Arc stable version performance regression test (#9785)

* add arc stable version regression test

* empty gpu mem between different models

* triger ci

* comment spr test

* triger ci

* address kai's comments and disable ci

* merge fp8 and int4

* disable ci

.github/workflows/llm_performance_tests_stable_version.yml

@@ -32,7 +32,7 @@ jobs:
       OMP_NUM_THREADS: 16
       THREAD_NUM: 16
       ANALYTICS_ZOO_ROOT: ${{ github.workspace }}
-      CSV_SAVE_PATH: ${{ github.event.schedule && '/mnt/disk1/nightly_perf_gpu/' || '/mnt/disk1/pr_perf_gpu/' }}
+      CSV_SAVE_PATH: '/mnt/disk1/stable_version_perf_gpu/'
 
     steps:
       - uses: actions/checkout@v3
@@ -69,36 +69,41 @@ jobs:
           source /home/arda/intel/oneapi/setvars.sh
           bash python/llm/test/run-llm-install-tests.sh
 
-      - name: Test on xpu
+      - name: Test on xpu (int4)
         shell: bash
         run: |
           source /home/arda/intel/oneapi/setvars.sh
           export USE_XETLA=OFF
           export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
-          mv python/llm/test/benchmark/stable-version-arc-perf-test.yaml python/llm/dev/benchmark/all-in-one/config.yaml
+          mv python/llm/test/benchmark/stable-version-arc-perf-test-sym_int4.yaml python/llm/dev/benchmark/all-in-one/config.yaml
           cd python/llm/dev/benchmark/all-in-one
           # hide time info
           sed -i 's/str(end - st)/"xxxxxx"/g' run.py
-          # change csv name
-          sed -i 's/{today}/{today}_test1/g' run.py
           python run.py
-          # upgrade transformers for model Mistral-7B-v0.1
-          python -m pip install transformers==4.34.0
-          mv ../../../test/benchmark/arc-perf-transformers-434.yaml ./config.yaml
-          # change csv name
-          sed -i 's/test1/test2/g' run.py
-          python run.py
-          python ../../../test/benchmark/concat_csv.py
-          cp ./*.csv $CSV_SAVE_PATH
+          cp ./*.csv $CSV_SAVE_PATH/int4
+          rm ./*.csv
           cd ../../../test/benchmark
           python -m pip install pandas==1.5.3
-          python csv_to_html.py -f $CSV_SAVE_PATH
-          cd ../../dev/benchmark/all-in-one/
-          python ../../../test/benchmark/check_results.py -n 45
-          if [ ${{ github.event.schedule}} ]; then
-            curl -T ./*.csv ${LLM_FTP_URL}/llm/nightly_perf/gpu/
-          fi
-          
+          python csv_to_html.py -f $CSV_SAVE_PATH/int4 -b $CSV_SAVE_PATH/int4/transformer_int4_gpu-results-1baseline.csv -t 5.0
+
+      - name: Test on xpu (fp8)
+        shell: bash
+        run: |
+          source /home/arda/intel/oneapi/setvars.sh
+          export USE_XETLA=OFF
+          export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+          mv python/llm/test/benchmark/stable-version-arc-perf-test-fp8.yaml python/llm/dev/benchmark/all-in-one/config.yaml
+          cd python/llm/dev/benchmark/all-in-one
+          # hide time info
+          sed -i 's/str(end - st)/"xxxxxx"/g' run.py
+          python run.py
+          cp ./*.csv $CSV_SAVE_PATH/fp8
+          rm ./*.csv
+          cd ../../../test/benchmark
+          python -m pip install pandas==1.5.3
+          python csv_to_html.py -f $CSV_SAVE_PATH/fp8 -b $CSV_SAVE_PATH/fp8/transformer_int4_gpu-results-1baseline.csv -t 5.0
+
+
   llm-performance-test-on-spr:
     needs: llm-cpp-build
     strategy:

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

@@ -413,8 +413,11 @@ def run_transformer_int4_gpu(repo_id,
             thread = threading.Thread(target=run_model_in_thread, args=(model, in_out, tokenizer, result, warm_up, num_beams, input_ids, out_len, actual_in_len, num_trials, reserved_mem_list))
             thread.start()
             thread.join()
-    del model
+    model.to('cpu')
+    torch.xpu.synchronize()
     torch.xpu.empty_cache()
+    del model
+    gc.collect()
     return result
 
 

python/llm/test/benchmark/stable-version-arc-perf-test-fp8.yaml

@@ -0,0 +1,18 @@
+repo_id:
+  - 'meta-llama/Llama-2-7b-chat-hf'
+  - 'THUDM/chatglm2-6b'
+  - 'THUDM/chatglm3-6b'
+  - 'baichuan-inc/Baichuan2-7B-Chat'
+  - 'Qwen/Qwen-7B-Chat'
+local_model_hub: '/mnt/disk1/models'
+warm_up: 1
+num_trials: 3
+num_beams: 1 # default to greedy search
+low_bit: 'fp8' # default to use 'sym_int4' (i.e. symmetric int4)
+in_out_pairs:
+  - '32-32'
+  - '1024-128'
+  - '2048-256'
+test_api:
+  - "transformer_int4_gpu"  # on Intel GPU
+cpu_embedding: False # whether put embedding to CPU (only avaiable now for gpu win related test_api)

python/llm/test/benchmark/stable-version-arc-perf-test-sym_int4.yaml

@@ -16,7 +16,3 @@ in_out_pairs:
 test_api:
   - "transformer_int4_gpu"  # on Intel GPU
 cpu_embedding: False # whether put embedding to CPU (only avaiable now for gpu win related test_api)
-exclude:
-  - 'fnlp/moss-moon-003-sft:1024'
-  - 'fnlp/moss-moon-003-sft:2048'
-  - 'bigscience/bloomz-7b1:2048'