin
 $./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
 ```
 
-Here is a demo of an interactive session running on Pixel 5 phone:
+Here's a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
-#### Build on Android using Termux
-[Termux](https://github.com/termux/termux-app#installation) is an alternative to execute `llama.cpp` on an Android device (no root required).
-```
-apt update && apt upgrade -y
-apt install git
-```
-
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
-
-[Follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
-
 ### Docker
 
 #### Prerequisites

common/sampling.cpp

@@ -35,6 +35,8 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
     result->prev.resize(params.n_prev);
 
+    result->n_considered = 0;
+
     llama_sampling_set_rng_seed(result, params.seed);
 
     return result;
@@ -64,6 +66,7 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
 
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
     ctx->cur.clear();
+    ctx->n_considered = 0;
 }
 
 void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
@@ -253,6 +256,8 @@ static llama_token llama_sampling_sample_impl(
         }
     }
 
+    ctx_sampling->n_considered = cur_p.size;
+
     return id;
 }
 

common/sampling.h

@@ -81,6 +81,7 @@ struct llama_sampling_context {
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
+    size_t n_considered;
 
     std::mt19937 rng;
 };

convert-hf-to-gguf-update.py

@@ -67,6 +67,7 @@ models = [
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
     {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
     {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
+    {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
 ]
 
 # make directory "models/tokenizers" if it doesn't exist

convert-hf-to-gguf.py

@@ -314,6 +314,9 @@ class Model(ABC):
         if chkhsh == "9c2227e4dd922002fb81bde4fc02b0483ca4f12911410dee2255e4987644e3f8":
             # ref: https://huggingface.co/CohereForAI/c4ai-command-r-v01
             res = "command-r"
+        if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
+            # ref: https://huggingface.co/allenai/OLMo-1.7-7B-hf
+            res = "olmo"
 
         if res is None:
             logger.warning("\n")
@@ -2831,8 +2834,9 @@ class OlmoModel(Model):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_layer_norm_eps(1e-5)
-        if "clip_qkv" in self.hparams is not None:
-            self.gguf_writer.add_clamp_kqv(self.hparams["clip_qkv"])
+        clip_qkv = self.hparams.get("clip_qkv")
+        if clip_qkv is not None:
+            self.gguf_writer.add_clamp_kqv(clip_qkv)
 
     # Same as super class, but permuting q_proj, k_proj
     # Copied from: LlamaModel

examples/finetune/finetune.cpp

@@ -575,7 +575,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
 
     auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) {
+        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16 || a->type == GGML_TYPE_BF16) {
             return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
         } else if (a->type == GGML_TYPE_F32) {
             return ggml_add(ctx, a, b);

examples/imatrix/imatrix.cpp

@@ -19,6 +19,7 @@
 
 struct Stats {
     std::vector<float> values;
+    std::vector<int> counts;
     int ncall = 0;
 };
 
@@ -121,12 +122,10 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         auto & e = m_stats[wname];
 
         ++e.ncall;
-        // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
-        //       using the following line, we can correct for that if needed by replacing the line above with:
-        //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
 
         if (e.values.empty()) {
             e.values.resize(src1->ne[0]*n_as, 0);
+            e.counts.resize(src1->ne[0]*n_as, 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
@@ -153,6 +152,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
                     for (int j = 0; j < (int)src1->ne[0]; ++j) {
                         e.values[e_start + j] += x[j]*x[j];
+                        e.counts[e_start + j]++;
                     }
                 }
             }
@@ -170,6 +170,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         auto& e = m_stats[wname];
         if (e.values.empty()) {
             e.values.resize(src1->ne[0], 0);
+            e.counts.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
@@ -183,6 +184,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             const float * x = data + row * src1->ne[0];
             for (int j = 0; j < (int)src1->ne[0]; ++j) {
                 e.values[j] += x[j]*x[j];
+                e.counts[j]++;
             }
         }
         if (e.ncall > m_last_call) {
@@ -222,7 +224,13 @@ void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) co
         out.write((const char *) &p.second.ncall, sizeof(p.second.ncall));
         int nval = p.second.values.size();
         out.write((const char *) &nval, sizeof(nval));
-        if (nval > 0) out.write((const char *) p.second.values.data(), nval * sizeof(float));
+        if (nval > 0) {
+            std::vector<float> tmp(nval);
+            for (int i = 0; i < nval; i++) {
+                tmp[i] = (p.second.values[i] / static_cast<float>(p.second.counts[i])) * static_cast<float>(p.second.ncall);
+            }
+            out.write((const char*)tmp.data(), nval*sizeof(float));
+        }
     }
 
     // Write the number of call the matrix was computed with
@@ -270,14 +278,28 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
             imatrix_data = {};
             return false;
         }
-        e.values.resize(nval);
-        in.read((char*)e.values.data(), nval*sizeof(float));
+
+        // When re-called from load_imatrix() with add set, this will already be created.
