metal : minor

common : don't do warm-up with more than n_batch tokens (close #3058 )
ggml-ci
2026-04-16 16:27:32 +03:00 · 2023-09-07 15:33:21 +03:00 · 2023-09-07 15:32:38 +03:00 · 2023-09-07 15:20:07 +03:00 · 2023-09-07 14:59:48 +03:00 · 2023-09-07 14:11:55 +03:00
29 changed files with 2768 additions and 2432 deletions
--- a/.editorconfig
+++ b/.editorconfig
@@ -17,3 +17,6 @@ indent_style = tab

 [prompts/*.txt]
 insert_final_newline = unset
+
+[examples/server/public/*]
+indent_size = 2
--- a/.gitignore
+++ b/.gitignore
@@ -31,27 +31,29 @@ tmp/
 models/*
 models-mnt

-/main
-/quantize
-/quantize-stats
-/result
-/perplexity
-/embedding
-/train-text-from-scratch
-/convert-llama2c-to-ggml
-/simple
-/benchmark-matmult
-/vdot
-/server
 /Pipfile
+/baby-llama
+/beam-search
+/benchmark-matmult
+/convert-llama2c-to-ggml
 /embd-input-test
+/embedding
 /gguf
 /gguf-llama-simple
 /libllama.so
 /llama-bench
-/baby-llama
-/beam-search
+/main
+/metal
+/perplexity
+/quantize
+/quantize-stats
+/result
 /save-load-state
+/server
+/simple
+/speculative
+/train-text-from-scratch
+/vdot
 build-info.h
 arm_neon.h
 compile_commands.json
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -36,6 +36,12 @@ endif()
 # Option list
 #

+if (APPLE)
+    set(LLAMA_METAL_DEFAULT ON)
+else()
+    set(LLAMA_METAL_DEFAULT OFF)
+endif()
+
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
 option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
@@ -76,7 +82,8 @@ option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
-option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
+option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
+option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
@@ -158,6 +165,33 @@ if (APPLE AND LLAMA_ACCELERATE)
    endif()
 endif()

+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+    message(STATUS "Metal framework found")
+
+    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+    add_compile_definitions(GGML_USE_METAL)
+    if (LLAMA_METAL_NDEBUG)
+        add_compile_definitions(GGML_METAL_NDEBUG)
+    endif()
+
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        )
+endif()
+
 if (LLAMA_BLAS)
    if (LLAMA_STATIC)
        set(BLA_STATIC ON)
@@ -293,29 +327,6 @@ if (LLAMA_CUBLAS)
    endif()
 endif()

-if (LLAMA_METAL)
-    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-
-    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
-
-    add_compile_definitions(GGML_USE_METAL)
-    #add_compile_definitions(GGML_METAL_NDEBUG)
-
-    # get full path to the file
-    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
-    # copy ggml-metal.metal to bin directory
-    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
-        ${FOUNDATION_LIBRARY}
-        ${METAL_FRAMEWORK}
-        ${METALKIT_FRAMEWORK}
-        )
-endif()
-
 if (LLAMA_MPI)
    cmake_minimum_required(VERSION 3.10)
    find_package(MPI)
--- a/89
+++ b/89
@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative tests/test-c.o

 # Binaries only useful for tests
 TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1
@@ -7,6 +7,39 @@ TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-dou
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report

+ifndef UNAME_S
+UNAME_S := $(shell uname -s)
+endif
+
+ifndef UNAME_P
+UNAME_P := $(shell uname -p)
+endif
+
+ifndef UNAME_M
+UNAME_M := $(shell uname -m)
+endif
+
+# Mac OS + Arm can report x86_64
+# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
+ifeq ($(UNAME_S),Darwin)
+	ifndef LLAMA_NO_METAL
+		LLAMA_METAL := 1
+	endif
+
+	ifneq ($(UNAME_P),arm)
+		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
+		ifeq ($(SYSCTL_M),1)
+			# UNAME_P := arm
+			# UNAME_M := arm64
+			warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789)
+		endif
+	endif
+endif
+
+ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
+BUILD_TARGETS += metal
+endif
+
 default: $(BUILD_TARGETS)

 test:
@@ -38,18 +71,6 @@ gcovr-report: coverage ## Generate gcovr report
 	mkdir -p gcovr-report
 	gcovr --root . --html --html-details --output gcovr-report/coverage.html

-ifndef UNAME_S
-UNAME_S := $(shell uname -s)
-endif
-
-ifndef UNAME_P
-UNAME_P := $(shell uname -p)
-endif
-
-ifndef UNAME_M
-UNAME_M := $(shell uname -m)
-endif
-
 ifdef RISCV_CROSS_COMPILE
 CC	:= riscv64-unknown-linux-gnu-gcc
 CXX	:= riscv64-unknown-linux-gnu-g++
@@ -58,19 +79,6 @@ endif
 CCV := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)

-# Mac OS + Arm can report x86_64
-# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
-ifeq ($(UNAME_S),Darwin)
-	ifneq ($(UNAME_P),arm)
-		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
-		ifeq ($(SYSCTL_M),1)
-			# UNAME_P := arm
-			# UNAME_M := arm64
-			warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789)
-		endif
-	endif
-endif
-
 #
 # Compile flags
 #
@@ -101,12 +109,11 @@ endif


 ifdef LLAMA_CODE_COVERAGE
-	CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
+	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
 endif

 ifdef LLAMA_DISABLE_LOGS
-	CFLAGS   += -DLOG_DISABLE_LOGS
-	CXXFLAGS += -DLOG_DISABLE_LOGS
+	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
 endif # LLAMA_DISABLE_LOGS

 # warnings
@@ -116,7 +123,7 @@ MK_CXXFLAGS  += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-m

 ifeq '' '$(findstring clang++,$(CXX))'
 	# g++ only
-	CXXFLAGS += -Wno-format-truncation
+	MK_CXXFLAGS += -Wno-format-truncation
 endif

 # OS specific
@@ -180,8 +187,8 @@ endif
 # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
 # https://github.com/ggerganov/llama.cpp/issues/2922
 ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
-	CFLAGS   += -Xassembler -muse-unaligned-vector-move
-	CXXFLAGS += -Xassembler -muse-unaligned-vector-move
+	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
+	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
 endif

 ifneq ($(filter aarch64%,$(UNAME_M)),)
@@ -218,8 +225,8 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 endif

 else
-	CFLAGS += -march=rv64gcv -mabi=lp64d
-	CXXFLAGS +=  -march=rv64gcv -mabi=lp64d
+	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
+	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif

 ifndef LLAMA_NO_K_QUANTS
@@ -231,8 +238,8 @@ endif
 endif

 ifndef LLAMA_NO_ACCELERATE
-	# Mac M1 - include Accelerate framework.
-	# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
+	# Mac OS - include Accelerate framework.
+	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
 		MK_CPPFLAGS += -DGGML_USE_ACCELERATE
 		MK_LDFLAGS  += -framework Accelerate
@@ -350,9 +357,12 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 endif # LLAMA_HIPBLAS

 ifdef LLAMA_METAL
-	MK_CPPFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG
+	MK_CPPFLAGS += -DGGML_USE_METAL
 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
 	OBJS		+= ggml-metal.o
+ifdef LLAMA_METAL_NDEBUG
+	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
+endif
 endif # LLAMA_METAL

 ifdef LLAMA_METAL
@@ -477,9 +487,8 @@ baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS)
 beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
-BUILD_TARGETS += metal
-endif
+speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 ifdef LLAMA_METAL
 metal: examples/metal/metal.cpp ggml.o $(OBJS)
--- a/README.md
+++ b/README.md
@@ -280,29 +280,11 @@ In order to build llama.cpp you have three different options.

 ### Metal Build

-Using Metal allows the computation to be executed on the GPU for Apple devices:
+On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
+To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or the `LLAMA_METAL=OFF` cmake option.

- Using `make`:
-
-  ```bash
-  LLAMA_METAL=1 make
-  ```
-
- Using `CMake`:
-
-    ```bash
-    mkdir build-metal
-    cd build-metal
-    cmake -DLLAMA_METAL=ON ..
-    cmake --build . --config Release
-    ```
-
-When built with Metal support, you can enable GPU inference with the `--gpu-layers|-ngl` command-line argument.
-Any value larger than 0 will offload the computation to the GPU. For example:
-
-```bash
-./main -m ./models/7B/ggml-model-q4_0.gguf -n 128 -ngl 1
-```
+When built with Metal support, you can explicitly disable GPU inference with the `--gpu-layers|-ngl 0` command-line
+argument.

