diff --git a/CMakePresets.json b/CMakePresets.json index d22ffa490..ae45d60af 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -48,10 +48,23 @@ } }, + { + "name": "arm64-apple-clang", "hidden": true, + "architecture": { "value": "arm64", "strategy": "external" }, + "toolset": { "value": "host=x64", "strategy": "external" }, + "cacheVariables": { + "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake" + } + }, + { "name": "arm64-windows-llvm-debug" , "inherits": [ "base", "arm64-windows-llvm", "debug" ] }, { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] }, { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] }, + { "name": "arm64-apple-clang-debug" , "inherits": [ "base", "arm64-apple-clang", "debug" ] }, + { "name": "arm64-apple-clang-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg" ] }, + { "name": "arm64-apple-clang+static-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] }, + { "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] }, { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] }, { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] }, diff --git a/Makefile b/Makefile index 719f45d16..051436344 100644 --- a/Makefile +++ b/Makefile @@ -34,6 +34,7 @@ BUILD_TARGETS = \ llama-save-load-state \ llama-server \ llama-simple \ + llama-simple-chat \ llama-speculative \ llama-tokenize \ llama-vdot \ @@ -1287,6 +1288,11 @@ llama-simple: examples/simple/simple.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) +llama-simple-chat: examples/simple-chat/simple-chat.cpp \ + $(OBJ_ALL) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + llama-tokenize: examples/tokenize/tokenize.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) diff --git a/README.md b/README.md index 8fe1f4b4b..0378a674e 100644 --- a/README.md +++ b/README.md @@ -17,7 +17,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ## Hot topics -- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669** +- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123 +- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669 - Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor) ---- diff --git a/ci/run.sh b/ci/run.sh index dc26d94ee..21b62dd1e 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -326,36 +326,36 @@ function gg_run_open_llama_7b_v2 { ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k - (time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log + (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" @@ -460,34 +460,34 @@ function gg_run_pythia_1_4b { ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k - (time ./bin/llama-cli --model ${model_f16} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-cli --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/llama-cli --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/llama-cli --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/llama-cli --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/llama-cli --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/llama-cli --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/llama-cli --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/llama-cli --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/llama-cli --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/llama-cli --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/llama-cli --model ${model_f16} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-cli --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-cli --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/llama-cli --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/llama-cli --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/llama-cli --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/llama-cli --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/llama-cli --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/llama-cli --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/llama-cli --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/llama-cli --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log + (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/llama-save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" @@ -591,36 +591,36 @@ function gg_run_pythia_2_8b { ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k - (time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log + (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" @@ -706,8 +706,8 @@ function gg_run_embd_bge_small { ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0 - (time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log set +e } @@ -752,7 +752,7 @@ function gg_run_rerank_tiny { model_f16="${path_models}/ggml-model-f16.gguf" # for this model, the SEP token is "" - (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?hi\nwhat is panda?it's a bear\nwhat is panda?The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log + (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?hi\nwhat is panda?it's a bear\nwhat is panda?The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log # sample output # rerank score 0: 0.029 diff --git a/cmake/arm64-apple-clang.cmake b/cmake/arm64-apple-clang.cmake new file mode 100644 index 000000000..5fcd2882a --- /dev/null +++ b/cmake/arm64-apple-clang.cmake @@ -0,0 +1,16 @@ +set( CMAKE_SYSTEM_NAME Darwin ) +set( CMAKE_SYSTEM_PROCESSOR arm64 ) + +set( target arm64-apple-darwin-macho ) + +set( CMAKE_C_COMPILER clang ) +set( CMAKE_CXX_COMPILER clang++ ) + +set( CMAKE_C_COMPILER_TARGET ${target} ) +set( CMAKE_CXX_COMPILER_TARGET ${target} ) + +set( arch_c_flags "-march=armv8.4-a -fvectorize -ffp-model=fast -fno-finite-math-only" ) +set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function" ) + +set( CMAKE_C_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" ) +set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" ) diff --git a/common/common.h b/common/common.h index cd5a8e051..727f85baa 100644 --- a/common/common.h +++ b/common/common.h @@ -155,7 +155,7 @@ struct common_sampler_params { struct common_params { int32_t n_predict = -1; // new tokens to predict - int32_t n_ctx = 0; // context size + int32_t n_ctx = 4096; // context size int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS) int32_t n_ubatch = 512; // physical batch size for prompt processing (must be >=32 to use BLAS) int32_t n_keep = 0; // number of tokens to keep from initial prompt diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index a34dabe23..76ee6cef5 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -72,7 +72,8 @@ class Model: def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, use_temp_file: bool = False, eager: bool = False, metadata_override: Path | None = None, model_name: str | None = None, - split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): + split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, + small_first_shard: bool = False, hparams: dict[str, Any] | None = None): if type(self) is Model: raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") @@ -87,7 +88,7 @@ class Model: self.is_safetensors = len(self.part_names) > 0 if not self.is_safetensors: self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin") - self.hparams = Model.load_hparams(self.dir_model) + self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"]) self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) self.tensor_names = None @@ -1541,6 +1542,17 @@ class LlamaModel(Model): special_vocab._set_special_token("eot", 32010) special_vocab.add_to_gguf(self.gguf_writer) + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + if "add_prefix_space" in tokenizer_config_json: + self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"]) + + # Apply to granite small models only + if self.hparams.get("vocab_size", 32000) == 49152: + self.gguf_writer.add_add_bos_token(False) + def set_gguf_parameters(self): super().set_gguf_parameters() hparams = self.hparams @@ -1557,17 +1569,6 @@ class LlamaModel(Model): self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) - tokenizer_config_file = self.dir_model / 'tokenizer_config.json' - if tokenizer_config_file.is_file(): - with open(tokenizer_config_file, "r", encoding="utf-8") as f: - tokenizer_config_json = json.load(f) - if "add_prefix_space" in tokenizer_config_json: - self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"]) - - # Apply to granite small models only - if self.hparams.get("vocab_size", 32000) == 49152: - self.gguf_writer.add_add_bos_token(False) - @staticmethod def permute(weights: Tensor, n_head: int, n_head_kv: int | None): if n_head_kv is not None and n_head != n_head_kv: diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py index 915e21836..ed1014cae 100755 --- a/convert_lora_to_gguf.py +++ b/convert_lora_to_gguf.py @@ -12,6 +12,7 @@ import json from math import prod from pathlib import Path from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast +from transformers import AutoConfig import torch @@ -256,8 +257,8 @@ def parse_args() -> argparse.Namespace: help="only print out what will be done, without writing any new files", ) parser.add_argument( - "--base", type=Path, required=True, - help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required", + "--base", type=Path, + help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config", ) parser.add_argument( "lora_path", type=Path, @@ -267,6 +268,12 @@ def parse_args() -> argparse.Namespace: return parser.parse_args() +def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]: + # normally, adapter does not come with base model config, we need to load it from AutoConfig + config = AutoConfig.from_pretrained(hf_model_id) + return config.to_dict() + + if __name__ == '__main__': args = parse_args() logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) @@ -281,7 +288,7 @@ if __name__ == '__main__': ftype = ftype_map[args.outtype] - dir_base_model: Path = args.base + dir_base_model: Path | None = args.base dir_lora: Path = args.lora_path lora_config = dir_lora / "adapter_config.json" input_model = dir_lora / "adapter_model.safetensors" @@ -301,9 +308,29 @@ if __name__ == '__main__': input_model = os.path.join(dir_lora, "adapter_model.bin") lora_model = torch.load(input_model, map_location="cpu", weights_only=True) + # load LoRA config + with open(lora_config, "r") as f: + lparams: dict[str, Any] = json.load(f) + # load base model - logger.info(f"Loading base model: {dir_base_model.name}") - hparams = Model.load_hparams(dir_base_model) + if dir_base_model is None: + if "base_model_name_or_path" in lparams: + model_id = lparams["base_model_name_or_path"] + logger.info(f"Loading base model from Hugging Face: {model_id}") + try: + hparams = load_hparams_from_hf(model_id) + except OSError as e: + logger.error(f"Failed to load base model config: {e}") + logger.error("Please try downloading the base model and add its path to --base") + sys.exit(1) + else: + logger.error("'base_model_name_or_path' is not found in adapter_config.json") + logger.error("Base model config is required. Please download the base model and add its path to --base") + sys.exit(1) + else: + logger.info(f"Loading base model: {dir_base_model.name}") + hparams = Model.load_hparams(dir_base_model) + with torch.inference_mode(): try: model_class = Model.from_model_architecture(hparams["architectures"][0]) @@ -323,13 +350,15 @@ if __name__ == '__main__': self.dir_model_card = dir_lora_model self.lora_alpha = float(lora_alpha) + def set_vocab(self): + pass + def set_type(self): self.gguf_writer.add_type(gguf.GGUFType.ADAPTER) self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora") def set_gguf_parameters(self): self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha) - super().set_gguf_parameters() def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: # Never add extra tensors (e.g. rope_freqs) for LoRA adapters @@ -350,7 +379,7 @@ if __name__ == '__main__': logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor") if ".embed_tokens.weight" in name or ".lm_head.weight" in name: logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning") - logger.error("Hint: if you are using TRL, make sure not to call setup_chat_format()") + logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948") sys.exit(1) if base_name in tensor_map: @@ -384,9 +413,6 @@ if __name__ == '__main__': yield (dest_name + ".lora_a", lora_a) yield (dest_name + ".lora_b", lora_b) - with open(lora_config, "r") as f: - lparams: dict[str, Any] = json.load(f) - alpha: float = lparams["lora_alpha"] model_instance = LoraModel( @@ -399,6 +425,7 @@ if __name__ == '__main__': dry_run=args.dry_run, dir_lora_model=dir_lora, lora_alpha=alpha, + hparams=hparams, ) logger.info("Exporting model...") diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index