+        if (e.values.empty()) {
+            e.values.resize(nval, 0);
+            e.counts.resize(nval, 0);
+        }
+
+        std::vector<float> tmp(nval);
+        in.read((char*)tmp.data(), nval*sizeof(float));
         if (in.fail()) {
             printf("%s: failed reading data for entry %d\n",__func__,i);
             imatrix_data = {};
             return false;
         }
-        e.ncall = ncall;
+
+        // Recreate the state as expected by save_imatrix(), and corerct for weighted sum.
+        for (int i = 0; i < nval; i++) {
+            e.values[i] += tmp[i];
+            e.counts[i] += ncall;
+        }
+        e.ncall += ncall;
+
     }
     return true;
 }

examples/quantize/quantize.cpp

@@ -46,7 +46,8 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
     { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, +0.0008 ppl @ LLaMA-v1-7B", },
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
-    { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
+    { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "14.00G, -0.0020 ppl @ Mistral-7B", },
+    { "BF16",   LLAMA_FTYPE_MOSTLY_BF16,   "14.00G, -0.0050 ppl @ Mistral-7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
     // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
     { "COPY",   LLAMA_FTYPE_ALL_F32,       "only copy tensors, no quantizing", },

examples/server/README.md

@@ -62,6 +62,18 @@ page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/
 - `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name.  Default: template taken from model's metadata. We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
 - `--log-disable`: Output logs to stdout only, not to `llama.log`. Default: enabled
 - `--log-format FORMAT`: Define the log output to FORMAT: json or text Default: `json`
+- `--rope-scaling` : RoPE scaling method. Defaults to linear unless otherwise specified by the model. Options are `none`, `linear`, `yarn`
+- `--rope-freq-base N` : RoPE frequency base (default: loaded from model)
+- `--rope-freq-scale N`: RoPE frequency scaling factor, expands context by a factor of 1/N (e.g. 0.25)
+- `--yarn-ext-factor N` : YaRN: extrapolation mix factor (Default: 1.0, 0.0 = full interpolation)
+- `--yarn-attn-factor N` : YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
+- `--yarn-beta-slow N`: YaRN: High correction dim or alpha (default: 1.0)
+- `--yarn-beta-fast N`: YaRN: low correction dim or beta (default: 32.0)
+- `--pooling` : Pooling type for embeddings, use model default if unspecified. Options are `none`, `mean`, `cls`
+- `-dt N`, `--defrag-thold N`: KV cache defragmentation threshold (default: -1.0, < 0 = disabled)
+- `-fa`, `--flash-attn` : enable flash attention (default: disabled).