 ### MPI Build

--- a/common/common.cpp
+++ b/common/common.cpp
@@ -584,109 +584,109 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }

 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stdout, "  -i, --interactive     run in interactive mode\n");
-    fprintf(stdout, "  --interactive-first   run in interactive mode and wait for input right away\n");
-    fprintf(stdout, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    fprintf(stdout, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
-    fprintf(stdout, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stdout, "                        halt generation at PROMPT, return control in interactive mode\n");
-    fprintf(stdout, "                        (can be specified more than once for multiple prompts).\n");
-    fprintf(stdout, "  --color               colorise output to distinguish prompt and user input from generations\n");
-    fprintf(stdout, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stdout, "  -p PROMPT, --prompt PROMPT\n");
-    fprintf(stdout, "                        prompt to start generation with (default: empty)\n");
-    fprintf(stdout, "  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stdout, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
-    fprintf(stdout, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
-    fprintf(stdout, "                        not supported with --interactive or other interactive options\n");
-    fprintf(stdout, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
-    fprintf(stdout, "  --random-prompt       start with a randomized prompt.\n");
-    fprintf(stdout, "  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
-    fprintf(stdout, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
-    fprintf(stdout, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
-    fprintf(stdout, "  -f FNAME, --file FNAME\n");
-    fprintf(stdout, "                        prompt file to start generation.\n");
-    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
-    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stdout, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
-    fprintf(stdout, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
-    fprintf(stdout, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
-    fprintf(stdout, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
-    fprintf(stdout, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
-    fprintf(stdout, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    fprintf(stdout, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
-    fprintf(stdout, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
-    fprintf(stdout, "  --mirostat N          use Mirostat sampling.\n");
-    fprintf(stdout, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    fprintf(stdout, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
-    fprintf(stdout, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
-    fprintf(stdout, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
-    fprintf(stdout, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
-    fprintf(stdout, "                        modifies the likelihood of token appearing in the completion,\n");
-    fprintf(stdout, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
-    fprintf(stdout, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
-    fprintf(stdout, "  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
-    fprintf(stdout, "  --grammar-file FNAME  file to read grammar from\n");
-    fprintf(stdout, "  --cfg-negative-prompt PROMPT\n");
-    fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
-    fprintf(stdout, "  --cfg-negative-prompt-file FNAME\n");
-    fprintf(stdout, "                        negative prompt file to use for guidance. (default: empty)\n");
-    fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
-    fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
-    fprintf(stdout, "  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
-    fprintf(stdout, "  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
-    fprintf(stdout, "  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
-    fprintf(stdout, "  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
-    fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --interactive-first   run in interactive mode and wait for input right away\n");
+    printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
+    printf("  -r PROMPT, --reverse-prompt PROMPT\n");
+    printf("                        halt generation at PROMPT, return control in interactive mode\n");
+    printf("                        (can be specified more than once for multiple prompts).\n");
+    printf("  --color               colorise output to distinguish prompt and user input from generations\n");
+    printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -p PROMPT, --prompt PROMPT\n");
+    printf("                        prompt to start generation with (default: empty)\n");
+    printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    printf("  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
+    printf("  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
+    printf("                        not supported with --interactive or other interactive options\n");
+    printf("  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
+    printf("  --random-prompt       start with a randomized prompt.\n");
+    printf("  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
+    printf("  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    printf("  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
+    printf("  -f FNAME, --file FNAME\n");
+    printf("                        prompt file to start generation.\n");
+    printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    printf("  --mirostat N          use Mirostat sampling.\n");
+    printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
+    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
+    printf("                        modifies the likelihood of token appearing in the completion,\n");
+    printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
+    printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
+    printf("  --grammar-file FNAME  file to read grammar from\n");
+    printf("  --cfg-negative-prompt PROMPT\n");
+    printf("                        negative prompt to use for guidance. (default: empty)\n");
+    printf("  --cfg-negative-prompt-file FNAME\n");
+    printf("                        negative prompt file to use for guidance. (default: empty)\n");
+    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
+    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
+    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
+    printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
+    printf("  --no-penalize-nl      do not penalize newline token\n");
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    printf("  --perplexity          compute perplexity over each ctx window of the prompt\n");
+    printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
+    printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
+    printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
+    printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
    if (llama_mlock_supported()) {
-        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
    }
    if (llama_mmap_supported()) {
-        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
    }
-    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
-    fprintf(stdout, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
-    fprintf(stdout, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    printf("  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
 #ifdef GGML_USE_CUBLAS
-    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
-    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif // GGML_USE_CUBLAS
 #endif
-    fprintf(stdout, "  --mtest               compute maximum memory usage\n");
-    fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
-    fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
+    printf("  --mtest               compute maximum memory usage\n");
+    printf("  --export              export the computation graph to 'llama.ggml'\n");
+    printf("  --verbose-prompt      print prompt before generation\n");
    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
-    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "  -md FNAME, --model-draft FNAME\n");
-    fprintf(stdout, "                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "  -ld LOGDIR, --logdir LOGDIR\n");
-    fprintf(stdout, "                        path under which to save YAML logs (no logging if unset)\n");
-    fprintf(stdout, "\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -md FNAME, --model-draft FNAME\n");
+    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
+    printf("  -ld LOGDIR, --logdir LOGDIR\n");
+    printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("\n");
 }

 std::string gpt_random_prompt(std::mt19937 & rng) {
@@ -717,7 +717,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param

    lparams.n_ctx           = params.n_ctx;
    lparams.n_batch         = params.n_batch;
-    lparams.n_gpu_layers    = params.n_gpu_layers;
+    if (params.n_gpu_layers != -1) {
+        lparams.n_gpu_layers = params.n_gpu_layers;
+    }
    lparams.main_gpu        = params.main_gpu;
    lparams.tensor_split    = params.tensor_split;
    lparams.low_vram        = params.low_vram;
@@ -770,8 +772,8 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
    {
        LOG("warming up the model with an empty run\n");

-        const std::vector<llama_token> tmp = { llama_token_bos(lctx), };
-        llama_eval(lctx, tmp.data(), tmp.size(), 0, params.n_threads);
+        const std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, params.n_threads);
        llama_reset_timings(lctx);
    }

@@ -1212,7 +1214,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
    fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false");
    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
-    fprintf(stream, "n_gpu_layers: %d # default: 0\n", params.n_gpu_layers);
+    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs);
    fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
--- a/common/common.h
+++ b/common/common.h
@@ -34,7 +34,7 @@ struct gpt_params {
    int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
    int32_t n_draft                         = 16;   // number of tokens to draft during speculative decoding
    int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
-    int32_t n_gpu_layers                    = 0;    // number of layers to store in VRAM
+    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
    int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
    float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
    int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
--- a/common/log.h
+++ b/common/log.h
@@ -513,16 +513,16 @@ inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string &

 inline void log_print_usage()
 {
-    fprintf(stdout, "log options:\n");
+    printf("log options:\n");
    /* format
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");*/
+    printf("  -h, --help            show this help message and exit\n");*/
    /* spacing
-    fprintf(stdout, "__-param----------------Description\n");*/
-    fprintf(stdout, "  --log-test            Run simple logging test\n");
-    fprintf(stdout, "  --log-disable         Disable trace logs\n");
-    fprintf(stdout, "  --log-enable          Enable trace logs\n");
-    fprintf(stdout, "  --log-file            Specify a log filename (without extension)\n");
-    fprintf(stdout, "                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+    printf("__-param----------------Description\n");*/
+    printf("  --log-test            Run simple logging test\n");
+    printf("  --log-disable         Disable trace logs\n");
+    printf("  --log-enable          Enable trace logs\n");
+    printf("  --log-file            Specify a log filename (without extension)\n");
+    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
 }

 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
--- a/convert-falcon-hf-to-gguf.py
+++ b/convert-falcon-hf-to-gguf.py
@@ -55,10 +55,10 @@ def count_model_parts(dir_model: Path) -> int:

 def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
-    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.bin)")
-    parser.add_argument("ftype",     type=int, choices=[0, 1],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    parser.add_argument("--outfile",    type=Path,           help="path to write to; default: based on input")
+    parser.add_argument("model",        type=Path,           help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
    return parser.parse_args()

 args = parse_args()
--- a/convert-llama-ggmlv3-to-gguf.py
+++ b/convert-llama-ggmlv3-to-gguf.py
@@ -5,6 +5,7 @@ import argparse
 import math
 import struct
 import sys
+from enum import IntEnum
 from pathlib import Path

 import numpy as np
@@ -34,10 +35,35 @@ GGML_QUANT_SIZES = {
    gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8),
 }

