ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Add extra warmup for chatglm3-6b in igpu-performance test (#11197)

Browse source 
* Add extra warmup for chatglm3-6b to record more stable performance (int4+fp32)

* Small updates
This commit is contained in:
Yuwen Hu 2024-06-04 14:06:09 +08:00 committed by GitHub
parent d90cd977d0
commit 9f8074c653
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 18 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

28
.github/workflows/llm_performance_tests.yml vendored
View file

@ -389,6 +389,14 @@ jobs:
sed -i "s/date.today()/\"$date_for_test_version\"/g" python/llm/dev/benchmark/all-in-one/run.py
- name: Add extra warmup for chatglm3-6b int4+fp32 for more stable results
shell: bash
run: |
sed -i '/^\s*result = run_transformer_int4_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)/ i\
if repo_id in ["THUDM/chatglm3-6b"]:\
run_transformer_int4_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)
' python/llm/dev/benchmark/all-in-one/run.py
- name: Prepare igpu perf test (32-32)
shell: bash
run: |
@ -414,13 +422,13 @@ jobs:
call conda deactivate
- name: Prepare igpu perf test for Qwen1.5 (32-32)
- name: Prepare igpu perf test for transformers 4.37 (32-32)
shell: bash
run: |
sed -i 's/{today}_test1/{today}_test2/g' python/llm/dev/benchmark/all-in-one/run.py
sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf/32-32_437.yaml
- name: Test on igpu for Qwen1.5 (32-32)
- name: Test on igpu for transformers 4.37 (32-32)
shell: cmd
run: |
call conda activate igpu-perf
@ -482,13 +490,13 @@ jobs:
call conda deactivate
- name: Prepare igpu perf test for Qwen 1.5 (1024-128)
- name: Prepare igpu perf test for transformers 4.37 (1024-128)
shell: bash
run: |
sed -i 's/{today}_test1/{today}_test2/g' python/llm/dev/benchmark/all-in-one/run.py
sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf/1024-128_437.yaml
- name: Test on igpu for Qwen 1.5 (1024-128)
- name: Test on igpu for transformers 4.37 (1024-128)
shell: cmd
run: |
call conda activate igpu-perf
@ -549,13 +557,13 @@ jobs:
call conda deactivate
- name: Prepare igpu perf test for Qwen 1.5 (2048-256)
- name: Prepare igpu perf test for transformers 4.37 (2048-256)
shell: bash
run: |
sed -i 's/{today}_test1/{today}_test2/g' python/llm/dev/benchmark/all-in-one/run.py
sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf/2048-256_437.yaml
- name: Test on igpu for Qwen 1.5 (2048-256)
- name: Test on igpu for transformers 4.37 (2048-256)
shell: cmd
run: |
call conda activate igpu-perf
@ -616,13 +624,13 @@ jobs:
call conda deactivate
- name: Prepare igpu perf test for Qwen 1.5 (load_low_bit 1024-128)
- name: Prepare igpu perf test for transformers 4.37 (load_low_bit 1024-128)
shell: bash
run: |
sed -i 's/{today}_test1/{today}_test2/g' python/llm/dev/benchmark/all-in-one/run.py
sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf/1024-128_loadlowbit_437.yaml
- name: Test on igpu for Qwen 1.5 (load_low_bit 1024-128)
- name: Test on igpu for transformers 4.37 (load_low_bit 1024-128)
shell: cmd
run: |
call conda activate igpu-perf
@ -681,13 +689,13 @@ jobs:
call conda deactivate
- name: Prepare igpu perf test for Qwen 1.5 (int4+fp16 1024-128)
- name: Prepare igpu perf test for transformers 4.37 (int4+fp16 1024-128)
shell: bash
run: |
sed -i 's/{today}_test1/{today}_test2/g' python/llm/dev/benchmark/all-in-one/run.py
sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf/1024-128_int4_fp16_437.yaml
- name: Test on igpu for Qwen 1.5 (int4+fp16 1024-128)
- name: Test on igpu for transformers 4.37 (int4+fp16 1024-128)
shell: cmd
run: |
call conda activate igpu-perf