+- `-ctk TYPE`, `--cache-type-k TYPE` : KV cache data type for K (default: `f16`, options `f32`, `f16`, `q8_0`, `q4_0`, `q4_1`, `iq4_nl`, `q5_0`, or `q5_1`)
+- `-ctv TYPE`, `--cache-type-v TYPE` : KV cache type for V (default `f16`, see `-ctk` for options)
 
 **If compiled with `LLAMA_SERVER_SSL=ON`**
 - `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key
@@ -260,7 +272,7 @@ node index.js
 
     `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. Default: `[]`
 
-    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token. Default: `0`
+    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token given the sampling settings. Note that for temperature < 0 the tokens are sampled greedily but token probabilities are still being calculated via a simple softmax of the logits without considering any other sampler settings. Default: `0`
 
     `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum. Default: `0`
 

examples/server/server.cpp

@@ -2266,17 +2266,31 @@ struct server_context {
                 llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
                 result.tok = id;
 
-                const int32_t n_probs = slot.sparams.n_probs;
-                if (slot.sparams.temp <= 0 && n_probs > 0) {
-                    // for llama_sample_token_greedy we need to sort candidates
-                    llama_sample_softmax(ctx, &cur_p);
-                }
+                const size_t n_probs = std::min(cur_p.size, (size_t) slot.sparams.n_probs);
+                if (n_probs > 0) {
+                    const size_t n_considered = slot.ctx_sampling->n_considered;
 
-                for (size_t i = 0; i < std::min(cur_p.size, (size_t) n_probs); ++i) {
-                    result.probs.push_back({
-                        cur_p.data[i].id,
-                        cur_p.data[i].p
-                    });
+                    // Make sure at least n_probs top tokens are at the front of the vector:
+                    if (slot.sparams.temp == 0.0f && n_probs > n_considered) {
+                        llama_sample_top_k(ctx, &cur_p, n_probs, 0);
+                    }
+
+                    if (slot.sparams.temp == 0.0f) {
+                        // With greedy sampling the probabilities have possibly not been calculated.
+                        for (size_t i = 0; i < n_probs; ++i) {
+                            result.probs.push_back({
+                                cur_p.data[i].id,
+                                i == 0 ? 1.0f : 0.0f
+                            });
+                        }
+                    } else {
+                        for (size_t i = 0; i < n_probs; ++i) {
+                            result.probs.push_back({
+                                cur_p.data[i].id,
+                                i >= n_considered ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
+                            });
+                        }
+                    }
                 }
 
                 if (!process_token(result, slot)) {

ggml-impl.h

@@ -17,6 +17,83 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
+ * Subnormals shall be flushed to zero, and NANs will be quiet.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    if (!(u.i & 0x7f800000)) { /* subnormal */
+        h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
+
 #ifdef __cplusplus
 extern "C" {
 #endif

ggml-metal.m

@@ -803,7 +803,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_GET_ROWS:
             {
-                return op->ne[3] == 1;
+                return op->src[0]->type != GGML_TYPE_BF16 && op->ne[3] == 1;
             }
         default:
             return false;

ggml-metal.metal

@@ -2175,7 +2175,7 @@ kernel void kernel_flash_attn_ext_f16(
 
     const short D4 = D/4;
     const short D8 = D/8;
-    const short Q8 = Q/8;
+  //const short Q8 = Q/8;
     const short NW = N_SIMDWIDTH;
     const short SH = (C + Q); // shared memory per simdgroup in (half)
 

ggml-quants.c

@@ -12450,6 +12450,24 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
     const size_t nb = nbytes/ggml_type_size(type);
 
     switch (type) {
+        case GGML_TYPE_BF16:
+            {
+                int nans = 0;
+                int infs = 0;
+                const unsigned short * f = (const unsigned short *) data;
+                for (size_t i = 0; i < nb; ++i) {
+                    nans += (f[i] & 0x7fff) > 0x7f80;
+                    infs += (f[i] & 0x7fff) == 0x7f80;
+                }