+class GGMLFormat(IntEnum):
+    GGML = 0
+    GGMF = 1
+    GGJT = 2
+
+class GGMLFType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1
+    MOSTLY_Q4_0          = 2
+    MOSTLY_Q4_1          = 3
+    MOSTLY_Q4_1_SOME_F16 = 4
+    MOSTLY_Q8_0          = 7
+    MOSTLY_Q5_0          = 8
+    MOSTLY_Q5_1          = 9
+    MOSTLY_Q2_K          = 10
+    MOSTLY_Q3_K_S        = 11
+    MOSTLY_Q3_K_M        = 12
+    MOSTLY_Q3_K_L        = 13
+    MOSTLY_Q4_K_S        = 14
+    MOSTLY_Q4_K_M        = 15
+    MOSTLY_Q5_K_S        = 16
+    MOSTLY_Q5_K_M        = 17
+    MOSTLY_Q6_K          = 18
+
 class Hyperparameters:
    def __init__(self):
-        self.n_vocab = self.n_embd = self.n_mult = self.n_head = self.n_layer = self.n_rot = self.ftype = 0
-        self.n_ff = 0
+        self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
+        self.n_layer = self.n_rot = self.n_ff = 0
+        self.ftype = GGMLFType.ALL_F32

    def set_n_ff(self, model):
        ff_tensor_idx = model.tensor_map.get(b'layers.0.feed_forward.w1.weight')
@@ -53,16 +79,21 @@ class Hyperparameters:
            self.n_head,
            self.n_layer,
            self.n_rot,
-            self.ftype,
+            ftype,
        ) = struct.unpack('<7I', data[offset:offset + (4 * 7)])
+        try:
+            self.ftype = GGMLFType(ftype)
+        except ValueError:
+            raise ValueError(f'Invalid ftype {ftype}')
        return 4 * 7

    def __str__(self):
-        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype}>'
+        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'

 class Vocab:
-    def __init__(self):
+    def __init__(self, load_scores = True):
        self.items = []
+        self.load_scores = load_scores

    def load(self, data, offset, n_vocab):
        orig_offset = offset
@@ -70,20 +101,24 @@ class Vocab:
            itemlen = struct.unpack('<I', data[offset:offset + 4])[0]
            assert itemlen < 4096, 'Absurd vocab item length'
            offset += 4
-            vocab = bytes(data[offset:offset + itemlen])
+            item_text = bytes(data[offset:offset + itemlen])
            offset += itemlen
-            score = struct.unpack('<f', data[offset:offset + 4])[0]
-            offset += 4
-            self.items.append((vocab, score))
+            if self.load_scores:
+                item_score = struct.unpack('<f', data[offset:offset + 4])[0]
+                offset += 4
+            else:
+                item_score = 0.0
+            self.items.append((item_text, item_score))
        return offset - orig_offset

 class Tensor:
-    def __init__(self):
+    def __init__(self, use_padding = True):
        self.name = None
        self.dims: tuple[int, ...] = ()
        self.dtype = None
        self.start_offset = 0
        self.len_bytes = np.int64(0)
+        self.use_padding = use_padding

    def load(self, data, offset):
        orig_offset = offset
@@ -99,7 +134,7 @@ class Tensor:
        offset += 4 * n_dims
        self.name = bytes(data[offset:offset + name_len])
        offset += name_len
-        pad = ((offset + 31) & ~31) - offset
+        pad = ((offset + 31) & ~31) - offset if self.use_padding else 0
        offset += pad
        n_elems = np.prod(self.dims)
        n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
@@ -109,7 +144,7 @@ class Tensor:
        # print(n_dims, name_len, dtype, self.dims, self.name, pad)
        return offset - orig_offset

-class GGMLV3Model:
+class GGMLModel:
    def __init__(self):
        self.hyperparameters = None
        self.vocab = None
@@ -117,20 +152,52 @@ class GGMLV3Model:
        self.tensors = []

    def validate_header(self, data, offset):
-        if bytes(data[offset:offset + 4]) != b'tjgg' or struct.unpack('<I', data[offset + 4:offset + 8])[0] != 3:
-            raise ValueError('Only GGJTv3 supported')
-        return 8
+        magic = bytes(data[offset:offset + 4])
+        if magic == b'GGUF':
+            raise ValueError('File is already in GGUF format.')
+        if magic == b'lmgg':
+            self.file_format = GGMLFormat.GGML
+            self.format_version = 1
+            return 4
+        version = struct.unpack('<I', data[offset + 4:offset + 8])[0]
+        if magic == b'fmgg':
+            if version != 1:
+                raise ValueError(f'Cannot handle unexpected GGMF file version {version}')
+            self.file_format = GGMLFormat.GGMF
+            self.format_version = version
+            return 8
+        if magic == b'tjgg':
+            if version < 1 or version > 3:
+                raise ValueError(f'Cannot handle unexpected GGJT file version {version}')
+            self.file_format = GGMLFormat.GGJT
+            self.format_version = version
+            return 8
+        raise ValueError(f"Unexpected file magic {magic!r}! This doesn't look like a GGML format file.")
+
+    def validate_conversion(self, ftype):
+        err = ''
+        if (self.file_format < GGMLFormat.GGJT or self.format_version < 2):
+            if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
+                err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
+        elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
+            if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                          GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+                err = 'Q4 and Q8 quantizations changed in GGJTv3.'
+        if len(err) > 0:
+            raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')

    def load(self, data, offset):
        offset += self.validate_header(data, offset)
        hp = Hyperparameters()
        offset += hp.load(data, offset)
-        vocab = Vocab()
+        print(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
+        self.validate_conversion(hp.ftype)
+        vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
        offset += vocab.load(data, offset, hp.n_vocab)
        tensors: list[Tensor] = []
        tensor_map = {}
        while offset < len(data):
-            tensor = Tensor()
+            tensor = Tensor(use_padding = self.file_format > GGMLFormat.GGMF)
            offset += tensor.load(data, offset)
            tensor_map[tensor.name] = len(tensors)
            tensors.append(tensor)
@@ -168,7 +235,10 @@ class GGMLToGGUF:

    def save(self):
        print('* Preparing to save GGUF file')
-        gguf_writer = gguf.GGUFWriter(self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+        gguf_writer = gguf.GGUFWriter(
+            self.cfg.output,
+            gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
+            use_temp_file = False )
        self.add_params(gguf_writer)
        self.add_vocab(gguf_writer)
        if self.special_vocab is not None:
@@ -185,7 +255,10 @@ class GGMLToGGUF:
    def add_params(self, gguf_writer):
        hp = self.model.hyperparameters
        cfg = self.cfg
-        desc = cfg.desc if cfg.desc is not None else 'converted from legacy GGJTv3 format'
+        if cfg.desc is not None:
+            desc = cfg.desc
+        else:
+            desc = f'converted from legacy {self.model.file_format.name}v{self.model.format_version} {hp.ftype.name} format'
        try:
            # Filenames aren't necessarily valid UTF8.
            name = cfg.name if cfg.name is not None else cfg.input.name
@@ -195,6 +268,7 @@ class GGMLToGGUF:
        if name is not None:
            gguf_writer.add_name(name)
        gguf_writer.add_description(desc)
+        gguf_writer.add_file_type(int(hp.ftype))
        if self.params_override is not None:
            po = self.params_override
            assert po.n_embd == hp.n_embd, 'Model hyperparams mismatch'
@@ -231,7 +305,8 @@ class GGMLToGGUF:
                tokens.append(vbytes)
                scores.append(score)
                toktypes.append(ttype)
-            assert len(tokens) == hp.n_vocab, f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
+            assert len(tokens) == hp.n_vocab, \
+                f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
            gguf_writer.add_token_list(tokens)
            gguf_writer.add_token_scores(scores)
            if len(toktypes) > 0:
@@ -283,7 +358,11 @@ class GGMLToGGUF:
                tempdims[1] = tempdims[0]
                tempdims[0] = temp
            # print(f'+ {tensor.name} | {mapped_name} {tensor.dims} :: {tempdims}')
-            gguf_writer.add_tensor(mapped_name, data[tensor.start_offset:tensor.start_offset + tensor.len_bytes], raw_shape = tempdims, raw_dtype = tensor.dtype)
+            gguf_writer.add_tensor(
+                mapped_name,
+                data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
+                raw_shape = tempdims,
+                raw_dtype = tensor.dtype )

