vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'ggerganov:master' into avx_opt

Browse source
This commit is contained in:
Eve 2024-11-03 01:53:49 +00:00 • committed by GitHub
commit a83ac00565
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
27 changed files with 1066 additions and 633 deletions
Download patch file Download diff file Expand all files Collapse all files

13
CMakePresets.json
View file

@ -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" ] },

6
Makefile
View file

@ -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, $<)

3
README.md
View file

@ -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)
----

164
ci/run.sh
View file

@ -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 "</s>"
(time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>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?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>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

16
cmake/arm64-apple-clang.cmake Normal file
View file

@ -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}" )

2
common/common.h
View file

@ -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

27
convert_hf_to_gguf.py
View file

@ -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:

47
convert_lora_to_gguf.py
View file

@ -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...")

1
examples/CMakeLists.txt
View file

@ -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()

57
examples/server/server.cpp
View file

@ -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<float>(lcp_len) / static_cast<int>(slot.prompt_tokens.size());
// fraction of the common subsequence length compared to the current slot's prompt length
float cur_similarity = static_cast<float>(cur_lcs_len) / static_cast<int>(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, &params, &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();

52
examples/server/utils.hpp
View file

@ -439,18 +439,60 @@ static std::string gen_chatcmplid() {
// other common utils
//
static size_t longest_common_prefix(const std::vector<llama_token> & a, const std::vector<llama_token> & 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<size_t> prev_row(b_len + 1, 0);
std::vector<size_t> 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) {

5
examples/simple-chat/CMakeLists.txt Normal file
View file

@ -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)

7
examples/simple-chat/README.md Normal file
View file

@ -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
...

197
examples/simple-chat/simple-chat.cpp Normal file
View file

@ -0,0 +1,197 @@
#include "llama.h"
#include <cstdio>
#include <cstring>
#include <iostream>
#include <string>
#include <vector>
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<llama_token> 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<llama_chat_message> messages;
std::vector<char> 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<char *>(msg.content));
}
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
return 0;
}

38
ggml/include/ggml-cpp.h Normal file
View file

@ -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 <memory>
// 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, ggml_context_deleter> ggml_context_ptr;
typedef std::unique_ptr<gguf_context, gguf_context_deleter> 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_t, ggml_gallocr_deleter> 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, ggml_backend_deleter> ggml_backend_ptr;
typedef std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> ggml_backend_buffer_ptr;
typedef std::unique_ptr<ggml_backend_event, ggml_backend_event_deleter> ggml_backend_event_ptr;
typedef std::unique_ptr<ggml_backend_sched, ggml_backend_sched_deleter> ggml_backend_sched_ptr;

22
ggml/include/ggml.h
View file

@ -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);

5
ggml/src/CMakeLists.txt
View file

@ -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()

2
ggml/src/ggml-backend.cpp
View file

@ -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__)

6
ggml/src/ggml-cuda.cu
View file

@ -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:

42
ggml/src/ggml-kompute.cpp
View file

@ -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>(args)...);
}
static void ggml_vk_mul_mat_q4_k(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& 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<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__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<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_q6_k(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& 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,

5
ggml/src/ggml-vulkan.cpp
View file

@ -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);

176
ggml/src/ggml.c
View file

@ -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) {

9
ggml/src/kompute-shaders/common.comp
View file

@ -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

133
ggml/src/kompute-shaders/op_mul_mat_q4_k.comp Normal file
View file

@ -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;
}
}
}

2
scripts/sync-ggml.last
View file

@ -1 +1 @@
162e232411ee98ceb0cccfa84886118d917d2123
bb78a40dc60e04c626bac2b65840b509988e990d

1
spm-headers/ggml-cpp.h Symbolic link
View file

@ -0,0 +1 @@
../ggml/include/ggml-cpp.h

661
src/llama.cpp
View file

File diff suppressed because it is too large Load diff