[LLM] iGPU long context tests (intel-analytics#9598)

* Temp enable PR

* Enable tests for 256-64

* Try again 128-64

* Empty cache after each iteration for igpu benchmark scripts

* Try tests for 512

* change order for 512

* Skip chatglm3 and llama2 for now

* Separate tests for 512-64

* Small fix

* Further fixes

* Change back to nightly again

.github/workflows/llm_performance_tests.yml

@@ -249,6 +249,7 @@ jobs:
 
           pip install %whl_name%[xpu] -i %INTERNAL_PYPI_URL% --trusted-host %INTERNAL_PYPI_TRUSTED_HOST% -q
           if %ERRORLEVEL% neq 0 (exit /b 1)
+          pip list
 
           call conda deactivate
 
@@ -263,15 +264,15 @@ jobs:
           cur_date=$(date +%Y-%m-%d)
           echo "LOG_FILE=${cur_date}_output.txt" >> "$GITHUB_ENV"
 
-      - name: Prepare igpu perf test
+      - name: Prepare igpu perf test (32)
         shell: bash
         run: |
           # hide time info
           sed -i 's/str(end - st)/"xxxxxx"/g' python/llm/dev/benchmark/all-in-one/run.py
           sed -i 's/{today}/{today}_test1/g' python/llm/dev/benchmark/all-in-one/run.py
-          sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf-test.yaml
+          sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/32-igpu-perf-test.yaml
 
-      - name: Test on igpu
+      - name: Test on igpu (32)
         shell: cmd
         run: |
           call conda activate igpu-perf
@@ -281,20 +282,20 @@ jobs:
           REM for llava
           set TRANSFORMERS_OFFLINE=1
 
-          move python\llm\test\benchmark\igpu-perf-test.yaml python\llm\dev\benchmark\all-in-one\config.yaml
           cd python\llm\dev\benchmark\all-in-one
-          python run.py >> %LOG_FILE% 2>&1
+          move ..\..\..\test\benchmark\32-igpu-perf-test.yaml config.yaml
+          python run.py >> %CSV_SAVE_PATH%\32\log\%LOG_FILE% 2>&1
           if %ERRORLEVEL% neq 0 (exit /b 1)
 
           call conda deactivate
 
-      - name: Prepare igpu perf test for Mistral
+      - name: Prepare igpu perf test for Mistral (32)
         shell: bash
         run: |
           sed -i 's/test1/test2/g' python/llm/dev/benchmark/all-in-one/run.py
-          sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/igpu-perf-test-434.yaml
+          sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/32-igpu-perf-test-434.yaml
 
-      - name: Test on igpu for Mistral
+      - name: Test on igpu for Mistral (32)
         shell: cmd
         run: |
           call conda activate igpu-perf
@@ -304,31 +305,71 @@ jobs:
           set SYCL_ENABLE_DEFAULT_CONTEXTS=1
           set SYCL_CACHE_PERSISTENT=1
 
-          move python\llm\test\benchmark\igpu-perf-test-434.yaml python\llm\dev\benchmark\all-in-one\config.yaml
           cd python\llm\dev\benchmark\all-in-one
-          python run.py >> %LOG_FILE% 2>&1
+          move ..\..\..\test\benchmark\32-igpu-perf-test-434.yaml config.yaml
+          python run.py >> %CSV_SAVE_PATH%\32\log\%LOG_FILE% 2>&1
           if %ERRORLEVEL% neq 0 (exit /b 1)
 
           call conda deactivate
 
-      - name: Concat csv and generate html
+      - name: Concat csv and generate html (32)
         shell: cmd
         run: |
           call conda activate igpu-perf
 
           cd python\llm\dev\benchmark\all-in-one
-          move %LOG_FILE% %CSV_SAVE_PATH%\log\
           python ..\..\..\test\benchmark\concat_csv.py
-          copy *.csv %CSV_SAVE_PATH%
+          copy *.csv %CSV_SAVE_PATH%\32\
           del /q *.csv
           cd ..\..\..\test\benchmark
-          python csv_to_html.py -f %CSV_SAVE_PATH%
+          python csv_to_html.py -f %CSV_SAVE_PATH%\32\
           if %ERRORLEVEL% neq 0 (exit /b 1)
 
           call conda deactivate
 
-      - name: Remove conda env
-        if: ${{ always() }}
+      # TODO: create a action function here for different input
+      - name: Prepare igpu perf test (512)
+        shell: bash
+        run: |
+          sed -i 's/{today}_test2/{today}/g' python/llm/dev/benchmark/all-in-one/run.py
+          sed -i "s/path to your local model hub/$MODEL_HUB_DIR/g" python/llm/test/benchmark/512-igpu-perf-test.yaml
+
+      - name: Test on igpu (512)
         shell: cmd
         run: |
-          call conda env remove -n igpu-perf -y
+          call conda activate igpu-perf
+          pip install transformers==4.31.0
+
+          call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
+          set SYCL_ENABLE_DEFAULT_CONTEXTS=1
+          set SYCL_CACHE_PERSISTENT=1
+          REM for llava
+          set TRANSFORMERS_OFFLINE=1
+
+          cd python\llm\dev\benchmark\all-in-one
+          move ..\..\..\test\benchmark\512-igpu-perf-test.yaml config.yaml
+          python run.py >> %CSV_SAVE_PATH%\512\log\%LOG_FILE% 2>&1
+          if %ERRORLEVEL% neq 0 (exit /b 1)
+
+          call conda deactivate
+
+      - name: Generate html (512)
+        shell: cmd
+        run: |
+          call conda activate igpu-perf
+
+          cd python\llm\dev\benchmark\all-in-one
+          copy *.csv %CSV_SAVE_PATH%\512\
+          del /q *.csv
+          cd ..\..\..\test\benchmark
+          python csv_to_html.py -f %CSV_SAVE_PATH%\512\
+          if %ERRORLEVEL% neq 0 (exit /b 1)
+
+          call conda deactivate
+
+      # for test on machine when encountering error
+      # - name: Remove conda env
+      #   if: ${{ always() }}
+      #   shell: cmd
+      #   run: |
+      #     call conda env remove -n igpu-perf -y