 def handle_metadata(cfg, hp):
    import convert
@@ -305,32 +384,46 @@ def handle_metadata(cfg, hp):
        params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
    else:
        raise ValueError('Unable to load metadata')
-    vocab = convert.load_vocab(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, cfg.vocabtype)
+    vocab = convert.load_vocab(
+        cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
+        cfg.vocabtype )
    # FIXME: Respect cfg.vocab_dir?
    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
    convert.check_vocab_size(params, vocab)
    return (params, vocab, svocab)

 def handle_args():
-    parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF')
-    parser.add_argument('--input', '-i', type = Path, required = True, help = 'Input GGMLv3 filename')
-    parser.add_argument('--output', '-o', type = Path, required = True, help ='Output GGUF filename')
-    parser.add_argument('--name', help = 'Set model name')
-    parser.add_argument('--desc', help = 'Set model description')
-    parser.add_argument('--gqa', type = int, default = 1, help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
-    parser.add_argument('--eps', default = '5.0e-06', help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
-    parser.add_argument('--context-length', '-c', type=int, default = 2048, help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
-    parser.add_argument('--model-metadata-dir', '-m', type = Path, help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
-    parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
-    parser.add_argument("--vocabtype", choices=["spm", "bpe"], help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)", default="spm")
+    parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
+    parser.add_argument('--input', '-i', type = Path, required = True,
+        help = 'Input GGMLv3 filename')
+    parser.add_argument('--output', '-o', type = Path, required = True,
+        help ='Output GGUF filename')
+    parser.add_argument('--name',
+        help = 'Set model name')
+    parser.add_argument('--desc',
+        help = 'Set model description')
+    parser.add_argument('--gqa', type = int, default = 1,
+        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+    parser.add_argument('--eps', default = '5.0e-06',
+        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+    parser.add_argument('--context-length', '-c', type=int, default = 2048,
+        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+    parser.add_argument('--model-metadata-dir', '-m', type = Path,
+        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+    parser.add_argument("--vocab-dir", type=Path,
+        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
+    parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
+        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
    return parser.parse_args()

 def main():
    cfg = handle_args()
    print(f'* Using config: {cfg}')
    print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n')
+    if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
+        print('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
    data = np.memmap(cfg.input, mode = 'r')
-    model = GGMLV3Model()
+    model = GGMLModel()
    print('* Scanning GGML input file')
    offset = model.load(data, 0)
    print(f'* GGML model hyperparameters: {model.hyperparameters}')
@@ -345,7 +438,12 @@ def main():
        print(f'* Special vocab: {special_vocab}')
    else:
        print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
-    converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override, special_vocab = special_vocab)
+        if model.file_format == GGMLFormat.GGML:
+            print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
+    converter = GGMLToGGUF(model, data, cfg,
+        params_override = params_override,
+        vocab_override = vocab_override,
+        special_vocab = special_vocab )
    converter.save()
    print(f'* Successful completion. Output saved to: {cfg.output}')

--- a/convert.py
+++ b/convert.py
@@ -673,7 +673,7 @@ class LazyUnpickler(pickle.Unpickler):
        assert isinstance(pid[1], LazyStorageKind)
        data_type = pid[1].data_type
        filename_stem = pid[2]
-        filename = self.data_base_path + '/' + filename_stem
+        filename = f'{self.data_base_path}/{filename_stem}'
        info = self.zip_file.getinfo(filename)

        def load(offset: int, elm_count: int) -> NDArray:
@@ -689,7 +689,6 @@ class LazyUnpickler(pickle.Unpickler):

    @staticmethod
    def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
-                               # pyright: ignore[reportSelfClsParameterName]
                               requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
        assert isinstance(storage, LazyStorage)

--- a/examples/gguf/gguf.cpp
+++ b/examples/gguf/gguf.cpp
@@ -76,7 +76,7 @@ bool gguf_ex_write(const std::string & fname) {

    gguf_write_to_file(ctx, fname.c_str(), false);

-    fprintf(stdout, "%s: wrote file '%s;\n", __func__, fname.c_str());
+    printf("%s: wrote file '%s;\n", __func__, fname.c_str());

    ggml_free(ctx_data);
    gguf_free(ctx);
@@ -93,20 +93,20 @@ bool gguf_ex_read_0(const std::string & fname) {

    struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);

-    fprintf(stdout, "%s: version:      %d\n", __func__, gguf_get_version(ctx));
-    fprintf(stdout, "%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
-    fprintf(stdout, "%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
+    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
+    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
+    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));

    // kv
    {
        const int n_kv = gguf_get_n_kv(ctx);

-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);

        for (int i = 0; i < n_kv; ++i) {
            const char * key = gguf_get_key(ctx, i);

-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
        }
    }

@@ -116,10 +116,10 @@ bool gguf_ex_read_0(const std::string & fname) {

        const int keyidx = gguf_find_key(ctx, findkey);
        if (keyidx == -1) {
-            fprintf(stdout, "%s: find key: %s not found.\n", __func__, findkey);
+            printf("%s: find key: %s not found.\n", __func__, findkey);
        } else {
            const char * key_value = gguf_get_val_str(ctx, keyidx);
-            fprintf(stdout, "%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);
+            printf("%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);
        }
    }

@@ -127,13 +127,13 @@ bool gguf_ex_read_0(const std::string & fname) {
    {
        const int n_tensors = gguf_get_n_tensors(ctx);

-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);

        for (int i = 0; i < n_tensors; ++i) {
            const char * name   = gguf_get_tensor_name  (ctx, i);
            const size_t offset = gguf_get_tensor_offset(ctx, i);

-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
        }
    }

@@ -153,20 +153,20 @@ bool gguf_ex_read_1(const std::string & fname) {

    struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);

-    fprintf(stdout, "%s: version:      %d\n", __func__, gguf_get_version(ctx));
-    fprintf(stdout, "%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
-    fprintf(stdout, "%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
+    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
+    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
+    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));

    // kv
    {
        const int n_kv = gguf_get_n_kv(ctx);

-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);

        for (int i = 0; i < n_kv; ++i) {
            const char * key = gguf_get_key(ctx, i);

-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
        }
    }

@@ -174,13 +174,13 @@ bool gguf_ex_read_1(const std::string & fname) {
    {
        const int n_tensors = gguf_get_n_tensors(ctx);

-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);

        for (int i = 0; i < n_tensors; ++i) {
            const char * name   = gguf_get_tensor_name  (ctx, i);
            const size_t offset = gguf_get_tensor_offset(ctx, i);

-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
        }
    }

@@ -189,13 +189,13 @@ bool gguf_ex_read_1(const std::string & fname) {
        const int n_tensors = gguf_get_n_tensors(ctx);

        for (int i = 0; i < n_tensors; ++i) {
-            fprintf(stdout, "%s: reading tensor %d data\n", __func__, i);
+            printf("%s: reading tensor %d data\n", __func__, i);

            const char * name = gguf_get_tensor_name(ctx, i);

            struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);

-            fprintf(stdout, "%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);

            // print first 10 elements
            const float * data = (const float *) cur->data;
@@ -219,7 +219,7 @@ bool gguf_ex_read_1(const std::string & fname) {
        }
    }

-    fprintf(stdout, "%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));
+    printf("%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));

    ggml_free(ctx_data);
    gguf_free(ctx);
@@ -229,7 +229,7 @@ bool gguf_ex_read_1(const std::string & fname) {

 int main(int argc, char ** argv) {
    if (argc < 3) {
-        fprintf(stdout, "usage: %s data.gguf r|w\n", argv[0]);
+        printf("usage: %s data.gguf r|w\n", argv[0]);
        return -1;
    }

--- a/examples/gptneox-wip/falcon-main.cpp
+++ b/examples/gptneox-wip/falcon-main.cpp
@@ -305,9 +305,9 @@ struct ggml_tensor * get_tensor_ex( struct ggml_context * ctx, std::string name)

    struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
    if( cur == NULL ) {
-        fprintf(stdout, "%s: tensor '%s' not found!\n", __func__, name.c_str());
+        printf("%s: tensor '%s' not found!\n", __func__, name.c_str());
    } else {
-//        fprintf(stdout, "%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
+//        printf("%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
    }

    return cur;
@@ -333,21 +333,21 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        return false;
    }

-    fprintf(stdout, "%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    fprintf(stdout, "%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    fprintf(stdout, "%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
+    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
+    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
+    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));

    // print all kv
    #if 0
    {
        const int n_kv = gguf_get_n_kv(ggufctx);

-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);

        for (int i = 0; i < n_kv; ++i) {
            const char * key = gguf_get_key(ggufctx, i);