ead630661..6df318c19 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -49,6 +49,7 @@ else() endif() add_subdirectory(save-load-state) add_subdirectory(simple) + add_subdirectory(simple-chat) add_subdirectory(speculative) add_subdirectory(tokenize) endif() diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 7953b5065..8531a784d 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -247,6 +247,7 @@ struct server_slot { if (is_processing()) { SLT_INF(*this, "stop processing: n_past = %d, truncated = %d\n", n_past, truncated); + t_last_used = ggml_time_us(); t_token_generation = (ggml_time_us() - t_start_generation) / 1e3; state = SLOT_STATE_IDLE; callback_on_release(id); @@ -725,12 +726,12 @@ struct server_context { return nullptr; } - server_slot * get_available_slot(const std::string & prompt) { + server_slot * get_available_slot(const server_task & task) { server_slot * ret = nullptr; // find the slot that has at least n% prompt similarity - if (ret == nullptr && slot_prompt_similarity != 0.0f && !prompt.empty()) { - int max_lcp_len = 0; + if (ret == nullptr && slot_prompt_similarity != 0.0f) { + int lcs_len = 0; float similarity = 0; for (server_slot & slot : slots) { @@ -740,25 +741,26 @@ struct server_context { } // skip the slot if it does not contains cached tokens - if (slot.prompt_tokens.empty()) { + if (slot.cache_tokens.empty()) { continue; } - // length of the Longest Common Prefix between the current slot's prompt and the input prompt - int lcp_len = longest_common_prefix(slot.cache_tokens, slot.prompt_tokens); + // length of the Longest Common Subsequence between the current slot's prompt and the input prompt + int cur_lcs_len = longest_common_subsequence(slot.cache_tokens, task.prompt_tokens); - // fraction of the common substring length compared to the current slot's prompt length - similarity = static_cast(lcp_len) / static_cast(slot.prompt_tokens.size()); + // fraction of the common subsequence length compared to the current slot's prompt length + float cur_similarity = static_cast(cur_lcs_len) / static_cast(slot.cache_tokens.size()); // select the current slot if the criteria match - if (lcp_len > max_lcp_len && similarity > slot_prompt_similarity) { - max_lcp_len = lcp_len; + if (cur_lcs_len > lcs_len && cur_similarity > slot_prompt_similarity) { + lcs_len = cur_lcs_len; + similarity = cur_similarity; ret = &slot; } } if (ret != nullptr) { - SLT_DBG(*ret, "selected slot by lcp similarity, max_lcp_len = %d, similarity = %f\n", max_lcp_len, similarity); + SLT_DBG(*ret, "selected slot by lcs similarity, lcs_len = %d, similarity = %f\n", lcs_len, similarity); } } @@ -1514,18 +1516,7 @@ struct server_context { { const int id_slot = json_value(task.data, "id_slot", -1); - server_slot * slot; - - if (id_slot != -1) { - slot = get_slot_by_id(id_slot); - } else { - std::string prompt; - if (task.data.contains("prompt") && task.data.at("prompt").is_string()) { - prompt = json_value(task.data, "prompt", std::string()); - } - - slot = get_available_slot(prompt); - } + server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task); if (slot == nullptr) { // if no slot is available, we defer this task for processing later @@ -2714,8 +2705,8 @@ int main(int argc, char ** argv) { }; const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) { - if (ctx_server.params.embedding || ctx_server.params.reranking) { - res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED)); + if (ctx_server.params.embedding) { + res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return; } @@ -2820,8 +2811,8 @@ int main(int argc, char ** argv) { // TODO: maybe merge this function with "handle_completions_generic" const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) { - if (ctx_server.params.embedding || ctx_server.params.reranking) { - res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED)); + if (ctx_server.params.embedding) { + res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return; } @@ -2946,11 +2937,6 @@ int main(int argc, char ** argv) { }; const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) { - // TODO: somehow clean up this checks in the future - if (!ctx_server.params.embedding || ctx_server.params.reranking) { - res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings` and without `--reranking`", ERROR_TYPE_NOT_SUPPORTED)); - return; - } const json body = json::parse(req.body); bool is_openai = false; @@ -3002,10 +2988,11 @@ int main(int argc, char ** argv) { }; const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) { - if (!ctx_server.params.reranking) { - res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED)); + if (!ctx_server.params.reranking || ctx_server.params.embedding) { + res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED)); return; } + const json body = json::parse(req.body); // TODO: implement @@ -3259,7 +3246,7 @@ int main(int argc, char ** argv) { ctx_server.queue_tasks.terminate(); }; - LOG_INF("%s: server is listening on %s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port); + LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port); ctx_server.queue_tasks.start_loop(); diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index 58f5a5684..c47ed3e47 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -439,18 +439,60 @@ static std::string gen_chatcmplid() { // other common utils // -static size_t longest_common_prefix(const std::vector & a, const std::vector & b) { +static size_t longest_common_prefix(const llama_tokens & a, const llama_tokens & b) { size_t i; for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {} return i; } -static size_t longest_common_prefix(const std::string & a, const std::string & b) { - size_t i; - for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {} +static size_t longest_common_subsequence(const llama_tokens & a, const llama_tokens & b) { + // check for empty sequences + if (a.empty() || b.empty()) { + return 0; + } - return i; + // get the lengths of the input sequences + size_t a_len = a.size(); + size_t b_len = b.size(); + + // initialize the maximum length of the longest common subsequence (LCS) + size_t max_length = 0; + + // use two rows instead of a 2D matrix to optimize space + std::vector prev_row(b_len + 1, 0); + std::vector curr_row(b_len + 1, 0); + + // iterate through the elements of a + for (size_t i = 1; i <= a_len; i++) { + // iterate through the elements of b + for (size_t j = 1; j <= b_len; j++) { + // if elements at the current positions match + if (a[i - 1] == b[j - 1]) { + // if it's the first element of either sequences, set LCS length to 1 + if (i == 1 || j == 1) { + curr_row[j] = 1; + } else { + // increment LCS length by 1 compared to the previous element + curr_row[j] = prev_row[j - 1] + 1; + } + + // update max_length if necessary + if (curr_row[j] > max_length) { + max_length = curr_row[j]; + } + } else { + // reset LCS length if elements don't match + curr_row[j] = 0; + } + } + + // update the previous row for the next iteration + prev_row = curr_row; + } + + // return the maximum length of the LCS + return max_length; } static bool ends_with(const std::string & str, const std::string & suffix) { diff --git a/examples/simple-chat/CMakeLists.txt b/examples/simple-chat/CMakeLists.txt new file mode 100644 index 000000000..87723533b --- /dev/null +++ b/examples/simple-chat/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET llama-simple-chat) +add_executable(${TARGET} simple-chat.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/simple-chat/README.md b/examples/simple-chat/README.md new file mode 100644 index 000000000..f0099ce3d --- /dev/null +++ b/examples/simple-chat/README.md @@ -0,0 +1,7 @@ +# llama.cpp/example/simple-chat + +The purpose of this example is to demonstrate a minimal usage of llama.cpp to create a simple chat program using the chat template from the GGUF file. + +```bash +./llama-simple-chat -m Meta-Llama-3.1-8B-Instruct.gguf -c 2048 +... diff --git a/examples/simple-chat/simple-chat.cpp b/examples/simple-chat/simple-chat.cpp new file mode 100644 index 000000000..5f9973163 --- /dev/null +++ b/examples/simple-chat/simple-chat.cpp @@ -0,0 +1,197 @@ +#include "llama.h" +#include +#include +#include +#include +#include + +static void print_usage(int, char ** argv) { + printf("\nexample usage:\n"); + printf("\n %s -m model.gguf [-c context_size] [-ngl n_gpu_layers]\n", argv[0]); + printf("\n"); +} + +int main(int argc, char ** argv) { + std::string model_path; + int ngl = 99; + int n_ctx = 2048; + + // parse command line arguments + for (int i = 1; i < argc; i++) { + try { + if (strcmp(argv[i], "-m") == 0) { + if (i + 1 < argc) { + model_path = argv[++i]; + } else { + print_usage(argc, argv); + return 1; + } + } else if (strcmp(argv[i], "-c") == 0) { + if (i + 1 < argc) { + n_ctx = std::stoi(argv[++i]); + } else { + print_usage(argc, argv); + return 1; + } + } else if (strcmp(argv[i], "-ngl") == 0) { + if (i + 1 < argc) { + ngl = std::stoi(argv[++i]); + } else { + print_usage(argc, argv); + return 1; + } + } else { + print_usage(argc, argv); + return 1; + } + } catch (std::exception & e) { + fprintf(stderr, "error: %s\n", e.what()); + print_usage(argc, argv); + return 1; + } + } + if (model_path.empty()) { + print_usage(argc, argv); + return 1; + } + + // only print errors + llama_log_set([](enum ggml_log_level level, const char * text, void * /* user_data */) { + if (level >= GGML_LOG_LEVEL_ERROR) { + fprintf(stderr, "%s", text); + } + }, nullptr); + + // initialize the model + llama_model_params model_params = llama_model_default_params(); + model_params.n_gpu_layers = ngl; + + llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params); + if (!model) { + fprintf(stderr , "%s: error: unable to load model\n" , __func__); + return 1; + } + + // initialize the context + llama_context_params ctx_params = llama_context_default_params(); + ctx_params.n_ctx = n_ctx; + ctx_params.n_batch = n_ctx; + + llama_context * ctx = llama_new_context_with_model(model, ctx_params); + if (!ctx) { + fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__); + return 1; + } + + // initialize the sampler + llama_sampler * smpl = llama_sampler_chain_init(llama_sampler_chain_default_params()); + llama_sampler_chain_add(smpl, llama_sampler_init_min_p(0.05f, 1)); + llama_sampler_chain_add(smpl, llama_sampler_init_temp(0.8f)); + llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED)); + + // helper function to evaluate a prompt and generate a response + auto generate = [&](const std::string & prompt) { + std::string response; + + // tokenize the prompt + const int n_prompt_tokens = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true); + std::vector prompt_tokens(n_prompt_tokens); + if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) { + GGML_ABORT("failed to tokenize the prompt\n"); + } + + // prepare a batch for the prompt + llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size()); + llama_token new_token_id; + while (true) { + // check if we have enough space in the context to evaluate this batch + int n_ctx = llama_n_ctx(ctx); + int n_ctx_used = llama_get_kv_cache_used_cells(ctx); + if (n_ctx_used + batch.n_tokens > n_ctx) { + printf("\033[0m\n"); + fprintf(stderr, "context size exceeded\n"); + exit(0); + } + + if (llama_decode(ctx, batch)) { + GGML_ABORT("failed to decode\n"); + } + + // sample the next token + new_token_id = llama_sampler_sample(smpl, ctx, -1); + + // is it an end of generation? + if (llama_token_is_eog(model, new_token_id)) { + break; + } + + // convert the token to a string, print it and add it to the response + char buf[256]; + int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true); + if (n < 0) { + GGML_ABORT("failed to convert token to piece\n"); + } + std::string piece(buf, n); + printf("%s", piece.c_str()); + fflush(stdout); + response += piece; + + // prepare the next batch with the sampled token + batch = llama_batch_get_one(&new_token_id, 1); + } + + return response; + }; + + std::vector messages; + std::vector formatted(llama_n_ctx(ctx)); + int prev_len = 0; + while (true) { + // get user input + printf("\033[32m> \033[0m"); + std::string user; + std::getline(std::cin, user); + + if (user.empty()) { + break; + } + + // add the user input to the message list and format it + messages.push_back({"user", strdup(user.c_str())}); + int new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size()); + if (new_len > (int)formatted.size()) { + formatted.resize(new_len); + new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size()); + } + if (new_len < 0) { + fprintf(stderr, "failed to apply the chat template\n"); + return 1; + } + + // remove previous messages to obtain the prompt to generate the response + std::string prompt(formatted.begin() + prev_len, formatted.begin() + new_len); + + // generate a response + printf("\033[33m"); + std::string response = generate(prompt); + printf("\n\033[0m"); + + // add the response to the messages + messages.push_back({"assistant", strdup(response.c_str())}); + prev_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), false, nullptr, 0); + if (prev_len < 0) { + fprintf(stderr, "failed to apply the chat template\n"); + return 1; + } + } + + // free resources + for (auto & msg : messages) { + free(const_cast(msg.content)); + } + llama_sampler_free(smpl); + llama_free(ctx); + llama_free_model(model); + + return 0; +} diff --git a/ggml/include/ggml-cpp.h b/ggml/include/ggml-cpp.h new file mode 100644 index 000000000..219361af4 --- /dev/null +++ b/ggml/include/ggml-cpp.h @@ -0,0 +1,38 @@ +#pragma once + +#ifndef __cplusplus +#error "This header is for C++ only" +#endif + +#include "ggml.h" +#include "ggml-alloc.h" +#include "ggml-backend.h" +#include + +// Smart pointers for ggml types + +// ggml + +struct ggml_context_deleter { void operator()(ggml_context * ctx) { ggml_free(ctx); } }; +struct gguf_context_deleter { void operator()(gguf_context * ctx) { gguf_free(ctx); } }; + +typedef std::unique_ptr ggml_context_ptr; +typedef std::unique_ptr gguf_context_ptr; + +// ggml-alloc + +struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } }; + +typedef std::unique_ptr ggml_gallocr_ptr; + +// ggml-backend + +struct ggml_backend_deleter { void operator()(ggml_backend_t backend) { ggml_backend_free(backend); } }; +struct ggml_backend_buffer_deleter { void operator()(ggml_backend_buffer_t buffer) { ggml_backend_buffer_free(buffer); } }; +struct ggml_backend_event_deleter { void operator()(ggml_backend_event_t event) { ggml_backend_event_free(event); } }; +struct ggml_backend_sched_deleter { void operator()(ggml_backend_sched_t sched) { ggml_backend_sched_free(sched); } }; + +typedef std::unique_ptr ggml_backend_ptr; +typedef std::unique_ptr ggml_backend_buffer_ptr; +typedef std::unique_ptr ggml_backend_event_ptr; +typedef std::unique_ptr ggml_backend_sched_ptr; diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index de3c706fc..2d93f31fa 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -217,7 +217,6 @@ #define GGML_MAX_DIMS 4 #define GGML_MAX_PARAMS 2048 -#define GGML_MAX_CONTEXTS 64 #define GGML_MAX_SRC 10 #define GGML_MAX_N_THREADS 512 #define GGML_MAX_OP_PARAMS 64 @@ -559,10 +558,10 @@ extern "C" { enum ggml_log_level { GGML_LOG_LEVEL_NONE = 0, - GGML_LOG_LEVEL_INFO = 1, - GGML_LOG_LEVEL_WARN = 2, - GGML_LOG_LEVEL_ERROR = 3, - GGML_LOG_LEVEL_DEBUG = 4, + GGML_LOG_LEVEL_DEBUG = 1, + GGML_LOG_LEVEL_INFO = 2, + GGML_LOG_LEVEL_WARN = 3, + GGML_LOG_LEVEL_ERROR = 4, GGML_LOG_LEVEL_CONT = 5, // continue previous log }; @@ -656,13 +655,6 @@ extern "C" { void * abort_callback_data; }; - // scratch buffer - struct ggml_scratch { - size_t offs; - size_t size; - void * data; - }; - struct ggml_init_params { // memory pool size_t mem_size; // bytes @@ -760,12 +752,12 @@ extern "C" { // main - GGML_API struct ggml_context * ggml_init(struct ggml_init_params params); - GGML_API void ggml_free(struct ggml_context * ctx); + GGML_API struct ggml_context * ggml_init (struct ggml_init_params params); + GGML_API void ggml_reset(struct ggml_context * ctx); + GGML_API void ggml_free (struct ggml_context * ctx); GGML_API size_t ggml_used_mem(const struct ggml_context * ctx); - GGML_API size_t ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch); GGML_API bool ggml_get_no_alloc(struct ggml_context * ctx); GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index aa405e4d0..0764a8d90 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -800,6 +800,7 @@ if (GGML_KOMPUTE) kompute-shaders/op_mul_mat_q8_0.comp kompute-shaders/op_mul_mat_q4_0.comp kompute-shaders/op_mul_mat_q4_1.comp + kompute-shaders/op_mul_mat_q4_k.comp kompute-shaders/op_mul_mat_q6_k.comp kompute-shaders/op_getrows_f32.comp kompute-shaders/op_getrows_f16.comp @@ -833,6 +834,7 @@ if (GGML_KOMPUTE) shaderop_mul_mat_q8_0.h shaderop_mul_mat_q4_0.h shaderop_mul_mat_q4_1.h + shaderop_mul_mat_q4_k.h shaderop_mul_mat_q6_k.h shaderop_getrows_f32.h shaderop_getrows_f16.h @@ -1366,6 +1368,7 @@ add_library(ggml ../include/ggml.h ../include/ggml-alloc.h ../include/ggml-backend.h + ../include/ggml-cpp.h ggml.c ggml-alloc.c ggml-backend.cpp @@ -1400,7 +1403,7 @@ list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads) find_library(MATH_LIBRARY m) if (MATH_LIBRARY) - if (NOT WIN32 OR NOT GGML_SYCL) + if (NOT WIN32 OR NOT DEFINED ENV{ONEAPI_ROOT}) list(APPEND GGML_EXTRA_LIBS_PRIVATE m) endif() endif() diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index f397f6252..c2afdf391 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -1508,7 +1508,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co return -1; } -#if 1 +#if 0 #define GGML_SCHED_MAX_SPLITS_DEBUG 4096 static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 087091516..b57f1b3b7 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -3107,18 +3107,20 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g } return false; } break; + case GGML_OP_NORM: + case GGML_OP_RMS_NORM: + return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0; + break; case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: - case GGML_OP_NORM: case GGML_OP_ADD: case GGML_OP_ADD1: case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: - case GGML_OP_RMS_NORM: case GGML_OP_SCALE: case GGML_OP_SQR: case GGML_OP_SQRT: diff --git a/ggml/src/ggml-kompute.cpp b/ggml/src/ggml-kompute.cpp index fea69fb04..2fea9e4cc 100644 --- a/ggml/src/ggml-kompute.cpp +++ b/ggml/src/ggml-kompute.cpp @@ -20,6 +20,7 @@ #include "shaderop_mul_mat_q8_0.h" #include "shaderop_mul_mat_q4_0.h" #include "shaderop_mul_mat_q4_1.h" +#include "shaderop_mul_mat_q4_k.h" #include "shaderop_mul_mat_q6_k.h" #include "shaderop_mul_mat_mat_f32.h" #include "shaderop_getrows_f32.h" @@ -1067,6 +1068,40 @@ static void ggml_vk_mul_mat_q8_0(Args&&... args) { ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward(args)...); } +static void ggml_vk_mul_mat_q4_k( + kp::Sequence& seq, + const std::shared_ptr& inA, + const std::shared_ptr& inB, + const std::shared_ptr& out, + uint32_t inAOff, uint32_t inBOff, uint32_t outOff, + int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10, + int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0, + int32_t ne1, int32_t r2, int32_t r3 +) { + const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv, + kp::shader_data::op_mul_mat_q4_k_comp_spv_len); + + struct PushConstants { + uint32_t inAOff, inBOff, outOff; + int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3; + } pushConsts { + 0, 0, 0, + ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3 + }; + + std::shared_ptr s_algo = nullptr; + if (!komputeManager()->hasAlgorithm(__func__)) { + s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}, {}, {pushConsts}); + } else { + s_algo = komputeManager()->getAlgorithm(__func__); + s_algo->setTensors({inA, inB, out}); + s_algo->setWorkgroup({unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}); + s_algo->setPushConstants({pushConsts}); + s_algo->updateDescriptors(s_kompute_context->pool.get()); + } + seq.record(s_algo); +} + static void ggml_vk_mul_mat_q6_k( kp::Sequence& seq, const std::shared_ptr& inA, @@ -1384,6 +1419,7 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons case GGML_TYPE_Q8_0: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: + case GGML_TYPE_Q4_K: return true; default: ; @@ -1635,6 +1671,12 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 ); break; + case GGML_TYPE_Q4_K: + ggml_vk_mul_mat_q4_k( + seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03 + ); + break; case GGML_TYPE_Q6_K: ggml_vk_mul_mat_q6_k( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index 83c37ea9c..a8e78c4db 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -1047,7 +1047,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor return buf; } - buf->size = size; vk::BufferCreateInfo buffer_create_info{ vk::BufferCreateFlags(), size, @@ -1075,7 +1074,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor if (memory_type_index == UINT32_MAX) { device->device.destroyBuffer(buf->buffer); - buf->size = 0; throw vk::OutOfDeviceMemoryError("No suitable memory type found"); } @@ -1092,13 +1090,11 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor } catch (const vk::SystemError& e) { device->device.destroyBuffer(buf->buffer); - buf->size = 0; throw e; } } else { // Out of Host/Device memory, clean up buffer device->device.destroyBuffer(buf->buffer); - buf->size = 0; throw e; } } @@ -1111,6 +1107,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0); buf->device = device; + buf->size = size; #ifdef GGML_VULKAN_MEMORY_DEBUG device->memory_logger->log_allocation(buf, size); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index f7fa6cf73..f47f867f6 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -306,6 +306,7 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) { } #define GGML_DEBUG 0 + #define GGML_GELU_FP16 #define GGML_GELU_QUICK_FP16 @@ -2014,18 +2015,14 @@ static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); struct ggml_context { size_t mem_size; - void* mem_buffer; + void * mem_buffer; bool mem_buffer_owned; bool no_alloc; - bool no_alloc_save; // this is used to save the no_alloc state when using scratch buffers int n_objects; struct ggml_object * objects_begin; struct ggml_object * objects_end; - - struct ggml_scratch scratch; - struct ggml_scratch scratch_save; }; struct ggml_context_container { @@ -3267,7 +3264,6 @@ struct ggml_numa_nodes { // struct ggml_state { - struct ggml_context_container contexts[GGML_MAX_CONTEXTS]; struct ggml_numa_nodes numa; }; @@ -3849,7 +3845,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { const uint64_t t_start = ggml_time_us(); UNUSED(t_start); g_state = (struct ggml_state) { - /*.contexts =*/ { { 0 } }, /*.numa =*/ { .n_nodes = 0, .total_cpus = 0, @@ -3868,26 +3863,9 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { is_first_call = false; } - // find non-used context in g_state - struct ggml_context * ctx = NULL; + ggml_critical_section_end(); - for (int i = 0; i < GGML_MAX_CONTEXTS; i++) { - if (!g_state.contexts[i].used) { - g_state.contexts[i].used = true; - ctx = &g_state.contexts[i].context; - - GGML_PRINT_DEBUG("%s: found unused context %d\n", __func__, i); - break; - } - } - - if (ctx == NULL) { - GGML_PRINT_DEBUG("%s: no unused context found\n", __func__); - - ggml_critical_section_end(); - - return NULL; - } + struct ggml_context * ctx = GGML_MALLOC(sizeof(struct ggml_context)); // allow to call ggml_init with 0 size if (params.mem_size == 0) { @@ -3901,12 +3879,9 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { /*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : ggml_aligned_malloc(mem_size), /*.mem_buffer_owned =*/ params.mem_buffer ? false : true, /*.no_alloc =*/ params.no_alloc, - /*.no_alloc_save =*/ params.no_alloc, /*.n_objects =*/ 0, /*.objects_begin =*/ NULL, /*.objects_end =*/ NULL, - /*.scratch =*/ { 0, 0, NULL, }, - /*.scratch_save =*/ { 0, 0, NULL, }, }; GGML_ASSERT(ctx->mem_buffer != NULL); @@ -3915,56 +3890,35 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { GGML_PRINT_DEBUG("%s: context initialized\n", __func__); - ggml_critical_section_end(); - return ctx; } +void ggml_reset(struct ggml_context * ctx) { + if (ctx == NULL) { + return; + } + + ctx->n_objects = 0; + ctx->objects_begin = NULL; + ctx->objects_end = NULL; +} + void ggml_free(struct ggml_context * ctx) { if (ctx == NULL) { return; } - // make this function thread safe - ggml_critical_section_start(); - - bool found = false; - - for (int i = 0; i < GGML_MAX_CONTEXTS; i++) { - if (&g_state.contexts[i].context == ctx) { - g_state.contexts[i].used = false; - - GGML_PRINT_DEBUG("%s: context %d has been freed. memory used = %zu\n", - __func__, i, ggml_used_mem(ctx)); - - if (ctx->mem_buffer_owned) { - ggml_aligned_free(ctx->mem_buffer, ctx->mem_size); - } - - found = true; - break; - } + if (ctx->mem_buffer_owned) { + ggml_aligned_free(ctx->mem_buffer, ctx->mem_size); } - if (!found) { - GGML_PRINT_DEBUG("%s: context not found\n", __func__); - } - - ggml_critical_section_end(); + GGML_FREE(ctx); } size_t ggml_used_mem(const struct ggml_context * ctx) { return ctx->objects_end == NULL ? 