+                if (nans) {
+                    fprintf(stderr, "%s: found %d NaNs in row of %zu BF16 values\n", __func__, nans, nb);
+                    return false;
+                }
+                if (infs) {
+                    fprintf(stderr, "%s: found %d infinities in row of %zu BF16 values\n", __func__, infs, nb);
+                    return false;
+                }
+            } break;
         case GGML_TYPE_F16:
             {
                 const ggml_fp16_t * f = (const ggml_fp16_t *) data;

ggml.h

@@ -326,14 +326,20 @@ extern "C" {
     // get ggml_status name string
     GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status);
 
+    // ieee 754-2008 half-precision float16
+    // todo: make this not an integral type
     typedef uint16_t ggml_fp16_t;
+    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t);
+    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float);
+    GGML_API void        ggml_fp16_to_fp32_row(const ggml_fp16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_fp16_row(const float *, ggml_fp16_t *, int64_t);
 
-    // convert FP16 <-> FP32
-    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
-    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
-
-    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n);
-    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n);
+    // google brain half-precision bfloat16
+    typedef struct { uint16_t bits; } ggml_bf16_t;
+    GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
+    GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
+    GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
     struct ggml_object;
     struct ggml_context;
@@ -370,6 +376,7 @@ extern "C" {
         GGML_TYPE_I64     = 27,
         GGML_TYPE_F64     = 28,
         GGML_TYPE_IQ1_M   = 29,
+        GGML_TYPE_BF16    = 30,
         GGML_TYPE_COUNT,
     };
 
@@ -410,6 +417,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ2_S   = 21, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
+        GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
     };
 
     // available tensor operations:

gguf-py/gguf/constants.py

@@ -817,6 +817,7 @@ class GGMLQuantizationType(IntEnum):
     I64     = 27
     F64     = 28
     IQ1_M   = 29
+    BF16    = 30
 
 
 class GGUFEndian(IntEnum):
@@ -888,6 +889,7 @@ GGML_QUANT_SIZES = {
     GGMLQuantizationType.I64:     (1, 8),
     GGMLQuantizationType.F64:     (1, 8),
     GGMLQuantizationType.IQ1_M:   (256, QK_K // 8 + QK_K // 16  + QK_K // 32),
+    GGMLQuantizationType.BF16:    (1, 2),
 }
 
 

llama.cpp

@@ -3175,6 +3175,7 @@ struct llama_model_loader {
             switch (type_max) {
                 case GGML_TYPE_F32:     ftype = LLAMA_FTYPE_ALL_F32;        break;
                 case GGML_TYPE_F16:     ftype = LLAMA_FTYPE_MOSTLY_F16;     break;
+                case GGML_TYPE_BF16:    ftype = LLAMA_FTYPE_MOSTLY_BF16;    break;
                 case GGML_TYPE_Q4_0:    ftype = LLAMA_FTYPE_MOSTLY_Q4_0;    break;
                 case GGML_TYPE_Q4_1:    ftype = LLAMA_FTYPE_MOSTLY_Q4_1;    break;
                 case GGML_TYPE_Q5_0:    ftype = LLAMA_FTYPE_MOSTLY_Q5_0;    break;
@@ -3666,6 +3667,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
     switch (ftype) {
         case LLAMA_FTYPE_ALL_F32:     return "all F32";
         case LLAMA_FTYPE_MOSTLY_F16:  return "F16";
+        case LLAMA_FTYPE_MOSTLY_BF16: return "BF16";
         case LLAMA_FTYPE_MOSTLY_Q4_0: return "Q4_0";
         case LLAMA_FTYPE_MOSTLY_Q4_1: return "Q4_1";
         case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
@@ -4389,6 +4391,9 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "command-r") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_COMMAND_R;
+            } else if (
+                tokenizer_pre == "olmo") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_OLMO;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -6126,6 +6131,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
             || !(
                 model.ftype == LLAMA_FTYPE_ALL_F32 ||
                 model.ftype == LLAMA_FTYPE_MOSTLY_F16 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_BF16 ||
                 model.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
                 model.ftype == LLAMA_FTYPE_MOSTLY_Q4_1
             )
@@ -12248,6 +12254,7 @@ struct llm_tokenizer_bpe {
                         });
                         break;
                     case LLAMA_VOCAB_PRE_TYPE_GPT2:
+                    case LLAMA_VOCAB_PRE_TYPE_OLMO:
                         word_collection = unicode_regex_split(text, {
                             "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                         });
@@ -14154,13 +14161,16 @@ static void llama_tensor_dequantize_internal(
         if (qtype.to_float == NULL) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
         }
-    } else if (tensor->type != GGML_TYPE_F16) {
+    } else if (tensor->type != GGML_TYPE_F16 &&
+               tensor->type != GGML_TYPE_BF16) {
         throw std::runtime_error(format("cannot dequantize/convert tensor type %s", ggml_type_name(tensor->type)));
     }
 
     if (nthread < 2) {
         if (tensor->type == GGML_TYPE_F16) {
             ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor->data, f32_output, nelements);
+        } else if (tensor->type == GGML_TYPE_BF16) {
+            ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
         } else if (ggml_is_quantized(tensor->type)) {
             qtype.to_float(tensor->data, f32_output, nelements);
         } else {
@@ -14169,7 +14179,14 @@ static void llama_tensor_dequantize_internal(
         return;
     }
 
-    size_t block_size = tensor->type == GGML_TYPE_F16 ? 1 : (size_t)ggml_blck_size(tensor->type);
+    size_t block_size;
+    if (tensor->type == GGML_TYPE_F16 ||
+        tensor->type == GGML_TYPE_BF16) {
+        block_size = 1;
+    } else {
+        block_size = (size_t)ggml_blck_size(tensor->type);
+    }
+
     size_t block_size_bytes = ggml_type_size(tensor->type);
 
     GGML_ASSERT(nelements % block_size == 0);
@@ -14188,6 +14205,8 @@ static void llama_tensor_dequantize_internal(
         auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) {
             if (typ == GGML_TYPE_F16) {
                 ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
+            } else if (typ == GGML_TYPE_BF16) {
+                ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
             } else {
                 qtype.to_float(inbuf, outbuf, nels);
             }
@@ -14548,6 +14567,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_1: default_type = GGML_TYPE_Q5_1; break;
         case LLAMA_FTYPE_MOSTLY_Q8_0: default_type = GGML_TYPE_Q8_0; break;
         case LLAMA_FTYPE_MOSTLY_F16:  default_type = GGML_TYPE_F16;  break;
+        case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
         case LLAMA_FTYPE_ALL_F32:     default_type = GGML_TYPE_F32;  break;
 
         // K-quants

llama.h

@@ -81,6 +81,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
         LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
         LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
+        LLAMA_VOCAB_PRE_TYPE_OLMO           = 10,
     };
 
     // note: these values should be synchronized with ggml_rope
@@ -136,6 +137,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };

tests/test-backend-ops.cpp

@@ -50,7 +50,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
 
     if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
         ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
-    } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16) {
+    } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
         GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
         std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
         std::vector<float> imatrix(tensor->ne[0], 1.0f); // dummy importance matrix
@@ -92,6 +92,8 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                     size_t i = i3*t->nb[3] + i2*t->nb[2] + i1*t->nb[1] + i0/bs*t->nb[0];
                     if (t->type == GGML_TYPE_F16) {
                         tv.push_back(ggml_fp16_to_fp32(*(ggml_fp16_t*)&buf[i]));
+                    } else if (t->type == GGML_TYPE_BF16) {
+                        tv.push_back(ggml_bf16_to_fp32(*(ggml_bf16_t*)&buf[i]));
                     } else if (t->type == GGML_TYPE_F32) {
                         tv.push_back(*(float *) &buf[i]);
                     } else if (t->type == GGML_TYPE_I32) {
@@ -1898,7 +1900,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     std::default_random_engine rng(0);
 
     const ggml_type all_types[] = {
-        GGML_TYPE_F32, GGML_TYPE_F16,
+        GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16,
         GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
         GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
         GGML_TYPE_Q8_0,