-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
        }
    }
    #endif
@@ -357,21 +357,21 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        int keyidx;

        keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { fprintf(stdout, "%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { fprintf(stdout, "%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { fprintf(stdout, "%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { fprintf(stdout, "%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { fprintf(stdout, "%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { fprintf(stdout, "%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { fprintf(stdout, "%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
-        if (keyidx != -1) { fprintf(stdout, "%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
    }

    // check required metadata
@@ -382,11 +382,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        keyidx = gguf_find_key(ggufctx, "general.architecture");
        if (keyidx != -1) {
            if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "falcon") != 0) {
-                fprintf(stdout, "%s: model architecture not supported!\n", __func__);
+                printf("%s: model architecture not supported!\n", __func__);
                return false;
            }
        } else {
-            fprintf(stdout, "%s: gguf model architecture not found!\n", __func__);
+            printf("%s: gguf model architecture not found!\n", __func__);
            return false;
        }

@@ -394,11 +394,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        keyidx = gguf_find_key(ggufctx, "falcon.tensor_data_layout");
        if (keyidx != -1) {
            if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "jploski") != 0) {
-                fprintf(stdout, "%s: model tensor data layout not supported!\n", __func__);
+                printf("%s: model tensor data layout not supported!\n", __func__);
                return false;
            }
        } else {
-            fprintf(stdout, "%s: gguf model tensor data layout not found!\n", __func__);
+            printf("%s: gguf model tensor data layout not found!\n", __func__);
            return false;
        }

@@ -455,11 +455,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_

        if (keyidx != -1) {
            if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
-                fprintf(stdout, "%s: tokenizer model not supported!\n", __func__);
+                printf("%s: tokenizer model not supported!\n", __func__);
                return false;
            }
        } else {
-            fprintf(stdout, "%s: tokenizer model not found!\n", __func__);
+            printf("%s: tokenizer model not found!\n", __func__);
            return false;
        }

@@ -467,22 +467,22 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");

        if (tokens_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            printf("%s: gpt2 tokenizer vocab not found!\n", __func__);
            return false;
        }

        int merges_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.merges");

        if (merges_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer merges not found!\n", __func__);
+            printf("%s: gpt2 tokenizer merges not found!\n", __func__);
            return false;
        }

        hparams.n_vocab = gguf_get_arr_n(ggufctx,tokens_keyidx);
        hparams.n_merges = gguf_get_arr_n(ggufctx,merges_keyidx);

-        fprintf(stdout, "%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
-        fprintf(stdout, "%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
+        printf("%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
+        printf("%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);

        for (size_t i = 0; i < hparams.n_vocab; i++) {
            std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
@@ -523,12 +523,12 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
        keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
        keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) {   vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }

-        if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
-        if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
-        if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
-        if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
-        if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
-        if( vocab.linefeed_id    != -1 ) { fprintf(stdout, "%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
+        if( vocab.special_bos_id != -1 ) { printf("%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
+        if( vocab.special_eos_id != -1 ) { printf("%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
+        if( vocab.special_unk_id != -1 ) { printf("%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
+        if( vocab.special_sep_id != -1 ) { printf("%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
+        if( vocab.special_pad_id != -1 ) { printf("%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
+        if( vocab.linefeed_id    != -1 ) { printf("%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }

    }

@@ -543,13 +543,13 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
    {
        const int n_tensors = gguf_get_n_tensors(ggufctx);

-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);

        for (int i = 0; i < n_tensors; ++i) {
            const char * name   = gguf_get_tensor_name  (ggufctx, i);
            const size_t offset = gguf_get_tensor_offset(ggufctx, i);

-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
        }
    }
    #endif
--- a/examples/gptneox-wip/gptneox-main.cpp
+++ b/examples/gptneox-wip/gptneox-main.cpp
@@ -318,9 +318,9 @@ struct ggml_tensor * get_tensor_ex( struct ggml_context * ctx, std::string name)

    struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
    if( cur == NULL ) {
-        fprintf(stdout, "%s: tensor '%s' not found!\n", __func__, name.c_str());
+        printf("%s: tensor '%s' not found!\n", __func__, name.c_str());
    } else {
-//        fprintf(stdout, "%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
+//        printf("%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
    }

    return cur;
@@ -346,21 +346,21 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
        return false;
    }

-    fprintf(stdout, "%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    fprintf(stdout, "%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    fprintf(stdout, "%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
+    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
+    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
+    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));

    // print all kv
    #if 0
    {
        const int n_kv = gguf_get_n_kv(ggufctx);

-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);

        for (int i = 0; i < n_kv; ++i) {
            const char * key = gguf_get_key(ggufctx, i);

-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
        }
    }
    #endif
@@ -370,21 +370,21 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
        int keyidx;

        keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { fprintf(stdout, "%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { fprintf(stdout, "%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { fprintf(stdout, "%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { fprintf(stdout, "%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { fprintf(stdout, "%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { fprintf(stdout, "%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { fprintf(stdout, "%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
-        if (keyidx != -1) { fprintf(stdout, "%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
    }

    // check required metadata
@@ -395,11 +395,11 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
        keyidx = gguf_find_key(ggufctx, "general.architecture");
        if (keyidx != -1) {
            if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gptneox") != 0) {
-                fprintf(stdout, "%s: model architecture not supported!\n", __func__);
+                printf("%s: model architecture not supported!\n", __func__);
                return false;
            }
        } else {
-            fprintf(stdout, "%s: gguf model architecture not found!\n", __func__);
+            printf("%s: gguf model architecture not found!\n", __func__);
            return false;
        }

@@ -456,11 +456,11 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2

        if (keyidx != -1) {
            if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
-                fprintf(stdout, "%s: tokenizer model not supported!\n", __func__);
+                printf("%s: tokenizer model not supported!\n", __func__);
                return false;
            }
        } else {
-            fprintf(stdout, "%s: tokenizer model not found!\n", __func__);
+            printf("%s: tokenizer model not found!\n", __func__);
            return false;
        }

@@ -468,22 +468,22 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
        int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");

        if (tokens_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            printf("%s: gpt2 tokenizer vocab not found!\n", __func__);
            return false;
        }

        int merges_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.merges");

        if (merges_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer merges not found!\n", __func__);
+            printf("%s: gpt2 tokenizer merges not found!\n", __func__);
            return false;
        }

        hparams.n_vocab = gguf_get_arr_n(ggufctx,tokens_keyidx);
        hparams.n_merges = gguf_get_arr_n(ggufctx,merges_keyidx);

-        fprintf(stdout, "%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
-        fprintf(stdout, "%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
+        printf("%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
+        printf("%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);

        for (size_t i = 0; i < hparams.n_vocab; i++) {
            std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
@@ -524,12 +524,12 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
        keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
        keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) {   vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }

-        if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
-        if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
-        if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
-        if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
-        if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
-        if( vocab.linefeed_id    != -1 ) { fprintf(stdout, "%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
+        if( vocab.special_bos_id != -1 ) { printf("%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
+        if( vocab.special_eos_id != -1 ) { printf("%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
+        if( vocab.special_unk_id != -1 ) { printf("%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
+        if( vocab.special_sep_id != -1 ) { printf("%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
+        if( vocab.special_pad_id != -1 ) { printf("%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
+        if( vocab.linefeed_id    != -1 ) { printf("%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
    }


@@ -543,13 +543,13 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
    {
        const int n_tensors = gguf_get_n_tensors(ggufctx);

-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);

        for (int i = 0; i < n_tensors; ++i) {
            const char * name   = gguf_get_tensor_name  (ggufctx, i);
            const size_t offset = gguf_get_tensor_offset(ggufctx, i);

-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
        }
    }
    #endif
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -165,26 +165,26 @@ static const cmd_params cmd_params_defaults = {
 };

 static void print_usage(int /* argc */, char ** argv) {
-    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help\n");
-    fprintf(stdout, "  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    fprintf(stdout, "  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    fprintf(stdout, "  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    fprintf(stdout, "  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    fprintf(stdout, "  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
-    fprintf(stdout, "  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    fprintf(stdout, "  -ngl N, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    fprintf(stdout, "  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    fprintf(stdout, "  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
-    fprintf(stdout, "  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    fprintf(stdout, "  -ts, --tensor_split <ts0/ts1/..>               \n");
-    fprintf(stdout, "  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
-    fprintf(stdout, "  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
-    fprintf(stdout, "  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
-    fprintf(stdout, "\n");
-    fprintf(stdout, "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help\n");
+    printf("  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl N, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
+    printf("  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>               \n");
+    printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
+    printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("\n");
+    printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");

