Merge remote-tracking branch 'origin/master' into grammar-fast

Author: ochafik

.github/workflows/bench.yml

@@ -32,7 +32,7 @@ on:
     -  cron: '04 2 * * *'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}-${{ github.event.inputs.sha }}
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}-${{ github.event.inputs.sha }}
   cancel-in-progress: true
 
 jobs:

.github/workflows/server.yml

@@ -23,7 +23,7 @@ on:
     -  cron: '2 4 * * *'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -58,6 +58,7 @@ jobs:
             git \
             cmake \
             python3-pip \
+            python3-venv \
             curl \
             wget \
             language-pack-en \
@@ -100,10 +101,13 @@ jobs:
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
 
-      - name: Tests dependencies
-        id: test_dependencies
+      - name: Setup python env
+        id: pipenv
         run: |
-          pip install -r examples/server/tests/requirements.txt
+          cd examples/server/tests
+          python3 -m venv venv
+          . venv/bin/activate
+          pip install -r requirements.txt
 
       - name: Tests
         id: server_integration_tests

.gitignore

@@ -34,6 +34,7 @@ lcov-report/
 gcovr-report/
 
 build*
+!build.zig
 cmake-build-*
 out/
 tmp/
@@ -100,6 +101,9 @@ qnt-*.txt
 perf-*.txt
 
 examples/jeopardy/results.txt
+examples/server/*.html.hpp
+examples/server/*.js.hpp
+examples/server/*.mjs.hpp
 
 poetry.lock
 poetry.toml

CMakeLists.txt

@@ -43,17 +43,7 @@ else()
     set(LLAMA_METAL_DEFAULT OFF)
 endif()
 
-# TODO: fix this for Android CI
-#       https://github.com/ggerganov/llama.cpp/pull/6716#issuecomment-2061509191
-#if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
-#    set(LLAMA_LLAMAFILE_DEFAULT OFF)
-#else()
-#    set(LLAMA_LLAMAFILE_DEFAULT ON)
-#endif()
-
-# TODO: temporary disable until MoE is fixed
-#       https://github.com/ggerganov/llama.cpp/pull/6716
-set(LLAMA_LLAMAFILE_DEFAULT OFF)
+set(LLAMA_LLAMAFILE_DEFAULT ON)
 
 # general
 option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)

Makefile

@@ -384,10 +384,6 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
-# TODO: temporary disable until MoE is fixed
-#       https://github.com/ggerganov/llama.cpp/pull/6716
-LLAMA_NO_LLAMAFILE := 1
-
 ifndef LLAMA_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
 	OBJS        += sgemm.o
@@ -772,7 +768,7 @@ batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -800,10 +796,19 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/json-schema-to-grammar.mjs.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
+# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+examples/server/%.hpp: examples/server/public/% Makefile
+	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
+		echo "unsigned char $${NAME}[] = {" && \
+		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
+		echo "};" && \
+		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
+	) > $@
+
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

README-sycl.md

@@ -229,12 +229,12 @@ source /opt/intel/oneapi/setvars.sh
 # Build LLAMA with MKL BLAS acceleration for intel GPU
 mkdir -p build && cd build
 
-# Option 1: Use FP16 for better performance in long-prompt inference
-#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-# Option 2: Use FP32 by default
+# Option 1: Use FP32 (recommended for better performance in most cases)
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
+# Option 2: Use FP16
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
 #build all binary
 cmake --build . --config Release -j -v
 ```
@@ -250,12 +250,12 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
 mkdir -p build && cd build
 
-# Option 1: Use FP16 for better performance in long-prompt  inference
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-# Option 2: Use FP32 by default
+# Option 1: Use FP32 (recommended for better performance in most cases)
 cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
+# Option 2: Use FP16
+cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
 #build all binary
 cmake --build . --config Release -j -v
 
@@ -416,6 +416,10 @@ mkdir -p build
 cd build
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+
+# Option 2: Or FP16
 cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 make -j

README.md

@@ -93,6 +93,7 @@ Typically finetunes of the base models below are supported as well.
 
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
+- [x] LLaMA 3 🦙🦙🦙
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
@@ -119,8 +120,9 @@ Typically finetunes of the base models below are supported as well.
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
 - [x] [Gemma](https://ai.google.dev/gemma)
 - [x] [Mamba](https://github.com/state-spaces/mamba)
+- [x] [Grok-1](https://huggingface.co/keyfan/grok-1-hf)
 - [x] [Xverse](https://huggingface.co/models?search=xverse)
-- [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
+- [x] [Command-R models](https://huggingface.co/models?search=CohereForAI/c4ai-command-r)
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
@@ -135,6 +137,8 @@ Typically finetunes of the base models below are supported as well.
 - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
+- [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
+- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 
 **HTTP server**
 
@@ -1117,7 +1121,9 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
+- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
+
+![matmul](media/matmul.png)
 
 ### Docs
 

build.zig

@@ -140,4 +140,33 @@ pub fn build(b: *std.build.Builder) !void {
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }
+
+    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
+    for (server_assets) |asset| {
+        const input_path = b.fmt("examples/server/public/{s}", .{asset});
+        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
+
+        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
+
+        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
+        defer b.allocator.free(input);
+
+        var buf = std.ArrayList(u8).init(b.allocator);
+        defer buf.deinit();
+
+        for (input) |byte| {
+            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
+        }
+
+        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
+        defer b.allocator.free(name);
+        std.mem.replaceScalar(u8, name, '.', '_');
+
+        try std.fs.cwd().writeFile(output_path, b.fmt(
+            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
+            .{ name, buf.items, name, input.len },
+        ));
+
+        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
+    }
 }

ci/run.sh

@@ -160,7 +160,9 @@ function gg_run_test_scripts_debug {
 
     set -e
 
+    # TODO: too slow, run on dedicated node
     (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+   #(cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
 
     set +e
 }
@@ -184,6 +186,7 @@ function gg_run_test_scripts_release {
     set -e
 
     (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
 
     set +e
 }