0 : ctx->objects_end->offs + ctx->objects_end->size; } -size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) { - const size_t result = ctx->scratch.data ? ctx->scratch.offs : 0; - - ctx->scratch = scratch; - - return result; -} - bool ggml_get_no_alloc(struct ggml_context * ctx) { return ctx->no_alloc; } @@ -3992,27 +3946,6 @@ size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) { return max_size; } -// IMPORTANT: -// when creating "opt" tensors, always save and load the scratch buffer -// this is an error prone process, but it is necessary to support inplace -// operators when using scratch buffers -// TODO: implement a better way -static void ggml_scratch_save(struct ggml_context * ctx) { - // this is needed to allow opt tensors to store their data - // TODO: again, need to find a better way - ctx->no_alloc_save = ctx->no_alloc; - ctx->no_alloc = false; - - ctx->scratch_save = ctx->scratch; - ctx->scratch.data = NULL; -} - -static void ggml_scratch_load(struct ggml_context * ctx) { - ctx->no_alloc = ctx->no_alloc_save; - - ctx->scratch = ctx->scratch_save; -} - //////////////////////////////////////////////////////////////////////////////// static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml_object_type type, size_t size) { @@ -4093,29 +4026,13 @@ static struct ggml_tensor * ggml_new_tensor_impl( size_t obj_alloc_size = 0; if (view_src == NULL && !ctx->no_alloc) { - if (ctx->scratch.data != NULL) { - // allocate tensor data in the scratch buffer - if (ctx->scratch.offs + data_size > ctx->scratch.size) { - GGML_LOG_WARN("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", - __func__, ctx->scratch.offs + data_size, ctx->scratch.size); - assert(false); - return NULL; - } - - data = (char * const) ctx->scratch.data + ctx->scratch.offs; - - ctx->scratch.offs += data_size; - } else { - // allocate tensor data in the context's memory pool - obj_alloc_size = data_size; - } + // allocate tensor data in the context's memory pool + obj_alloc_size = data_size; } struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size); GGML_ASSERT(obj_new); - // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here - struct ggml_tensor * const result = (struct ggml_tensor *)((char *)ctx->mem_buffer + obj_new->offs); #ifdef __clang__ @@ -4211,24 +4128,16 @@ struct ggml_tensor * ggml_new_tensor_4d( } struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) { - ggml_scratch_save(ctx); - struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1); - ggml_scratch_load(ctx); - ggml_set_i32(result, value); return result; } struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) { - ggml_scratch_save(ctx); - struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1); - ggml_scratch_load(ctx); - ggml_set_f32(result, value); return result; @@ -7276,6 +7185,7 @@ struct ggml_tensor * ggml_ssm_conv( const int64_t n_s = sx->ne[2]; // TODO: maybe support other strides than 1? + // FIXME: this is always true? GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t); GGML_ASSERT(sx->ne[1] == d_inner); GGML_ASSERT(n_t >= 0); @@ -20295,7 +20205,6 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { uint64_t size_eval = 0; // compute size of intermediate results - // TODO: does not take into account scratch buffers !!!! for (int i = 0; i < cgraph->n_nodes; ++i) { size_eval += ggml_nbytes_pad(cgraph->nodes[i]); } @@ -22106,18 +22015,46 @@ static size_t gguf_type_size(enum gguf_type type) { return GGUF_TYPE_SIZE[type]; } -static void gguf_tensor_info_sanitize(struct gguf_tensor_info * info) { - GGML_ASSERT(info->n_dims <= GGML_MAX_DIMS); - GGML_ASSERT(0 <= info->type && info->type < GGML_TYPE_COUNT); +static bool gguf_tensor_info_sanitize(struct gguf_tensor_info * info) { + if (info->n_dims > GGML_MAX_DIMS) { + fprintf(stderr, "%s: invalid number of dimensions (%" PRIu32 ")\n", __func__, info->n_dims); + return false; + } + + if (info->type < 0 || info->type >= GGML_TYPE_COUNT) { + fprintf(stderr, "%s: invalid type (%d)\n", __func__, info->type); + return false; + } + + if (strlen(info->name.data) >= GGML_MAX_NAME) { + fprintf(stderr, "%s: tensor '%s' name is too long\n", __func__, info->name.data); + return false; + } for (uint32_t i = 0; i < info->n_dims; ++i) { - GGML_ASSERT(info->ne[i] > 0); + if (info->ne[i] <= 0) { + fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[i]); + return false; + } } // prevent overflow for total number of elements - GGML_ASSERT(INT64_MAX/info->ne[1] > info->ne[0]); - GGML_ASSERT(INT64_MAX/info->ne[2] > info->ne[0]*info->ne[1]); - GGML_ASSERT(INT64_MAX/info->ne[3] > info->ne[0]*info->ne[1]*info->ne[2]); + if (INT64_MAX/info->ne[1] <= info->ne[0]) { + fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[1]); + return false; + } + + if (INT64_MAX/info->ne[2] <= info->ne[0]*info->ne[1]) { + fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[2]); + return false; + } + + if (INT64_MAX/info->ne[3] <= info->ne[0]*info->ne[1]*info->ne[2]) { + fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[3]); + return false; + } + + return true; } static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) { @@ -22418,8 +22355,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ok = ok && gguf_fread_el (file, &info->type, sizeof(info->type), &offset); ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset), &offset); - // TODO: return an error instead of crashing with GGML_ASSERT - gguf_tensor_info_sanitize(info); + ok = ok && gguf_tensor_info_sanitize(info); // make sure there is no duplicated tensor names for (uint64_t j = 0; j < i && ok; ++j) { diff --git a/ggml/src/kompute-shaders/common.comp b/ggml/src/kompute-shaders/common.comp index 62d62b025..2aaddf704 100644 --- a/ggml/src/kompute-shaders/common.comp +++ b/ggml/src/kompute-shaders/common.comp @@ -15,6 +15,7 @@ #define TWOPI_F 6.283185307179586f #define QK_K 256 +#define K_SCALE_SIZE 12 #define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx]) #define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx) @@ -64,6 +65,14 @@ mat4 dequantize_q4_1(const block_q4_1 xb, uint il) { return reg; } +#define sizeof_block_q4_k 144 +struct block_q4_k { + float16_t d; + float16_t dmin; + uint8_t scales[K_SCALE_SIZE]; + uint8_t qs[QK_K/2]; +}; + #define sizeof_block_q6_k 210 struct block_q6_k { uint8_t ql[QK_K/2]; // quants, lower 4 bits diff --git a/ggml/src/kompute-shaders/op_mul_mat_q4_k.comp b/ggml/src/kompute-shaders/op_mul_mat_q4_k.comp new file mode 100644 index 000000000..fc8e45aa9 --- /dev/null +++ b/ggml/src/kompute-shaders/op_mul_mat_q4_k.comp @@ -0,0 +1,133 @@ +#version 450 + +#include "common.comp" + +#define N_DST 4 +#define SIZE_OF_BLOCK sizeof_block_q4_k + +layout(local_size_x = 4) in; +layout(local_size_y = 8) in; +layout(local_size_z = 1) in; + +layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; }; +layout (binding = 1) readonly buffer tensorInB { float inB[]; }; +layout (binding = 2) writeonly buffer tensorOut { float out_[]; }; + +layout (push_constant) uniform parameter { + uint inAOff; + uint inBOff; + uint outOff; + int ne00; + int ne10; + int ne0; + int ne1; + int ne01; + int ne02; + int ne12; + int r2; + int r3; +} pcs; + +void main() { + const uint16_t kmask1 = uint16_t(0x3f3f); + const uint16_t kmask2 = uint16_t(0x0f0f); + const uint16_t kmask3 = uint16_t(0xc0c0); + + const uint ix = gl_SubgroupInvocationID/8; // 0...3 + const uint it = gl_SubgroupInvocationID%8; // 0...7 + const uint iq = it/4; // 0 or 1 + const uint ir = it%4; // 0...3 + + const uint nb = pcs.ne00/QK_K; + + const uint r0 = gl_WorkGroupID.x; + const uint r1 = gl_WorkGroupID.y; + const uint im = gl_WorkGroupID.z; + + const uint first_row = r0 * N_DST; + const uint ib_row = first_row * nb; + + const uint i12 = im%pcs.ne12; + const uint i13 = im/pcs.ne12; + + const uint offset0 = (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02); + + const uint xblk = ib_row + offset0 + pcs.inAOff; + const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; + + float yl[16]; + float yh[16]; + float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f}; + float all_sum = 0.f; + + uint y4 = y + ix * QK_K + 64 * iq + 8 * ir; + + for (uint ib = ix; ib < nb; ib += 4) { + const uint blk_idx = ib + xblk; + + float sumy[4] = {0.f, 0.f, 0.f, 0.f}; + for (int i = 0; i < 8; ++i) { + yl[i+0] = inB[y4+i+ 0]; sumy[0] += yl[i+0]; + yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8]; + yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0]; + yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8]; + } + + for (int row = 0; row < N_DST; row++) { + uint row_idx = row * nb; + + uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0); + uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2); + uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4); + uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6); + uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8); + + uint16_t sc16[4]; + sc16[0] = sc_0 & kmask1; + sc16[1] = sc_2 & kmask1; + sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2); + sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2); + + float acc1[4] = {0.f, 0.f, 0.f, 0.f}; + float acc2[4] = {0.f, 0.f, 0.f, 0.f}; + for (int i = 0; i < 8; i += 2) { + uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i); + uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i); + acc1[0] += yl[i+0] * (q1 & 0x000F); + acc1[1] += yl[i+1] * (q1 & 0x0F00); + acc1[2] += yl[i+8] * (q1 & 0x00F0); + acc1[3] += yl[i+9] * (q1 & 0xF000); + acc2[0] += yh[i+0] * (q2 & 0x000F); + acc2[1] += yh[i+1] * (q2 & 0x0F00); + acc2[2] += yh[i+8] * (q2 & 0x00F0); + acc2[3] += yh[i+9] * (q2 & 0xF000); + } + + uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF); + uint8_t sc8_1 = uint8_t(sc16[0] >> 8 ); + uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF); + uint8_t sc8_3 = uint8_t(sc16[1] >> 8 ); + uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF); + uint8_t sc8_5 = uint8_t(sc16[2] >> 8 ); + uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF); + uint8_t sc8_7 = uint8_t(sc16[3] >> 8 ); + + float dall = float(inA[blk_idx + row_idx].d); + float dmin = float(inA[blk_idx + row_idx].dmin); + sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 + + (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f + + (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 + + (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) - + dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7); + } + + y4 += 4 * QK_K; + } + + for (int row = 0; row < N_DST; ++row) { + all_sum = subgroupAdd(sumf[row]); + if (subgroupElect()) { + out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum; + } + } +} diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index da40927e1..48863847c 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -162e232411ee98ceb0cccfa84886118d917d2123 +bb78a40dc60e04c626bac2b65840b509988e990d diff --git a/spm-headers/ggml-cpp.h b/spm-headers/ggml-cpp.h new file mode 120000 index 000000000..8a8604cc2 --- /dev/null +++ b/spm-headers/ggml-cpp.h @@ -0,0 +1 @@ +../ggml/include/ggml-cpp.h \ No newline at end of file diff --git a/src/llama.cpp b/src/llama.cpp index ef1b8ee59..3f534596e 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -7,6 +7,7 @@ #include "ggml.h" #include "ggml-alloc.h" #include "ggml-backend.h" +#include "ggml-cpp.h" // TODO: replace with ggml API call #define QK_K 256 @@ -2797,31 +2798,22 @@ struct llama_kv_cache { std::vector k_l; // per layer std::vector v_l; - std::vector ctxs; - std::vector bufs; + std::vector ctxs; + std::vector bufs; - size_t total_size() const { + size_t total_size() { size_t size = 0; - for (ggml_backend_buffer_t buf : bufs) { - size += ggml_backend_buffer_get_size(buf); + for (auto & buf : bufs) { + size += ggml_backend_buffer_get_size(buf.get()); } return size; } - - ~llama_kv_cache() { - for (struct ggml_context * ctx : ctxs) { - ggml_free(ctx); - } - for (ggml_backend_buffer_t buf : bufs) { - ggml_backend_buffer_free(buf); - } - } }; struct llama_control_vector { std::vector tensors; // per layer - std::vector ctxs; - std::vector bufs; + std::vector ctxs; + std::vector bufs; int32_t layer_start = -1; int32_t layer_end = -1; @@ -2840,15 +2832,6 @@ struct llama_control_vector { } return cur; } - - ~llama_control_vector() { - for (struct ggml_context * ctx : ctxs) { - ggml_free(ctx); - } - for (ggml_backend_buffer_t buf : bufs) { - ggml_backend_buffer_free(buf); - } - } }; struct llama_model { @@ -2908,10 +2891,10 @@ struct llama_model { std::vector dev_layer; // contexts where the model tensors metadata is stored - std::vector ctxs; + std::vector ctxs; // the model memory buffers for the tensor data - std::vector bufs; + std::vector bufs; // model memory mapped files llama_mmaps mappings; @@ -2930,13 +2913,7 @@ struct llama_model { std::set lora_adapters; ~llama_model() { - for (struct ggml_context * ctx : ctxs) { - ggml_free(ctx); - } - for (ggml_backend_buffer_t buf : bufs) { - ggml_backend_buffer_free(buf); - } - while (!lora_adapters.empty()) { + while (!lora_adapters.empty()) { llama_lora_adapter_free(*lora_adapters.begin()); } } @@ -3253,16 +3230,6 @@ struct llama_context { , t_start_us(model.t_start_us) , t_load_us(model.t_load_us) {} - ~llama_context() { - ggml_backend_sched_free(sched); - - for (ggml_backend_t backend : backends) { - ggml_backend_free(backend); - } - - ggml_backend_buffer_free(buf_output); - } - const struct llama_model & model; struct llama_cparams cparams; @@ -3272,7 +3239,7 @@ struct llama_context { std::unordered_map lora_adapters; - std::vector backends; + std::vector backends; std::vector> set_n_threads_fns; ggml_backend_t backend_cpu = nullptr; @@ -3294,7 +3261,7 @@ struct llama_context { mutable int32_t n_eval = 0; // number of eval calls // host buffer for the model output (logits and embeddings) - ggml_backend_buffer_t buf_output = nullptr; + ggml_backend_buffer_ptr buf_output; // decode output (2-dimensional array: [n_outputs][n_vocab]) size_t logits_size = 0; // capacity (of floats) for logits @@ -3324,7 +3291,7 @@ struct llama_context { // memory buffers used to evaluate the model std::vector buf_compute_meta; - ggml_backend_sched_t sched = nullptr; + ggml_backend_sched_ptr sched; ggml_abort_callback abort_callback = nullptr; void * abort_callback_data = nullptr; @@ -3358,8 +3325,8 @@ struct llama_lora_adapter { struct llama_model * base_model; // map tensor name to lora_a_b std::unordered_map ab_map; - std::vector ctxs; - std::vector bufs; + std::vector ctxs; + std::vector bufs; float alpha; @@ -3377,12 +3344,6 @@ struct llama_lora_adapter { } ~llama_lora_adapter() { - for (struct ggml_context * ctx : ctxs) { - ggml_free(ctx); - } - for (ggml_backend_buffer_t buf : bufs) { - ggml_backend_buffer_free(buf); - } auto pos = base_model->lora_adapters.find(this); if (pos != base_model->lora_adapters.end()) { base_model->lora_adapters.erase(pos); @@ -3401,24 +3362,21 @@ static bool buft_supported(ggml_backend_buffer_type_t buft, ggml_backend_dev_t d /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; - ggml_context * ctx = ggml_init(params); + ggml_context_ptr ctx { ggml_init(params) }; if (!ctx) { throw std::runtime_error(format("failed to create ggml context")); } - ggml_backend_buffer_t buf = ggml_backend_buft_alloc_buffer(buft, 0); - ggml_tensor * op_tensor = fn(ctx); + ggml_backend_buffer_ptr buf { ggml_backend_buft_alloc_buffer(buft, 0) }; + ggml_tensor * op_tensor = fn(ctx.get()); for (int i = 0; i < GGML_MAX_SRC; i++) { if (op_tensor->src[i] != nullptr) { assert(op_tensor->src[i]->buffer == nullptr); - op_tensor->src[i]->buffer = buf; + op_tensor->src[i]->buffer = buf.get(); } } bool op_supported = ggml_backend_dev_supports_op(dev, op_tensor); - ggml_free(ctx); - ggml_backend_buffer_free(buf); - return op_supported; } @@ -3470,7 +3428,8 @@ static bool llama_kv_cache_init( // create a context for each buffer type std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { - if (ctx_map.count(buft) == 0) { + auto it = ctx_map.find(buft); + if (it == ctx_map.end()) { struct ggml_init_params params = { /*.mem_size =*/ size_t(2u*n_layer*ggml_tensor_overhead()), /*.mem_buffer =*/ NULL, @@ -3481,9 +3440,10 @@ static bool llama_kv_cache_init( return nullptr; } ctx_map[buft] = ctx; - cache.ctxs.push_back(ctx); + cache.ctxs.emplace_back(ctx); + return ctx; } - return ctx_map.at(buft); + return it->second; }; cache.k_l.reserve(n_layer); @@ -3535,7 +3495,7 @@ static bool llama_kv_cache_init( } ggml_backend_buffer_clear(buf, 0); LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); - cache.bufs.push_back(buf); + cache.bufs.emplace_back(buf); } return true; @@ -3788,7 +3748,7 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { cache.used = 0; for (auto & buf : cache.bufs) { - ggml_backend_buffer_clear(buf, 0); + ggml_backend_buffer_clear(buf.get(), 0); } } @@ -4271,21 +4231,38 @@ struct llama_model_loader { ggml_tensor * tensor; - llama_tensor_weight(const llama_file * file, uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { - const int tensor_idx = gguf_find_tensor(gguf_ctx, name); - offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx); + llama_tensor_weight(const llama_file * file, uint16_t idx, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { + const int tensor_idx = gguf_find_tensor(gguf_ctx, ggml_get_name(tensor)); + if (tensor_idx < 0) { + throw std::runtime_error(format("tensor '%s' not found in the model", ggml_get_name(tensor))); + } + offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx); if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size) { - throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", name)); + throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", ggml_get_name(tensor))); } } }; - std::vector weights; + // custom comparator to sort weights more nicely by layer + struct weight_name_comparer { + bool operator()(const std::string & a, const std::string & b) const { + int a_layer = -1; + int b_layer = -1; + sscanf(a.c_str(), "blk.%d.", &a_layer); + sscanf(b.c_str(), "blk.%d.", &b_layer); + if (a_layer != b_layer) { + return a_layer < b_layer; + } + return a < b; + } + }; + + std::map weights_map; std::unordered_map kv_overrides; - struct gguf_context * meta = NULL; - std::vector contexts; + gguf_context_ptr meta; + std::vector contexts; std::string arch_name; LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN); @@ -4308,7 +4285,7 @@ struct llama_model_loader { /*.ctx = */ &ctx, }; - meta = gguf_init_from_file(fname.c_str(), params); + meta.reset(gguf_init_from_file(fname.c_str(), params)); if (!meta) { throw std::runtime_error(format("%s: failed to load model from %s\n", __func__, fname.c_str())); } @@ -4323,7 +4300,14 @@ struct llama_model_loader { // For subsidiary files, `meta` tensor data offset must not be used, // so we build a unified tensors index for weights. for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { - weights.emplace_back(files.back().get(), 0, cur->name, meta, cur); + std::string tensor_name = std::string(cur->name); + // make sure there is no duplicated tensor names + if (weights_map.find(tensor_name) != weights_map.end()) { + throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur))); + } + n_elements += ggml_nelements(cur); + n_bytes += ggml_nbytes(cur); + weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), 0, meta.get(), cur)); } uint16_t n_split = 0; get_key(llm_kv(LLM_KV_SPLIT_COUNT), n_split, false); @@ -4353,7 +4337,7 @@ struct llama_model_loader { /*.no_alloc = */ true, /*.ctx = */ &ctx, }; - struct gguf_context * ctx_gguf = gguf_init_from_file(split_path, split_params); + gguf_context_ptr ctx_gguf { gguf_init_from_file(split_path, split_params) }; if (!ctx_gguf) { throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path)); } @@ -4363,17 +4347,22 @@ struct llama_model_loader { // Save tensors data offset info of the shard. for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { - weights.emplace_back(files.back().get(), idx, cur->name, ctx_gguf, cur); + std::string tensor_name = std::string(cur->name); + // make sure there is no duplicated tensor names + if (weights_map.find(tensor_name) != weights_map.end()) { + throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur))); + } + n_elements += ggml_nelements(cur); + n_bytes += ggml_nbytes(cur); + weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), idx, ctx_gguf.get(), cur)); } - - gguf_free(ctx_gguf); } get_key(llm_kv(LLM_KV_SPLIT_TENSORS_COUNT), n_tensors); // sanity check { - const int n_tensors_loaded = (int) weights.size(); + const int n_tensors_loaded = (int) weights_map.size(); if (n_tensors != n_tensors_loaded) { throw std::runtime_error(format("corrupted model: %d tensors expected but %d found", n_tensors, n_tensors_loaded)); } @@ -4382,23 +4371,10 @@ struct llama_model_loader { LLAMA_LOG_INFO("%s: additional %d GGUFs metadata loaded.\n", __func__, n_split - 1); } - n_kv = gguf_get_n_kv(meta); - n_tensors = weights.size(); + n_kv = gguf_get_n_kv(meta.get()); + n_tensors = weights_map.size(); - fver = (enum llama_fver) gguf_get_version(meta); - - std::set tensor_names; - for (auto & w : weights) { - n_elements += ggml_nelements(w.tensor); - n_bytes += ggml_nbytes(w.tensor); - // make sure there is no duplicated tensor names - const std::string name(w.tensor->name); - auto found = tensor_names.find(name); - if (found != tensor_names.end()) { - throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", w.tensor->name)); - } - tensor_names.insert(name); - } + fver = (enum llama_fver) gguf_get_version(meta.get()); LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", __func__, n_kv, n_tensors, fname.c_str(), llama_file_version_name(fver)); @@ -4411,8 +4387,10 @@ struct llama_model_loader { uint32_t n_type_max = 0; enum ggml_type type_max = GGML_TYPE_F32; - for (int i = 0; i < n_tensors; i++) { - const ggml_tensor * tensor = weights.at(i).tensor; + for (const auto & it : weights_map) { + const llama_tensor_weight & w = it.second; + const ggml_tensor * tensor = w.tensor; + enum ggml_type type = tensor->type; n_type[type]++; @@ -4423,8 +4401,8 @@ struct llama_model_loader { } if (trace > 0) { - const uint16_t sid = weights.at(i).idx; - LLAMA_LOG_INFO("%s: - tensor %4d, split %2d: %32s %-8s [ %s ]\n", __func__, i, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str()); + const uint16_t sid = w.idx; + LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ]\n", __func__, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str()); } } @@ -4467,23 +4445,23 @@ struct llama_model_loader { ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED); { - const int kid = gguf_find_key(meta, "general.file_type"); // TODO: use LLM_KV + const int kid = gguf_find_key(meta.get(), "general.file_type"); // TODO: use LLM_KV if (kid >= 0) { - ftype = (llama_ftype) gguf_get_val_u32(meta, kid); + ftype = (llama_ftype) gguf_get_val_u32(meta.get(), kid); } } LLAMA_LOG_INFO("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); for (int i = 0; i < n_kv; i++) { - const char * name = gguf_get_key(meta, i); - const enum gguf_type type = gguf_get_kv_type(meta, i); + const char * name = gguf_get_key(meta.get(), i); + const enum gguf_type type = gguf_get_kv_type(meta.get(), i); const std::string type_name = type == GGUF_TYPE_ARRAY - ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta, i)), gguf_get_arr_n(meta, i)) + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta.get(), i)), gguf_get_arr_n(meta.get(), i)) : gguf_type_name(type); - std::string value = gguf_kv_to_str(meta, i); + std::string value = gguf_kv_to_str(meta.get(), i); const size_t MAX_VALUE_LEN = 40; if (value.size() > MAX_VALUE_LEN) { value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); @@ -4512,19 +4490,10 @@ struct llama_model_loader { this->check_tensors = check_tensors; } - ~llama_model_loader() { - if (meta) { - gguf_free(meta); - } - for (auto * ctx : contexts) { - ggml_free(ctx); - } - } - template typename std::enable_if::value, bool>::type get_arr_n(const std::string & key, T & result, const bool required = true) { - const int kid = gguf_find_key(meta, key.c_str()); + const int kid = gguf_find_key(meta.get(), key.c_str()); if (kid < 0) { if (required) { @@ -4534,7 +4503,7 @@ struct llama_model_loader { } struct GGUFMeta::ArrayInfo arr_info = - GGUFMeta::GKV::get_kv(meta, kid); + GGUFMeta::GKV::get_kv(meta.get(), kid); result = arr_info.length; @@ -4549,9 +4518,9 @@ struct llama_model_loader { template bool get_arr(const std::string & key, std::vector & result, const bool required = true) { - const int kid = gguf_find_key(meta, key.c_str()); + const int kid = gguf_find_key(meta.get(), key.c_str()); - if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { + if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) { if (required) { throw std::runtime_error(format("array key not found in model: %s", key.c_str())); } @@ -4559,7 +4528,7 @@ struct llama_model_loader { } struct GGUFMeta::ArrayInfo arr_info = - GGUFMeta::GKV::get_kv(meta, kid); + GGUFMeta::GKV::get_kv(meta.get(), kid); switch (arr_info.gt) { case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; @@ -4578,9 +4547,9 @@ struct llama_model_loader { template bool get_arr(const std::string & key, std::array & result, const bool required = true) { - const int kid = gguf_find_key(meta, key.c_str()); + const int kid = gguf_find_key(meta.get(), key.c_str()); - if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { + if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) { if (required) { throw std::runtime_error(format("array key not found in model: %s", key.c_str())); } @@ -4588,7 +4557,7 @@ struct llama_model_loader { } struct GGUFMeta::ArrayInfo arr_info = - GGUFMeta::GKV::get_kv(meta, kid); + GGUFMeta::GKV::get_kv(meta.get(), kid); switch (arr_info.gt) { case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; @@ -4620,7 +4589,7 @@ struct llama_model_loader { const struct llama_model_kv_override * override = it != kv_overrides.end() ? &it->second : nullptr; - const bool found = GGUFMeta::GKV::set(meta, key, result, override); + const bool found = GGUFMeta::GKV::set(meta.get(), key, result, override); if (required && !found) { throw std::runtime_error(format("key not found in model: %s", key.c_str())); @@ -4637,7 +4606,7 @@ struct llama_model_loader { // get array of n <= N_MAX elements, or a single element repeated n times template bool get_key_or_arr(const std::string & key, std::array & result, uint32_t n, const bool required = true) { - const int kid = gguf_find_key(meta, key.c_str()); + const int kid = gguf_find_key(meta.get(), key.c_str()); if (kid < 0) { if (required) { @@ -4650,9 +4619,9 @@ struct llama_model_loader { throw std::runtime_error(format("n > N_MAX: %u > %u for key %s", (uint32_t) n, (uint32_t) N_MAX, key.c_str())); } - if (gguf_get_kv_type(meta, kid) == GGUF_TYPE_ARRAY) { + if (gguf_get_kv_type(meta.get(), kid) == GGUF_TYPE_ARRAY) { struct GGUFMeta::ArrayInfo arr_info = - GGUFMeta::GKV::get_kv(meta, kid); + GGUFMeta::GKV::get_kv(meta.get(), kid); if (n != arr_info.length) { throw std::runtime_error(format("key %s has wrong array length; expected %u, got %u", key.c_str(), n, (uint32_t) arr_info.length)); @@ -4688,21 +4657,13 @@ struct llama_model_loader { return llm_kv.arch; } - const char * get_tensor_name(int i) const { - return weights.at(i).tensor->name; - } - const llama_tensor_weight * get_weight(const char * name) const { - for (const auto & weight : weights) { - if (strcmp(name, weight.tensor->name) == 0) { - return &weight; - } + auto pos = weights_map.find(name); + if (pos != weights_map.end()) { + return &pos->second; } - return nullptr; - } - const llama_tensor_weight * get_weight(int i) const { - return get_weight(get_tensor_name(i)); + return nullptr; } const llama_tensor_weight & require_weight(const char * name) const { @@ -4729,10 +4690,6 @@ struct llama_model_loader { return tensor; } - struct ggml_tensor * get_tensor_meta(int i) const { - return get_tensor_meta(get_tensor_name(i)); - } - const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector & ne, bool required) const { const struct ggml_tensor * cur = get_tensor_meta(name.c_str()); @@ -4839,8 +4796,8 @@ struct llama_model_loader { } // compute the total size of all tensors for progress reporting - for (auto & w : weights) { - size_data += ggml_nbytes(w.tensor); + for (const auto & it : weights_map) { + size_data += ggml_nbytes(it.second.tensor); } } @@ -4852,19 +4809,12 @@ struct llama_model_loader { *last = 0; *addr = mapping->addr; for (ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor; tensor = ggml_get_next_tensor(ctx, tensor)) { - try { - const auto * weight = get_weight(ggml_get_name(tensor)); - if (!weight) { - continue; - } - if (weight->idx != idx) { - continue; - } - *first = std::min(*first, weight->offs); - *last = std::max(*last, weight->offs + ggml_nbytes(tensor)); - } catch(...) { - // the tensor is not in the model + const auto * weight = get_weight(ggml_get_name(tensor)); + if (!weight || weight->idx != idx) { + continue; } + *first = std::min(*first, weight->offs); + *last = std::max(*last, weight->offs + ggml_nbytes(tensor)); } } @@ -5041,7 +4991,6 @@ struct llama_model_loader { ggml_backend_tensor_set(cur, data, 0, n_size); } } else { - GGML_ASSERT(weight->idx < files.size()); const auto & file = files.at(weight->idx); if (ggml_backend_buffer_is_host(cur->buffer)) { file->seek(weight->offs, SEEK_SET); @@ -5342,7 +5291,7 @@ static void llm_load_hparams( llama_model_loader & ml, llama_model & model) { auto & hparams = model.hparams; - const gguf_context * ctx = ml.meta; + const gguf_context * ctx = ml.meta.get(); // get metadata as string for (int i = 0; i < gguf_get_n_kv(ctx); i++) { @@ -6109,7 +6058,7 @@ static void llm_load_vocab( llama_model & model) { auto & vocab = model.vocab; - struct gguf_context * ctx = ml.meta; + struct gguf_context * ctx = ml.meta.get(); const auto kv = LLM_KV(model.arch); @@ -7104,10 +7053,11 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; - ggml_context * ctx = ggml_init(params); - if (!ctx) { + ggml_context_ptr ctx_ptr { ggml_init(params) }; + if (!ctx_ptr) { throw std::runtime_error(format("failed to create ggml context")); } + ggml_context * ctx = ctx_ptr.get(); ggml_tensor * op_tensor = nullptr; @@ -7119,7 +7069,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w } break; case GGML_OP_MUL_MAT: { - ggml_tensor * b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512); + ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], 512, w->ne[2], w->ne[3]); op_tensor = ggml_mul_mat(ctx, w, b); } break; case GGML_OP_MUL_MAT_ID: @@ -7159,18 +7109,38 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w } break; case GGML_OP_SSM_CONV: { - // TODO: ggml_ssm_conv(ctx, conv_x, model.layers[il].ssm_conv1d); - op_tensor = ggml_ssm_conv(ctx, nullptr, w); + // FIXME + ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 12345, w->ne[1], 6789); + op_tensor = ggml_ssm_conv(ctx, conv_x, w); } break; case GGML_OP_SSM_SCAN: { - // TODO: ggml_ssm_scan(ctx, ssm, x, dt, model.layers[il].ssm_a, B, C); - op_tensor = ggml_ssm_scan(ctx, nullptr, nullptr, nullptr, w, nullptr, nullptr); + // FIXME + const int64_t d_state = w->ne[0]; + const int64_t d_inner = w->ne[1]; + const int64_t n_seq_tokens = 512; + const int64_t n_seqs = 1; + ggml_tensor * s = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, d_inner, n_seqs); + ggml_tensor * x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs); + ggml_tensor * dt = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs); + ggml_tensor * B = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs); + ggml_tensor * C = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs); + op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C); } break; case GGML_OP_RWKV_WKV: { - // TODO: ggml_rwkv_wkv(ctx, k, v, r, layer->time_mix_first, w, *wkv_state); - op_tensor = ggml_rwkv_wkv(ctx, nullptr, nullptr, nullptr, w, nullptr, nullptr); + // FIXME + const int64_t S = 123; + const int64_t H = 123; + const int64_t n_tokens = 123; + const int64_t n_seqs = 123; + ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, S, 1, H, n_tokens); + ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, 1, S, H, n_tokens); + ggml_tensor * r = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, 1, S, H, n_tokens); + ggml_tensor * tf = w; + ggml_tensor * td = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, 1, S, H, n_tokens); + ggml_tensor * state = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, S, n_seqs, S, H); + op_tensor = ggml_rwkv_wkv(ctx, k, v, r, tf, td, state); } break; default: GGML_ABORT("%s: missing test for op %s for tensor %s", __func__, ggml_op_name(op), w->name); @@ -7183,8 +7153,6 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w ggml_backend_buffer_free(w->buffer); w->buffer = nullptr; - ggml_free(ctx); - return op_supported; } @@ -7375,7 +7343,8 @@ static bool llm_load_tensors( std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { - if (ctx_map.count(buft) == 0) { + auto it = ctx_map.find(buft); + if (it == ctx_map.end()) { ggml_init_params params = { /*.mem_size =*/ ctx_size, /*.mem_buffer =*/ NULL, @@ -7386,9 +7355,10 @@ static bool llm_load_tensors( throw std::runtime_error(format("failed to create ggml context")); } ctx_map[buft] = ctx; - model.ctxs.push_back(ctx); + model.ctxs.emplace_back(ctx); + return ctx; } - return ctx_map.at(buft); + return it->second; }; // create tensors for the weights @@ -7426,7 +7396,7 @@ static bool llm_load_tensors( if (flags & llama_model_loader::TENSOR_NOT_REQUIRED) { return nullptr; } - throw std::runtime_error(format("missing tensor %s", tn.str().c_str())); + throw std::runtime_error(format("missing tensor '%s'", tn.str().c_str())); } // some models use the token embedding tensor as the output, but since these are used in different layers and with different ops @@ -7445,7 +7415,7 @@ static bool llm_load_tensors( // tensors with "bias" suffix are always used with GGML_OP_ADD ggml_op op; - bool bias = strcmp(tn.suffix, "bias") == 0; + bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0; if (bias) { op = GGML_OP_ADD; } else { @@ -9114,7 +9084,7 @@ static bool llm_load_tensors( if (buf == nullptr) { throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } - model.bufs.push_back(buf); + model.bufs.emplace_back(buf); bufs.emplace(idx, buf); } } @@ -9123,7 +9093,7 @@ static bool llm_load_tensors( if (buf == nullptr) { throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } - model.bufs.push_back(buf); + model.bufs.emplace_back(buf); if (use_mlock && ggml_backend_buffer_is_host(buf)) { model.mlock_bufs.emplace_back(new llama_mlock); auto & mlock_buf = model.mlock_bufs.back(); @@ -9163,13 +9133,13 @@ static bool llm_load_tensors( } // print memory requirements per buffer type - for (ggml_backend_buffer_t buf : model.bufs) { - LLAMA_LOG_INFO("%s: %10s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf) / 1024.0 / 1024.0); + for (auto & buf : model.bufs) { + LLAMA_LOG_INFO("%s: %10s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0); } // populate tensors_by_name - for (ggml_context * ctx : model.ctxs) { - for (auto * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) { + for (auto & ctx : model.ctxs) { + for (auto * cur = ggml_get_first_tensor(ctx.get()); cur != NULL; cur = ggml_get_next_tensor(ctx.get(), cur)) { model.tensors_by_name.emplace_back(ggml_get_name(cur), cur); } } @@ -10274,10 +10244,8 @@ struct llm_build_context { } void free() { - if (ctx0) { - ggml_free(ctx0); - ctx0 = nullptr; - } + ggml_free(ctx0); + ctx0 = nullptr; } struct ggml_cgraph * build_k_shift() { @@ -10305,10 +10273,10 @@ struct llm_build_context { // dequantize to f32 -> RoPE -> quantize back tmp = ggml_cast(ctx0, k, GGML_TYPE_F32); cb(tmp, "K_f32", il); - for (auto * backend : lctx.backends) { + for (auto & backend : lctx.backends) { // Figure out which backend KV cache belongs to - if (ggml_backend_supports_buft(backend, ggml_backend_buffer_get_type(kv_self.k_l[il]->buffer))) { - ggml_backend_sched_set_tensor_backend(lctx.sched, tmp, backend); + if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(kv_self.k_l[il]->buffer))) { + ggml_backend_sched_set_tensor_backend(lctx.sched.get(), tmp, backend.get()); break; } } @@ -16423,7 +16391,7 @@ static struct ggml_cgraph * llama_build_graph( if (!lctx.cparams.offload_kqv) { if (strcmp(name, "kqv_merged_cont") == 0) { // all nodes between the KV store and the attention output are run on the CPU - ggml_backend_sched_set_tensor_backend(lctx.sched, cur, lctx.backend_cpu); + ggml_backend_sched_set_tensor_backend(lctx.sched.get(), cur, lctx.backend_cpu); } } @@ -16433,10 +16401,10 @@ static struct ggml_cgraph * llama_build_graph( if (ubatch.n_tokens < 32 || full_offload) { if (il != -1 && strcmp(name, "norm") == 0) { const auto & dev_layer = lctx.model.dev_layer.at(il); - for (auto * backend : lctx.backends) { - if (ggml_backend_get_device(backend) == dev_layer.dev) { - if (ggml_backend_supports_op(backend, cur)) { - ggml_backend_sched_set_tensor_backend(lctx.sched, cur, backend); + for (auto & backend : lctx.backends) { + if (ggml_backend_get_device(backend.get()) == dev_layer.dev) { + if (ggml_backend_supports_op(backend.get(), cur)) { + ggml_backend_sched_set_tensor_backend(lctx.sched.get(), cur, backend.get()); } } } @@ -17123,7 +17091,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { lctx.output_ids.resize(n_batch); } - const size_t prev_size = lctx.buf_output ? ggml_backend_buffer_get_size(lctx.buf_output) : 0; + const size_t prev_size = lctx.buf_output ? ggml_backend_buffer_get_size(lctx.buf_output.get()) : 0; const size_t new_size = (logits_size + embd_size) * sizeof(float); // alloc only when more than the current capacity is required @@ -17134,7 +17102,6 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark) LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0); #endif - ggml_backend_buffer_free(lctx.buf_output); lctx.buf_output = nullptr; lctx.logits = nullptr; lctx.embd = nullptr; @@ -17142,27 +17109,19 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { auto * buft = ggml_backend_cpu_buffer_type(); // try to use the host buffer of the device where the output tensor is allocated for faster transfer to system memory - ggml_tensor * output_tensor = lctx.model.output; - if (!output_tensor) { - // bert models don't have an output tensor, use the last layer - output_tensor = lctx.model.layers.back().layer_out_norm; + auto * output_dev = lctx.model.dev_output.dev; + auto * output_dev_host_buft = output_dev ? ggml_backend_dev_host_buffer_type(output_dev) : nullptr; + if (output_dev_host_buft) { + buft = output_dev_host_buft; } - if (output_tensor) { - auto * output_buft = ggml_backend_buffer_get_type(output_tensor->buffer); - auto * output_dev = ggml_backend_buft_get_device(output_buft); - auto * output_dev_host_buft = ggml_backend_dev_host_buffer_type(output_dev); - if (output_dev_host_buft) { - buft = output_dev_host_buft; - } - } - lctx.buf_output = ggml_backend_buft_alloc_buffer(buft, new_size); + lctx.buf_output.reset(ggml_backend_buft_alloc_buffer(buft, new_size)); if (lctx.buf_output == nullptr) { LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0)); return 0; } } - float * output_base = (float *) ggml_backend_buffer_get_base(lctx.buf_output); + float * output_base = (float *) ggml_backend_buffer_get_base(lctx.buf_output.get()); lctx.logits = has_logits ? output_base : nullptr; lctx.embd = has_embd ? output_base + logits_size : nullptr; @@ -17174,7 +17133,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { // set all ids as invalid (negative) std::fill(lctx.output_ids.begin(), lctx.output_ids.end(), -1); - ggml_backend_buffer_clear(lctx.buf_output, 0); + ggml_backend_buffer_clear(lctx.buf_output.get(), 0); lctx.n_outputs = 0; @@ -17234,7 +17193,7 @@ static void llama_graph_compute( set_n_threads_fn.second(set_n_threads_fn.first, n_threads); } - auto err = ggml_backend_sched_graph_compute_async(lctx.sched, gf); + auto err = ggml_backend_sched_graph_compute_async(lctx.sched.get(), gf); if (err != GGML_STATUS_SUCCESS) { LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, err); } @@ -17392,8 +17351,8 @@ static int llama_decode_internal( //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); - ggml_backend_sched_reset(lctx.sched); - ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); + ggml_backend_sched_reset(lctx.sched.get()); + ggml_backend_sched_set_eval_callback(lctx.sched.get(), lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false); @@ -17421,7 +17380,7 @@ static int llama_decode_internal( } // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); - ggml_backend_sched_alloc_graph(lctx.sched, gf); + ggml_backend_sched_alloc_graph(lctx.sched.get(), gf); llama_set_inputs(lctx, ubatch); @@ -17444,7 +17403,7 @@ static int llama_decode_internal( // extract logits if (res) { - ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched, res); + ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), res); GGML_ASSERT(backend_res != nullptr); GGML_ASSERT(lctx.logits != nullptr); @@ -17460,7 +17419,7 @@ static int llama_decode_internal( // extract embeddings if (embd) { - ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd); + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), embd); GGML_ASSERT(backend_embd != nullptr); switch (cparams.pooling_type) { @@ -17555,7 +17514,7 @@ static int llama_decode_internal( // Reset state for the next token before backend sync, to allow the CPU activities in the reset to // overlap with device computation. - ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_reset(lctx.sched.get()); return 0; } @@ -17633,8 +17592,8 @@ static int llama_encode_internal( GGML_ASSERT(n_threads > 0); - ggml_backend_sched_reset(lctx.sched); - ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); + ggml_backend_sched_reset(lctx.sched.get()); + ggml_backend_sched_set_eval_callback(lctx.sched.get(), lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false); @@ -17658,7 +17617,7 @@ static int llama_encode_internal( } } - ggml_backend_sched_alloc_graph(lctx.sched, gf); + ggml_backend_sched_alloc_graph(lctx.sched.get(), gf); llama_set_inputs(lctx, ubatch); @@ -17666,7 +17625,7 @@ static int llama_encode_internal( // extract embeddings if (embd) { - ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd); + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), embd); GGML_ASSERT(backend_embd != nullptr); if (llama_model_has_decoder(&lctx.model)) { @@ -17733,7 +17692,7 @@ static int llama_encode_internal( // Reset state for the next token before backend sync, to allow the CPU activities in the reset to // overlap with device computation. - ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_reset(lctx.sched.get()); return 0; } @@ -17947,7 +17906,7 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { #else // ggml_graph defrag - ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_reset(lctx.sched.get()); ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids); @@ -17969,11 +17928,11 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { } { - ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_reset(lctx.sched.get()); ggml_cgraph * gf = llama_build_graph_k_shift(lctx); - ggml_backend_sched_alloc_graph(lctx.sched, gf); + ggml_backend_sched_alloc_graph(lctx.sched.get(), gf); llama_set_k_shift(lctx); @@ -18013,8 +17972,8 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { ggml_cgraph * gf = llama_build_graph(lctx, ubatch, true); // initialize scheduler with the worst-case graph - ggml_backend_sched_reset(lctx.sched); - if (!ggml_backend_sched_reserve(lctx.sched, gf)) { + ggml_backend_sched_reset(lctx.sched.get()); + if (!ggml_backend_sched_reserve(lctx.sched.get(), gf)) { LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__); } } @@ -18565,40 +18524,57 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } const size_t align = GGUF_DEFAULT_ALIGNMENT; - struct gguf_context * ctx_out = gguf_init_empty(); + gguf_context_ptr ctx_out { gguf_init_empty() }; // copy the KV pairs from the input file - gguf_set_kv (ctx_out, ml.meta); - gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV - gguf_set_val_u32(ctx_out, "general.file_type", ftype); // TODO: use LLM_KV + gguf_set_kv (ctx_out.get(), ml.meta.get()); + gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV + gguf_set_val_u32(ctx_out.get(), "general.file_type", ftype); // TODO: use LLM_KV // Remove split metadata - gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_NO).c_str()); - gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str()); - gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str()); + gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str()); + gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str()); + gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str()); if (params->kv_overrides) { const std::vector & overrides = *(const std::vector *)params->kv_overrides; - for (auto & o : overrides) { + for (const auto & o : overrides) { if (o.key[0] == 0) break; if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) { - gguf_set_val_f32(ctx_out, o.key, o.val_f64); + gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) { - gguf_set_val_i32(ctx_out, o.key, o.val_i64); + gguf_set_val_i32(ctx_out.get(), o.key, o.val_i64); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) { - gguf_set_val_bool(ctx_out, o.key, o.val_bool); + gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) { - gguf_set_val_str(ctx_out, o.key, o.val_str); + gguf_set_val_str(ctx_out.get(), o.key, o.val_str); } else { LLAMA_LOG_WARN("%s: unknown KV override type for key %s\n", __func__, o.key); } } } - for (int i = 0; i < ml.n_tensors; ++i) { - const struct ggml_tensor * meta = ml.get_tensor_meta(i); + // make a list of weights + std::vector tensors; + tensors.reserve(ml.weights_map.size()); + for (const auto & it : ml.weights_map) { + tensors.push_back(&it.second); + } - const std::string name = ggml_get_name(meta); + // keep_split requires that the weights are sorted by split index + if (params->keep_split) { + std::sort(tensors.begin(), tensors.end(), [](const llama_model_loader::llama_tensor_weight * a, const llama_model_loader::llama_tensor_weight * b) { + if (a->idx == b->idx) { + return a->offs < b->offs; + } + return a->idx < b->idx; + }); + } + + for (const auto * it : tensors) { + const struct ggml_tensor * tensor = it->tensor; + + const std::string name = ggml_get_name(tensor); // TODO: avoid hardcoded tensor names - use the TN_* constants if (name.find("attn_v.weight") != std::string::npos || @@ -18636,32 +18612,32 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s std::vector> f32_conv_buf; uint16_t n_split = 1; + // Assume split index is continuous if (params->keep_split) { - for (int i = 0; i < ml.n_tensors; ++i) { - n_split = std::max(uint16_t(ml.get_weight(i)->idx+1), n_split); + for (const auto * it : tensors) { + n_split = std::max(uint16_t(it->idx + 1), n_split); } } - std::vector ctx_outs(n_split, NULL); - ctx_outs[0] = ctx_out; + std::vector ctx_outs(n_split); + ctx_outs[0] = std::move(ctx_out); // populate the original tensors so we get an initial meta data - for (int i = 0; i < ml.n_tensors; ++i) { - auto weight = ml.get_weight(i); - uint16_t i_split = params->keep_split ? weight->idx : 0; - struct ggml_tensor * tensor = weight->tensor; - if (ctx_outs[i_split] == NULL) { - ctx_outs[i_split] = gguf_init_empty(); + for (const auto * it : tensors) { + uint16_t i_split = params->keep_split ? it->idx : 0; + struct ggml_tensor * tensor = it->tensor; + if (!ctx_outs[i_split]) { + ctx_outs[i_split].reset(gguf_init_empty()); } - gguf_add_tensor(ctx_outs[i_split], tensor); + gguf_add_tensor(ctx_outs[i_split].get(), tensor); } // Set split info if needed if (n_split > 1) { for (size_t i = 0; i < ctx_outs.size(); ++i) { - gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i); - gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split); - gguf_set_val_i32(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors); + gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i); + gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split); + gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors); } } @@ -18671,8 +18647,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // Write metadata and close file handler if (fout.is_open()) { fout.seekp(0); - std::vector data(gguf_get_meta_size(ctx_outs[cur_split])); - gguf_get_meta_data(ctx_outs[cur_split], data.data()); + std::vector data(gguf_get_meta_size(ctx_outs[cur_split].get())); + gguf_get_meta_data(ctx_outs[cur_split].get(), data.data()); fout.write((const char *) data.data(), data.size()); fout.close(); } @@ -18689,19 +18665,19 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s fout = std::ofstream(fname, std::ios::binary); fout.exceptions(std::ofstream::failbit); // fail fast on write errors - const size_t meta_size = gguf_get_meta_size(ctx_outs[cur_split]); + const size_t meta_size = gguf_get_meta_size(ctx_outs[cur_split].get()); // placeholder for the meta data ::zeros(fout, meta_size); }; const auto tn = LLM_TN(model.arch); new_ofstream(0); - for (int i = 0; i < ml.n_tensors; ++i) { - auto weight = ml.get_weight(i); - struct ggml_tensor * tensor = weight->tensor; - if (weight->idx != cur_split && params->keep_split) { + for (const auto * it : tensors) { + const auto & weight = *it; + struct ggml_tensor * tensor = weight.tensor; + if (weight.idx != cur_split && params->keep_split) { close_ofstream(); - new_ofstream(weight->idx); + new_ofstream(weight.idx); } const std::string name = ggml_get_name(tensor); @@ -18874,17 +18850,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s total_size_new += new_size; // update the gguf meta data as we go - gguf_set_tensor_type(ctx_outs[cur_split], name.c_str(), new_type); - gguf_set_tensor_data(ctx_outs[cur_split], name.c_str(), new_data, new_size); + gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type); + gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data, new_size); // write tensor data + padding fout.write((const char *) new_data, new_size); zeros(fout, GGML_PAD(new_size, align) - new_size); } close_ofstream(); - for (auto & c:ctx_outs) { - gguf_free(c); - } LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); @@ -18898,51 +18871,51 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s static void llama_lora_adapter_init_internal(struct llama_model * model, const char * path_lora, struct llama_lora_adapter & adapter) { LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora); - ggml_context * ctx = nullptr; + ggml_context * ctx_init; struct gguf_init_params meta_gguf_params = { /* .no_alloc = */ true, - /* .ctx = */ &ctx, + /* .ctx = */ &ctx_init, }; - struct gguf_context * ctx_gguf = gguf_init_from_file(path_lora, meta_gguf_params); + + gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) }; if (!ctx_gguf) { throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora)); } + ggml_context_ptr ctx { ctx_init }; + // check metadata { auto get_kv_str = [&](const std::string & key) -> std::string { - int id = gguf_find_key(ctx_gguf, key.c_str()); - return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id)); + int id = gguf_find_key(ctx_gguf.get(), key.c_str()); + return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id)); }; auto