 }

--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -151,14 +151,6 @@ int main(int argc, char ** argv) {
        LOG_TEE("%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
    }

-    if (params.n_ctx > 2048) {
-        // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
-        LOG_TEE("%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
-    } else if (params.n_ctx < 8) {
-        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
-        params.n_ctx = 8;
-    }
-
    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);

    if (params.seed == LLAMA_DEFAULT_SEED) {
@@ -194,6 +186,13 @@ int main(int argc, char ** argv) {
        return 1;
    }

+    if (params.n_ctx > llama_n_ctx(ctx)) {
+        LOG_TEE("%s: warning: base model only supports context sizes no greater than %d tokens (%d specified)\n", __func__, llama_n_ctx(ctx), params.n_ctx);
+    } else if (params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+
    // print system information
    {
        LOG_TEE("\n");
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -368,7 +368,7 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
        // Example, we have a context window of 512, we will compute perplexity for each of the
        // last 256 tokens.  Then, we split the input up into context window size chunks to
        // process the entire prompt.
-        const int first = std::min(512, params.n_ctx/2);
+        const int first = params.n_ctx/2;
        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first,
                       workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
        count += params.n_ctx - first - 1;
@@ -668,11 +668,6 @@ int main(int argc, char ** argv) {
        params.n_ctx += params.ppl_stride/2;
    }

-    if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
-                "expect poor results\n", __func__, params.n_ctx);
-    }
-
    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);

    if (params.seed == LLAMA_DEFAULT_SEED) {
@@ -698,6 +693,11 @@ int main(int argc, char ** argv) {
        return 1;
    }

+    if (params.n_ctx > llama_n_ctx(ctx)) {
+        fprintf(stderr, "%s: warning: model might not support context sizes greater than %d tokens (%d specified);"
+                "expect poor results\n", __func__, llama_n_ctx(ctx), params.n_ctx);
+    }
+
    // print system information
    {
        fprintf(stderr, "\n");
--- a/examples/server/index.html.hpp
+++ b/examples/server/index.html.hpp
--- a/examples/server/public/index.html
+++ b/examples/server/public/index.html
@@ -145,7 +145,29 @@
      color: #888;
    }

+
+    @keyframes loading-bg-wipe {
+      0% {
+        background-position: 0%;
+      }
+      100% {
+        background-position: 100%;
+      }
+    }
+
+    .loading {
+      --loading-color-1: #eeeeee00;
+      --loading-color-2: #eeeeeeff;
+      background-size: 50% 100%;
+      background-image: linear-gradient(90deg, var(--loading-color-1), var(--loading-color-2), var(--loading-color-1));
+      animation: loading-bg-wipe 2s linear infinite;
+    }
+
    @media (prefers-color-scheme: dark) {
+      .loading {
+        --loading-color-1: #22222200;
+        --loading-color-2: #222222ff;
+      }
      .popover-content {
        background-color: black;
      }
@@ -321,7 +343,10 @@
    const llamaStats = signal(null)
    const controller = signal(null)

-    const generating = computed(() => controller.value == null )
+    // currently generating a completion?
+    const generating = computed(() => controller.value != null)
+
+    // has the user started a chat?
    const chatStarted = computed(() => session.value.transcript.length > 0)

    const transcriptUpdate = (transcript) => {
@@ -430,11 +455,19 @@
      return html`
        <form onsubmit=${submit}>
          <div>
-            <textarea type="text" rows=2 onkeypress=${enterSubmits} value="${message}" oninput=${(e) => message.value = e.target.value} placeholder="Say something..."/>
+            <textarea
+               className=${generating.value ? "loading" : null}
+               oninput=${(e) => message.value = e.target.value}
+               onkeypress=${enterSubmits}
+               placeholder="Say something..."
+               rows=2
+               type="text"
+               value="${message}"
+            />
          </div>
          <div class="right">
-            <button type="submit" disabled=${!generating.value} >Send</button>
-            <button onclick=${stop} disabled=${generating}>Stop</button>
+            <button type="submit" disabled=${generating.value}>Send</button>
+            <button onclick=${stop} disabled=${!generating.value}>Stop</button>
            <button onclick=${reset}>Reset</button>
          </div>
        </form>
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -118,7 +118,7 @@ static void server_log(const char *level, const char *function, int line,
    }

    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
-    fprintf(stdout, "%.*s\n", (int)str.size(), str.data());
+    printf("%.*s\n", (int)str.size(), str.data());
    fflush(stdout);
 }

@@ -694,50 +694,50 @@ struct llama_server_context
 static void server_print_usage(const char *argv0, const gpt_params &params,
                               const server_params &sparams)
 {
-    fprintf(stdout, "usage: %s [options]\n", argv0);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stdout, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("usage: %s [options]\n", argv0);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
    if (llama_mlock_supported())
    {
-        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
    }
    if (llama_mmap_supported())
    {
-        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
    }
-    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
-    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    printf("  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
-    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "  -a ALIAS, --alias ALIAS\n");
-    fprintf(stdout, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stdout, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
-    fprintf(stdout, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
-    fprintf(stdout, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
-    fprintf(stdout, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
-    fprintf(stdout, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
-    fprintf(stdout, "\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -a ALIAS, --alias ALIAS\n");
+    printf("                        set an alias for the model, will be added as `model` field in completion response\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+    printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
+    printf("\n");
 }

 static void server_params_parse(int argc, char **argv, server_params &sparams,
@@ -1595,7 +1595,7 @@ int main(int argc, char **argv)
    svr.set_base_dir(sparams.public_path);

    // to make it ctrl+clickable:
-    fprintf(stdout, "\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+    printf("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);

    LOG_INFO("HTTP server listening", {
                                          {"hostname", sparams.hostname},
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -6,6 +6,7 @@

 #include "common.h"
 #include "llama.h"
+#include "grammar-parser.h"

 #include <cmath>
 #include <cstdio>
@@ -109,16 +110,35 @@ int main(int argc, char ** argv) {
    // used to determine end of generation
    bool has_eos = false;

+    // grammar stuff
+    struct llama_grammar * grammar_dft = NULL;
+    struct llama_grammar * grammar_tgt = NULL;
+
+    grammar_parser::parse_state parsed_grammar;
+
+    // if requested - load the grammar, error checking is omitted for brevity
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
    const auto t_dec_start = ggml_time_us();

    while (true) {
        LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));

-        // sample from the drafted tokens if any
        int i_dft = 0;
        while (true) {
-            const llama_token id = llama_sample_token(ctx_tgt, NULL, NULL, params, last_tokens, candidates, i_dft);
+            // sample from the target model
+            const llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);

+            // remember which tokens were sampled - used for repetition penalties during sampling
            last_tokens.erase(last_tokens.begin());
            last_tokens.push_back(id);

@@ -134,8 +154,9 @@ int main(int argc, char ** argv) {

            ++n_predict;

+            // check if the draft matches the target
            if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
-                LOG("drafted token %d accepted\n", id);
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
                ++n_accept;
                ++n_past_tgt;
                ++n_past_dft;
@@ -145,6 +166,14 @@ int main(int argc, char ** argv) {
            }

            // the drafted token was rejected or we are out of drafted tokens
+
+            if (i_dft < (int) drafted.size()) {
+                LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
+                        i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
+            } else {
+                LOG("out of drafted tokens\n");
+            }
+
            llama_eval(ctx_dft, &id, 1, n_past_dft, params.n_threads);
            ++n_past_dft;

@@ -158,7 +187,16 @@ int main(int argc, char ** argv) {
            break;
        }

-        // sample n_draft tokens from the draft model picking the best token
+        if (grammar_tgt) {
+            if (grammar_dft) {
+                llama_grammar_free(grammar_dft);
+            }
+            grammar_dft = llama_grammar_copy(grammar_tgt);
+
+            LOG("copied target grammar to draft grammar\n");
+        }
+
+        // sample n_draft tokens from the draft model using greedy decoding
        int n_past_cur = n_past_dft;
        for (int i = 0; i < n_draft; ++i) {
            float * logits = llama_get_logits(ctx_dft);
@@ -170,25 +208,40 @@ int main(int argc, char ** argv) {

            llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };

+            if (grammar_dft != NULL) {
+                llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
+            }
+
            // computes softmax and sorts the candidates
            llama_sample_softmax(ctx_dft, &cur_p);

            for (int i = 0; i < 3; ++i) {
-                LOG(" - draft candidate %d: %d (%.3f)\n", i, cur_p.data[i].id, cur_p.data[i].p);
+                LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
            }

