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

cuda : rename build flag to LLAMA_CUDA (#6299)

Browse source
This commit is contained in:
slaren 2024-03-26 01:16:01 +01:00 • committed by GitHub
parent b06c16ef9f
commit 280345968d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
28 changed files with 129 additions and 115 deletions
Download patch file Download diff file Expand all files Collapse all files

4
scripts/LlamaConfig.cmake.in
View file

@ -3,7 +3,7 @@ set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(LLAMA_BLAS @LLAMA_BLAS@)
set(LLAMA_CUBLAS @LLAMA_CUBLAS@)
set(LLAMA_CUDA @LLAMA_CUDA@)
set(LLAMA_METAL @LLAMA_METAL@)
set(LLAMA_MPI @LLAMA_MPI@)
set(LLAMA_CLBLAST @LLAMA_CLBLAST@)
@ -27,7 +27,7 @@ if (LLAMA_BLAS)
find_package(BLAS REQUIRED)
endif()
if (LLAMA_CUBLAS)
if (LLAMA_CUDA)
find_package(CUDAToolkit REQUIRED)
endif()

2
scripts/compare-commits.sh
View file

@ -23,7 +23,7 @@ fi
make_opts=""
if [[ "$backend" == "cuda" ]]; then
make_opts="LLAMA_CUBLAS=1"
make_opts="LLAMA_CUDA=1"
fi
git checkout $1

16
scripts/pod-llama.sh
View file

@ -42,7 +42,7 @@ git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
LLAMA_CUBLAS=1 make -j
LLAMA_CUDA=1 make -j
ln -sfn /workspace/TinyLlama-1.1B-Chat-v0.3 ./models/tinyllama-1b
ln -sfn /workspace/CodeLlama-7b-hf ./models/codellama-7b
@ -60,7 +60,7 @@ cd /workspace/llama.cpp
mkdir build-cublas
cd build-cublas
cmake -DLLAMA_CUBLAS=1 ../
cmake -DLLAMA_CUDA=1 ../
make -j
if [ "$1" -eq "0" ]; then
@ -186,17 +186,17 @@ if [ "$1" -eq "1" ]; then
# batched
cd /workspace/llama.cpp
LLAMA_CUBLAS=1 make -j && ./batched ./models/tinyllama-1b/ggml-model-f16.gguf "Hello, my name is" 8 128 999
LLAMA_CUDA=1 make -j && ./batched ./models/tinyllama-1b/ggml-model-f16.gguf "Hello, my name is" 8 128 999
# batched-bench
cd /workspace/llama.cpp
LLAMA_CUBLAS=1 make -j && ./batched-bench ./models/tinyllama-1b/ggml-model-f16.gguf 4608 1 99 0 512 128 1,2,3,4,5,6,7,8,16,32
LLAMA_CUDA=1 make -j && ./batched-bench ./models/tinyllama-1b/ggml-model-f16.gguf 4608 1 99 0 512 128 1,2,3,4,5,6,7,8,16,32
# parallel
cd /workspace/llama.cpp
LLAMA_CUBLAS=1 make -j && ./parallel -m ./models/tinyllama-1b/ggml-model-f16.gguf -t 1 -ngl 100 -c 4096 -b 512 -s 1 -np 8 -ns 128 -n 100 -cb
LLAMA_CUDA=1 make -j && ./parallel -m ./models/tinyllama-1b/ggml-model-f16.gguf -t 1 -ngl 100 -c 4096 -b 512 -s 1 -np 8 -ns 128 -n 100 -cb
fi
@ -204,10 +204,10 @@ fi
#if [ "$1" -eq "7" ]; then
# cd /workspace/llama.cpp
#
# LLAMA_CUBLAS=1 make -j && ./speculative -m ./models/codellama-34b-instruct/ggml-model-f16.gguf -md ./models/codellama-7b-instruct/ggml-model-q4_0.gguf -p "# Dijkstra's shortest path algorithm in Python (4 spaces indentation) + complexity analysis:\n\n" -e -ngl 999 -ngld 999 -t 4 -n 512 -c 4096 -s 21 --draft 16 -np 1 --temp 0.0
# LLAMA_CUDA=1 make -j && ./speculative -m ./models/codellama-34b-instruct/ggml-model-f16.gguf -md ./models/codellama-7b-instruct/ggml-model-q4_0.gguf -p "# Dijkstra's shortest path algorithm in Python (4 spaces indentation) + complexity analysis:\n\n" -e -ngl 999 -ngld 999 -t 4 -n 512 -c 4096 -s 21 --draft 16 -np 1 --temp 0.0
#fi
# more benches
#LLAMA_CUBLAS=1 make -j && ./batched-bench ./models/codellama-7b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1
#LLAMA_CUBLAS=1 make -j && ./batched-bench ./models/codellama-13b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1
#LLAMA_CUDA=1 make -j && ./batched-bench ./models/codellama-7b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1
#LLAMA_CUDA=1 make -j && ./batched-bench ./models/codellama-13b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1

2
scripts/server-llm.sh
View file

@ -380,7 +380,7 @@ fi
if [[ "$backend" == "cuda" ]]; then
printf "[+] Building with CUDA backend\n"
LLAMA_CUBLAS=1 make -j server $log
LLAMA_CUDA=1 make -j server $log
elif [[ "$backend" == "cpu" ]]; then
printf "[+] Building with CPU backend\n"
make -j server $log