get_kv_f32 = [&](const std::string & key) -> float { - int id = gguf_find_key(ctx_gguf, key.c_str()); - return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id); + int id = gguf_find_key(ctx_gguf.get(), key.c_str()); + return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id); }; LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN); auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE)); if (general_type != "adapter") { - gguf_free(ctx_gguf); throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type); } auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE)); auto general_arch = llm_arch_from_string(general_arch_str); if (general_arch != model->arch) { - gguf_free(ctx_gguf); throw std::runtime_error("model arch and LoRA arch mismatch"); } auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE)); if (adapter_type != "lora") { - gguf_free(ctx_gguf); throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type); } adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA)); } - int n_tensors = gguf_get_n_tensors(ctx_gguf); + int n_tensors = gguf_get_n_tensors(ctx_gguf.get()); // contexts for each buffer type std::map ctx_map; @@ -18956,7 +18929,11 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c /*.no_alloc =*/ true, }; ggml_context * buft_ctx = ggml_init(params); + if (!buft_ctx) { + return nullptr; + } ctx_map[buft] = buft_ctx; + adapter.ctxs.emplace_back(buft_ctx); return buft_ctx; }; return it->second; @@ -18967,7 +18944,7 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c auto str_endswith = [](const std::string & str, const std::string & suffix) { return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0; }; - for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { + for (ggml_tensor * cur = ggml_get_first_tensor(ctx.get()); cur; cur = ggml_get_next_tensor(ctx.get(), cur)) { std::string name(cur->name); if (str_endswith(name, ".lora_a")) { replace_all(name, ".lora_a", ""); @@ -18984,8 +18961,6 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c ab_map[name].b = cur; } } else { - gguf_free(ctx_gguf); - ggml_free(ctx); throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix"); } } @@ -18996,28 +18971,20 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c llama_lora_weight & w = it.second; if (!w.a || !w.b) { - gguf_free(ctx_gguf); - ggml_free(ctx); throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component"); } // device buft and device ctx auto * model_tensor = llama_get_model_tensor(model, name.c_str()); if (!model_tensor) { - gguf_free(ctx_gguf); - ggml_free(ctx); throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model"); } struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer)); // validate tensor shape if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) { - gguf_free(ctx_gguf); - ggml_free(ctx); throw std::runtime_error("tensor '" + name + "' has incorrect shape"); } if (w.a->ne[1] != w.b->ne[0]) { - gguf_free(ctx_gguf); - ggml_free(ctx); throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)"); } // save tensor to adapter @@ -19032,18 +18999,15 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c { adapter.ctxs.reserve(ctx_map.size()); adapter.bufs.reserve(ctx_map.size()); - for (auto it : ctx_map) { + for (auto & it : ctx_map) { ggml_backend_buffer_type_t buft = it.first; ggml_context * ctx_dev = it.second; - ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft); + ggml_backend_buffer_ptr buf { ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft) }; if (!buf) { - gguf_free(ctx_gguf); - ggml_free(ctx); throw std::runtime_error("failed to allocate buffer for lora adapter\n"); } - LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); - adapter.ctxs.push_back(ctx_dev); - adapter.bufs.push_back(buf); + LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get())/1024.0/1024.0); + adapter.bufs.emplace_back(std::move(buf)); } } @@ -19052,7 +19016,7 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c llama_file gguf_file(path_lora, "rb"); std::vector read_buf; auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) { - size_t offs = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, gguf_find_tensor(ctx_gguf, orig->name)); + size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name)); size_t size = ggml_nbytes(orig); read_buf.resize(size); gguf_file.seek(offs, SEEK_SET); @@ -19068,10 +19032,6 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c } LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2); - - // free ctx for reading gguf - gguf_free(ctx_gguf); - ggml_free(ctx); } int32_t llama_lora_adapter_set( @@ -19480,12 +19440,26 @@ struct llama_context * llama_new_context_with_model( cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL; } - LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); - LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); - LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); - LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); - LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); - LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); + const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; + + LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max); + LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); + LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq); + LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); + LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); + LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); + LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); + LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); + + if (n_ctx_per_seq < hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", + __func__, n_ctx_per_seq, hparams.n_ctx_train); + } + + if (n_ctx_per_seq > hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", + __func__, n_ctx_per_seq, hparams.n_ctx_train); + } ctx->abort_callback = params.abort_callback; ctx->abort_callback_data = params.abort_callback_data; @@ -19520,7 +19494,7 @@ struct llama_context * llama_new_context_with_model( llama_free(ctx); return nullptr; } - ctx->backends.push_back(backend); + ctx->backends.emplace_back(backend); } // add ACCEL backends (such as BLAS) @@ -19533,7 +19507,7 @@ struct llama_context * llama_new_context_with_model( llama_free(ctx); return nullptr; } - ctx->backends.push_back(backend); + ctx->backends.emplace_back(backend); } } @@ -19544,16 +19518,16 @@ struct llama_context * llama_new_context_with_model( llama_free(ctx); return nullptr; } - ctx->backends.push_back(ctx->backend_cpu); + ctx->backends.emplace_back(ctx->backend_cpu); // create a list of the set_n_threads functions in the backends - for (auto * backend : ctx->backends) { - ggml_backend_dev_t dev = ggml_backend_get_device(backend); + for (auto & backend : ctx->backends) { + ggml_backend_dev_t dev = ggml_backend_get_device(backend.get()); ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr; if (reg) { auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"); if (ggml_backend_set_n_threads_fn) { - ctx->set_n_threads_fns.emplace_back(backend, ggml_backend_set_n_threads_fn); + ctx->set_n_threads_fns.emplace_back(backend.get(), ggml_backend_set_n_threads_fn); } } } @@ -19592,17 +19566,18 @@ struct llama_context * llama_new_context_with_model( } LLAMA_LOG_INFO("%s: %10s output buffer size = %8.2f MiB\n", __func__, - ggml_backend_buffer_name(ctx->buf_output), - ggml_backend_buffer_get_size(ctx->buf_output) / 1024.0 / 1024.0); + ggml_backend_buffer_name(ctx->buf_output.get()), + ggml_backend_buffer_get_size(ctx->buf_output.get()) / 1024.0 / 1024.0); } // scheduler and compute buffers { // buffer types used for the compute buffer of each backend std::vector backend_buft; - for (auto * backend : ctx->backends) { - auto * buft = ggml_backend_get_default_buffer_type(backend); - if (ggml_backend_is_cpu(backend) && !model->devices.empty()) { + std::vector backend_ptrs; + for (auto & backend : ctx->backends) { + auto * buft = ggml_backend_get_default_buffer_type(backend.get()); + if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) { // use the host buffer of the first device CPU for faster transfer of the intermediate state auto * dev = model->devices[0]; auto * host_buft = ggml_backend_dev_host_buffer_type(dev); @@ -19611,6 +19586,7 @@ struct llama_context * llama_new_context_with_model( } } backend_buft.push_back(buft); + backend_ptrs.push_back(backend.get()); } const size_t max_nodes = llama_model_max_nodes(*model); @@ -19628,12 +19604,12 @@ struct llama_context * llama_new_context_with_model( // pipeline parallelism requires support for async compute and events in all devices if (pipeline_parallel) { - for (auto * backend : ctx->backends) { - if (ggml_backend_is_cpu(backend)) { + for (auto & backend : ctx->backends) { + if (ggml_backend_is_cpu(backend.get())) { // ignore CPU backend continue; } - auto * dev = ggml_backend_get_device(backend); + auto * dev = ggml_backend_get_device(backend.get()); ggml_backend_dev_props props; ggml_backend_dev_get_props(dev, &props); if (!props.caps.async || !props.caps.events) { @@ -19644,10 +19620,10 @@ struct llama_context * llama_new_context_with_model( } } - ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), max_nodes, pipeline_parallel); + ctx->sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel)); if (pipeline_parallel) { - LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched)); + LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched.get())); } // initialize scheduler with the worst-case graph @@ -19659,29 +19635,29 @@ struct llama_context * llama_new_context_with_model( ggml_cgraph * gf_pp = llama_build_graph(*ctx, ubatch_pp, true); // reserve pp graph first so that buffers are only allocated once - ggml_backend_sched_reserve(ctx->sched, gf_pp); - int n_splits_pp = ggml_backend_sched_get_n_splits(ctx->sched); + ggml_backend_sched_reserve(ctx->sched.get(), gf_pp); + int n_splits_pp = ggml_backend_sched_get_n_splits(ctx->sched.get()); int n_nodes_pp = ggml_graph_n_nodes(gf_pp); // reserve with tg graph to get the number of splits and nodes llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; ggml_cgraph * gf_tg = llama_build_graph(*ctx, ubatch_tg, true); - ggml_backend_sched_reserve(ctx->sched, gf_tg); - int n_splits_tg = ggml_backend_sched_get_n_splits(ctx->sched); + ggml_backend_sched_reserve(ctx->sched.get(), gf_tg); + int n_splits_tg = ggml_backend_sched_get_n_splits(ctx->sched.get()); int n_nodes_tg = ggml_graph_n_nodes(gf_tg); // reserve again with pp graph to avoid ggml-alloc reallocations during inference - gf_pp = llama_build_graph(*ctx, ubatch_pp, false); - if (!ggml_backend_sched_reserve(ctx->sched, gf_pp)) { + gf_pp = llama_build_graph(*ctx, ubatch_pp, true); + if (!ggml_backend_sched_reserve(ctx->sched.get(), gf_pp)) { LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__); llama_free(ctx); return nullptr; } - for (size_t i = 0; i < ctx->backends.size(); i++) { - ggml_backend_t backend = ctx->backends[i]; + for (size_t i = 0; i < backend_ptrs.size(); ++i) { + ggml_backend_t backend = backend_ptrs[i]; ggml_backend_buffer_type_t buft = backend_buft[i]; - size_t size = ggml_backend_sched_get_buffer_size(ctx->sched, backend); + size_t size = ggml_backend_sched_get_buffer_size(ctx->sched.get(), backend); if (size > 1) { LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__, ggml_backend_buft_name(buft), @@ -19961,7 +19937,8 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const // create a context for each buffer type std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { - if (ctx_map.count(buft) == 0) { + auto it = ctx_map.find(buft); + if (it == ctx_map.end()) { struct ggml_init_params params = { /*.mem_size =*/ model.hparams.n_layer*ggml_tensor_overhead(), /*.mem_buffer =*/ NULL, @@ -19972,12 +19949,12 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const return nullptr; } ctx_map[buft] = ctx; - cvec.ctxs.push_back(ctx); + cvec.ctxs.emplace_back(ctx); + return ctx; } - return ctx_map.at(buft); + return it->second; }; - // make tensors cvec.tensors.reserve(model.hparams.n_layer); cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0 @@ -20008,7 +19985,7 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const return false; } ggml_backend_buffer_clear(buf, 0); - cvec.bufs.push_back(buf); + cvec.bufs.emplace_back(buf); } return true; @@ -21276,7 +21253,7 @@ int32_t llama_decode( } void llama_synchronize(struct llama_context * ctx) { - ggml_backend_sched_synchronize(ctx->sched); + ggml_backend_sched_synchronize(ctx->sched.get()); // FIXME: if multiple single tokens are evaluated without a synchronization, // the stats will be added to the prompt evaluation stats