-            // too low probability, stop drafting
+            // TODO: better logic?
            if (cur_p.data[0].p < 2*cur_p.data[1].p) {
+                LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
                break;
            }

-            drafted.push_back(cur_p.data[0].id);
+            // drafted token
+            const llama_token id = cur_p.data[0].id;
+
+            drafted.push_back(id);
            ++n_drafted;

-            if (i < n_draft - 1) {
-                // evaluate the drafted token on the draft model
-                llama_eval(ctx_dft, &drafted.back(), 1, n_past_cur, params.n_threads);
-                ++n_past_cur;
+            // no need to evaluate the last drafted token, since we won't use the result
+            if (i == n_draft - 1) {
+                break;
+            }
+
+            // evaluate the drafted token on the draft model
+            llama_eval(ctx_dft, &drafted.back(), 1, n_past_cur, params.n_threads);
+            ++n_past_cur;
+
+            if (grammar_dft != NULL) {
+                llama_grammar_accept_token(ctx_dft, grammar_dft, id);
            }
        }

@@ -196,6 +249,7 @@ int main(int argc, char ** argv) {
        llama_eval(ctx_tgt, drafted.data(), drafted.size(), n_past_tgt, params.n_threads);
        ++n_past_tgt;

+        // the first token is always proposed by the traget model before the speculation loop
        drafted.erase(drafted.begin());
    }

@@ -226,6 +280,10 @@ int main(int argc, char ** argv) {
    llama_free(ctx_dft);
    llama_free_model(model_dft);

+    if (grammar_dft != NULL) {
+        llama_grammar_free(grammar_dft);
+        llama_grammar_free(grammar_tgt);
+    }
    llama_backend_free();

    fprintf(stderr, "\n\n");
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -327,7 +327,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {

 void * ggml_metal_host_malloc(size_t n) {
    void * data = NULL;
-    const int result = posix_memalign((void **) &data, getpagesize(), n);
+    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
    if (result != 0) {
        metal_printf("%s: error: posix_memalign failed\n", __func__);
        return NULL;
@@ -401,7 +401,7 @@ bool ggml_metal_add_buffer(
            }
        }

-        const size_t size_page = getpagesize();
+        const size_t size_page = sysconf(_SC_PAGESIZE);

        size_t size_aligned = size;
        if ((size_aligned % size_page) != 0) {
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -220,27 +220,32 @@ kernel void kernel_norm(
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }
-    //// broadcast
-    //if (tpitg == 0) {
-    //    sum[0] /= ne00;
-    //}
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    // broadcast
+    if (tpitg == 0) {
+        sum[0] /= ne00;
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
    const float mean  = sum[0];

-    // recenter and VARIANCE
+    // recenter
    device float * y = dst + tgpig*ne00;
-    sum[tpitg] = 0.0f;
    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
        y[i00] = x[i00] - mean;
+    }
+
+    // VARIANCE
+    // parallel sum
+    //
+    // WARNING: combining this loop with the one above will give you wrong results for nth == 256
+    //          I have no idea why, so for now I am keeping them separate. But this behavior is very concerning.
+    //          Tested with:
+    //          ./perplexity -m ./falcon-7b/ggml-model-q4_0.gguf -f wiki.test.raw -ngl 1 -t 4
+    //
+    sum[tpitg] = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
        sum[tpitg] += y[i00] * y[i00];
    }

-    //// VARIANCE
-    //// parallel sum
-    //sum[tpitg] = 0.0f;
-    //for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-    //    sum[tpitg] += y[i00] * y[i00];
-    //}
    // reduce
    threadgroup_barrier(mem_flags::mem_threadgroup);
    for (uint i = ntg/2; i > 0; i /= 2) {
@@ -249,11 +254,11 @@ kernel void kernel_norm(
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }
-    //// broadcast
-    //if (tpitg == 0) {
-    //    sum[0] /= ne00;
-    //}
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    // broadcast
+    if (tpitg == 0) {
+        sum[0] /= ne00;
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
    const float variance = sum[0];

    const float scale = 1.0f/sqrt(variance + eps);
@@ -262,7 +267,6 @@ kernel void kernel_norm(
    }
 }

-
 kernel void kernel_rms_norm(
        device const  void * src0,
        device       float * dst,
@@ -630,7 +634,6 @@ kernel void kernel_mul_mat_f16_f32(
            }
        }
    }
-
 }

 kernel void kernel_alibi_f32(
--- a/ggml-opencl.cpp
+++ b/ggml-opencl.cpp
@@ -1334,7 +1334,7 @@ void ggml_cl_free_data(const struct ggml_tensor* tensor) {
        return;
    }

-    cl_mem mem = (cl_mem)tensor->data;
+    cl_mem mem = (cl_mem)tensor->extra;
    clReleaseMemObject(mem);
 }

@@ -1393,7 +1393,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
    size_t d_size;

    cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
-    cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
    cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst


@@ -1491,7 +1491,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
    size_t d_size;
    cl_mem d_X;
    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
    } else {
        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
    }
@@ -1567,7 +1567,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
    size_t d_size;
    cl_mem d_X;
    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
    } else {
        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
    }
@@ -1697,7 +1697,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                events.emplace_back();
                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
            } else if (src0->backend == GGML_BACKEND_GPU) {
-                d_Q = (cl_mem) src0->data;
+                d_Q = (cl_mem) src0->extra;
            } else {
                GGML_ASSERT(false);
            }
@@ -1860,6 +1860,6 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {

    CL_CHECK(clFinish(queue));

-    tensor->data = dst;
+    tensor->extra = dst;
    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }
--- a/ggml.c
+++ b/ggml.c
@@ -194,7 +194,7 @@ typedef void * thread_ret_t;
 inline static void * ggml_aligned_malloc(size_t size) {
    void * aligned_memory = NULL;
 #ifdef GGML_USE_METAL
-    int result = posix_memalign(&aligned_memory, getpagesize(), size);
+    int result = posix_memalign(&aligned_memory, sysconf(_SC_PAGESIZE), size);
 #else
    int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
 #endif
--- a/grammars/json_arr.gbnf
+++ b/grammars/json_arr.gbnf
@@ -0,0 +1,34 @@
+# This is the same as json.gbnf but we restrict whitespaces at the end of the root array
+# Useful for generating JSON arrays
+
+root   ::= arr
+value  ::= object | array | string | number | ("true" | "false" | "null") ws
+
+arr  ::=
+  "[\n" ws (
+            value
+    (",\n" ws value)*
+  )? "]"
+
+object ::=
+  "{" ws (
+            string ":" ws value
+    ("," ws string ":" ws value)*
+  )? "}" ws
+
+array  ::=
+  "[" ws (
+            value
+    ("," ws value)*
+  )? "]" ws
+
+string ::=
+  "\"" (
+    [^"\\] |
+    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
+  )* "\"" ws
+
+number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
+
+# Optional space: by convention, applied in this grammar after literal chars when allowed
+ws ::= ([ \t\n] ws)?
--- a/k_quants.c
+++ b/k_quants.c
@@ -83,7 +83,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
        float ax = fabsf(x[i]);
        if (ax > amax) { amax = ax; max = x[i]; }
    }
-    if (!amax) { // all zero
+    if (amax < 1e-30f) { // all zero
        for (int i = 0; i < n; ++i) {
            L[i] = 0;
        }
@@ -1086,6 +1086,13 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict

        }

+        if (!max_abs_scale) {
+            memset(&y[i], 0, sizeof(block_q6_K));
+            y[i].d = ggml_fp32_to_fp16(0.f);
+            x += QK_K;
+            continue;
+        }
+
        float iscale = -128.f/max_scale;
        y[i].d = ggml_fp32_to_fp16(1/iscale);
        for (int ib = 0; ib < QK_K/16; ++ib) {
--- a/llama.cpp
+++ b/llama.cpp
@@ -606,16 +606,16 @@ struct llama_mmap {

        if (prefetch > 0) {
            // Advise the kernel to preload the mapped memory
-            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
-                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+            if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
                        strerror(errno));
            }
        }
        if (numa) {
            // advise the kernel not to use readahead
            // (because the next page might not belong on the same node)
-            if (madvise(addr, file->size, MADV_RANDOM)) {
-                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+            if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
                        strerror(errno));
            }
        }
@@ -2942,7 +2942,12 @@ static bool llama_eval_internal(