python/llm/dev/benchmark/all-in-one/run.py

@@ -731,6 +731,7 @@ def run_transformer_int4_gpu_win(repo_id,
                     if i >= warm_up:
                         result[in_out].append([model.first_cost, model.rest_cost_mean, model.encoder_time,
                                             actual_in_len, actual_out_len, gpu_peak_mem])
+                    torch.xpu.empty_cache()
             except RuntimeError:
                 pass
     model.to('cpu')

python/llm/test/benchmark/igpu-perf-test-434.yaml → python/llm/test/benchmark/32-igpu-perf-test-434.yaml

@@ -7,7 +7,6 @@ num_beams: 1 # default to greedy search
 low_bit: 'sym_int4' # default to use 'sym_int4' (i.e. symmetric int4)
 in_out_pairs:
   - '32-32'
-  # - '512-64'
   # - '1024-128'
 test_api:
   # - "transformer_int4"

python/llm/test/benchmark/igpu-perf-test.yaml → python/llm/test/benchmark/32-igpu-perf-test.yaml

@@ -18,7 +18,6 @@ num_beams: 1 # default to greedy search
 low_bit: 'sym_int4' # default to use 'sym_int4' (i.e. symmetric int4)
 in_out_pairs:
   - '32-32'
-  # - '512-64'
   # - '1024-128'
 test_api:
   # - "transformer_int4"

python/llm/test/benchmark/512-igpu-perf-test-434.yaml

@@ -0,0 +1,21 @@
+repo_id:
+  - 'mistralai/Mistral-7B-Instruct-v0.1'
+local_model_hub: 'path to your local model hub'
+warm_up: 1
+num_trials: 3
+num_beams: 1 # default to greedy search
+low_bit: 'sym_int4' # default to use 'sym_int4' (i.e. symmetric int4)
+in_out_pairs:
+  - '512-64'
+  # - '1024-128'
+test_api:
+  # - "transformer_int4"
+  # - "native_int4"
+  # - "optimize_model"
+  # - "pytorch_autocast_bf16"
+  # - "ipex_fp16_gpu" # on Intel GPU
+  # - "transformer_int4_gpu"  # on Intel GPU
+  # - "optimize_model_gpu"  # on Intel GPU
+  # - "deepspeed_transformer_int4_cpu" # on Intel SPR Server
+  - "transformer_int4_gpu_win" # on Intel GPU for Windows (catch GPU peak memory)
+cpu_embedding: True # whether put embedding to CPU (only avaiable now for gpu win related test_api)

python/llm/test/benchmark/512-igpu-perf-test.yaml

@@ -0,0 +1,32 @@
+repo_id:
+  - 'THUDM/chatglm2-6b'
+  # - 'THUDM/chatglm3-6b'
+  # - 'baichuan-inc/Baichuan2-7B-Chat'
+  # - 'internlm/internlm-chat-7b-8k'
+  # - 'Qwen/Qwen-7B-Chat-10-12'
+  # - 'BAAI/AquilaChat2-7B'
+  - '01-ai/Yi-6B'
+  # - 'meta-llama/Llama-2-7b-chat-hf'
+  # - 'WisdomShell/CodeShell-7B-Chat'
+  - 'tiiuae/falcon-7b-instruct-with-patch'
+  - 'mosaicml/mpt-7b-chat'
+  # - 'liuhaotian/llava-v1.5-7b'
+local_model_hub: 'path to your local model hub'
+warm_up: 1
+num_trials: 3
+num_beams: 1 # default to greedy search
+low_bit: 'sym_int4' # default to use 'sym_int4' (i.e. symmetric int4)
+in_out_pairs:
+  - '512-64'
+  # - '1024-128'
+test_api:
+  # - "transformer_int4"
+  # - "native_int4"
+  # - "optimize_model"
+  # - "pytorch_autocast_bf16"
+  # - "ipex_fp16_gpu" # on Intel GPU
+  # - "transformer_int4_gpu"  # on Intel GPU
+  # - "optimize_model_gpu"  # on Intel GPU
+  # - "deepspeed_transformer_int4_cpu" # on Intel SPR Server
+  - "transformer_int4_gpu_win" # on Intel GPU for Windows (catch GPU peak memory)
+cpu_embedding: True # whether put embedding to CPU (only avaiable now for gpu win related test_api)