    // for big prompts, if BLAS is enabled, it is better to use only one thread
    // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
-    n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
+    // TODO: this is mostly important for Apple Silicon where CBLAS is still performing very well
+    //       we still need some threads to process all non-mul_mat ops, but not too much to avoid interfering
+    //       with the BLAS calls. need a better solution
+    if (N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas()) {
+        n_threads = std::min(4, n_threads);
+    }

    struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
@@ -3850,6 +3855,25 @@ void llama_grammar_free(struct llama_grammar * grammar) {
    delete grammar;
 }

+struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar) {
+    llama_grammar * result = new llama_grammar{ grammar->rules, grammar->stacks, grammar->partial_utf8 };
+
+    // redirect elements in stacks to point to new rules
+    for (size_t is = 0; is < result->stacks.size(); is++) {
+        for (size_t ie = 0; ie < result->stacks[is].size(); ie++) {
+            for (size_t ir0 = 0; ir0 < grammar->rules.size(); ir0++) {
+                for (size_t ir1 = 0; ir1 < grammar->rules[ir0].size(); ir1++) {
+                    if (grammar->stacks[is][ie] == &grammar->rules[ir0][ir1]) {
+                         result->stacks[is][ie]  =  &result->rules[ir0][ir1];
+                    }
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
 //
 // sampling
 //
@@ -5340,7 +5364,7 @@ struct llama_context_params llama_context_default_params() {
        /*.seed                        =*/ LLAMA_DEFAULT_SEED,
        /*.n_ctx                       =*/ 512,
        /*.n_batch                     =*/ 512,
-        /*.gpu_layers                  =*/ 0,
+        /*.n_gpu_layers                =*/ 0,
        /*.main_gpu                    =*/ 0,
        /*.tensor_split                =*/ nullptr,
        /*.rope_freq_base              =*/ 10000.0f,
@@ -5357,6 +5381,10 @@ struct llama_context_params llama_context_default_params() {
        /*.embedding                   =*/ false,
    };

+#ifdef GGML_USE_METAL
+    result.n_gpu_layers = 1;
+#endif
+
    return result;
 }

@@ -5549,43 +5577,43 @@ struct llama_context * llama_new_context_with_model(
            }
 #endif
        }
-    }

 #ifdef GGML_USE_METAL
-    if (params.n_gpu_layers > 0) {
-        // this allocates all Metal resources and memory buffers
+        if (params.n_gpu_layers > 0) {
+            // this allocates all Metal resources and memory buffers

-        void * data_ptr  = NULL;
-        size_t data_size = 0;
+            void * data_ptr  = NULL;
+            size_t data_size = 0;

-        if (params.use_mmap) {
-            data_ptr  = ctx->model.mapping->addr;
-            data_size = ctx->model.mapping->size;
-        } else {
-            data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
-            data_size = ggml_get_mem_size  (ctx->model.ctx);
-        }
+            if (params.use_mmap) {
+                data_ptr  = ctx->model.mapping->addr;
+                data_size = ctx->model.mapping->size;
+            } else {
+                data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
+                data_size = ggml_get_mem_size  (ctx->model.ctx);
+            }

-        const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
+            const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);

-        LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+            LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);

 #define LLAMA_METAL_CHECK_BUF(result)                            \
-    if (!(result)) {                                             \
-        LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
-        llama_free(ctx);                                         \
-        return NULL;                                             \
-    }
+            if (!(result)) {                                             \
+                LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
+                llama_free(ctx);                                         \
+                return NULL;                                             \
+            }

-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));

-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));

-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
 #undef LLAMA_METAL_CHECK_BUF
-    }
+        }
 #endif
+    }

 #ifdef GGML_USE_MPI
    ctx->ctx_mpi = ggml_mpi_init();
--- a/llama.h
+++ b/llama.h
@@ -410,6 +410,8 @@ extern "C" {

    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);

+    LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
+
    //
    // Sampling functions
    //
Author	SHA1	Message	Date
Georgi Gerganov	2f689dee06	metal : minor	2023-09-07 15:33:21 +03:00
Georgi Gerganov	efac2d469f	common : don't do warm-up with more than n_batch tokens (close #3058 ) ggml-ci	2023-09-07 15:32:38 +03:00
Georgi Gerganov	783379670a	metal : restore original F16 mat-vec multiplication It works after the norm fixes	2023-09-07 15:20:07 +03:00
Georgi Gerganov	ed92c3d4b2	metal : put warning in kernel_norm to not combine the loops	2023-09-07 14:59:48 +03:00
Georgi Gerganov	5e1c4089d8	metal : fix kernel_norm ggml-ci	2023-09-07 14:11:55 +03:00
Przemysław Pawełczyk	fec2fb19e4	ggml : posixify madvise and pagesize (#3037 ) * llama : use posix_madvise() instead of madvise() derived from BSD sed -i 's,\<madvise\>,posix_&,g;s,\<MADV_,POSIX_&,g' llama.cpp * ggml : use sysconf(_SC_PAGESIZE) instead of getpagesize() derived from BSD sed -i 's,getpagesize(),sysconf(_SC_PAGESIZE),g' ggml.c * metal : use sysconf(_SC_PAGESIZE) instead of getpagesize() derived from BSD sed -i 's,getpagesize(),sysconf(_SC_PAGESIZE),g' ggml-metal.m	2023-09-07 11:15:06 +03:00
Georgi Gerganov	178b1850eb	k-quants : fix zero-weight guard in Q6_K (ref #3040 )	2023-09-06 12:40:57 +03:00
Kerfuffle	ea2c85d5d2	convert-llama-ggml-to-gguf: Try to handle files older than GGJTv3 (#3023 ) * convert-llama-ggmlv3-to-gguf: Try to handle files older than GGJTv3 * Better error messages for files that cannot be converted * Add file type to GGUF output * Rename to convert-llama-ggml-to-gguf.py * Include original file type information in description * Improve some informational output	2023-09-06 02:49:11 -06:00
Cebtenzzre	9912b9efc8	build : add LLAMA_METAL_NDEBUG flag (#3033 )	2023-09-05 18:21:10 -04:00
Cebtenzzre	9e2023156e	make : use new flag variables for recent changes (#3019 )	2023-09-05 15:12:00 -04:00
Cebtenzzre	de2fe892af	examples : replace fprintf to stdout with printf (#3017 )	2023-09-05 15:10:27 -04:00
Erik Scholz	c9c3220c48	convert: fix convert.py not working with int filename_stem (#3028 ) * fix implicit int to string conversion * convert : remove an obsolete pyright comment --------- Co-authored-by: Cebtenzzre <cebtenzzre@gmail.com>	2023-09-05 19:41:00 +02:00
Kawrakow	d59bd97065	Guard against all weights in a super-block being zero (#3010 ) * Guard against all weights in a super-block being zero * Also guard against extremely small weights Closes #2982 --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>	2023-09-05 09:55:33 +02:00
Georgi Gerganov	35938ee3b0	llama : update logic for number of threads when using BLAS	2023-09-05 10:46:39 +03:00
Georgi Gerganov	921772104b	speculative : add grammar support (#2991 ) * speculative : add grammar support * grammars : add json_arr.gbnf * grammar : add comments to new grammar file * grammar : remove one nested level * common : warm-up with 2 tokens - seems to work better * speculative : print draft token pieces * speculative : reuse grammar parser + better logs and comments * speculative : avoid grammar_mem * make : fix speculative build	2023-09-05 08:46:17 +03:00
Georgi Gerganov	2ba85c8609	py : minor	2023-09-04 22:50:50 +03:00
Georgi Gerganov	e36ecdccc8	build : on Mac OS enable Metal by default (#2901 ) * build : on Mac OS enable Metal by default * make : try to fix build on Linux * make : move targets back to the top * make : fix target clean * llama : enable GPU inference by default with Metal * llama : fix vocab_only logic when GPU is enabled * common : better `n_gpu_layers` assignment * readme : update Metal instructions * make : fix merge conflict remnants * gitignore : metal	2023-09-04 22:26:24 +03:00
slaren	bd33e5ab92	ggml-opencl : store GPU buffer in ggml_tensor::extra (#2994 )	2023-09-04 14:59:52 +02:00
Cebtenzzre	3103568144	llama-bench : make cpp file non-executable (#2999 )	2023-09-04 13:40:18 +03:00
Leng Yue	5b8530d88c	make : add speculative example (#3003 )	2023-09-04 13:39:57 +03:00
Aarni Koskela	e4386f417f	server : add a subtle loading animation to the edit box (#2466 ) * editorconfig: add override for the server HTML (which already is 2-space indented) * server: add a subtle loading animation to the edit box	2023-09-04 16:28:55 +08:00