[LLM] win igpu performance for ipex 2.1 and oneapi 2024.0 (#9679)

* Change igpu win tests for ipex 2.1 and oneapi 2024.0

* Qwen model repo id updates; updates model list for 512-64

* Add .eval for win igpu all-in-one benchmark for best performance

.github/workflows/llm_performance_tests.yml

@@ -250,7 +250,7 @@ jobs:
           if not exist dist\bigdl_llm*.whl (exit /b 1)
           for %%i in (dist\bigdl_llm*.whl) do set whl_name=%%i
 
-          pip install %whl_name%[xpu] -i %INTERNAL_PYPI_URL% --trusted-host %INTERNAL_PYPI_TRUSTED_HOST% -q
+          pip install --pre --upgrade %whl_name%[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
           if %ERRORLEVEL% neq 0 (exit /b 1)
           pip list
 
@@ -291,7 +291,6 @@ jobs:
         run: |
           call conda activate igpu-perf
           call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
-          set SYCL_ENABLE_DEFAULT_CONTEXTS=1
           set SYCL_CACHE_PERSISTENT=1
           set BIGDL_LLM_XMX_DISABLED=1
           REM for llava
@@ -317,7 +316,6 @@ jobs:
           pip install transformers==4.34.0
 
           call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
-          set SYCL_ENABLE_DEFAULT_CONTEXTS=1
           set SYCL_CACHE_PERSISTENT=1
           set BIGDL_LLM_XMX_DISABLED=1
 
@@ -359,7 +357,6 @@ jobs:
           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
           set BIGDL_LLM_XMX_DISABLED=1
           REM for llava
@@ -385,7 +382,6 @@ jobs:
           pip install transformers==4.34.0
 
           call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
-          set SYCL_ENABLE_DEFAULT_CONTEXTS=1
           set SYCL_CACHE_PERSISTENT=1
           set BIGDL_LLM_XMX_DISABLED=1
 
@@ -427,7 +423,6 @@ jobs:
           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
           set BIGDL_LLM_XMX_DISABLED=1
           REM for llava
@@ -453,7 +448,6 @@ jobs:
           pip install transformers==4.34.0
 
           call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
-          set SYCL_ENABLE_DEFAULT_CONTEXTS=1
           set SYCL_CACHE_PERSISTENT=1
           set BIGDL_LLM_XMX_DISABLED=1
 
@@ -494,7 +488,6 @@ jobs:
           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
           set BIGDL_LLM_XMX_DISABLED=1
           REM for llava

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

@@ -666,12 +666,12 @@ def run_transformer_int4_gpu_win(repo_id,
     st = time.perf_counter()
     if repo_id in CHATGLM_IDS:
         model = AutoModel.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
-                                          trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding)
+                                          trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
         model = model.to('xpu')
     elif repo_id in LLAMA_IDS:
         model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, trust_remote_code=True,
-                                                     use_cache=True, cpu_embedding=cpu_embedding)
+                                                     use_cache=True, cpu_embedding=cpu_embedding).eval()
         tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
         model = model.to('xpu')
     elif repo_id in LLAVA_IDS:
@@ -679,12 +679,12 @@ def run_transformer_int4_gpu_win(repo_id,
         sys.path.append(rf"{llava_repo_dir}")
         from llava.model.language_model.llava_llama import LlavaLlamaForCausalLM
         model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
-                                          trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding)
+                                          trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
         model = model.to('xpu')
     else:
         model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
-                                                     trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding)
+                                                     trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
         model = model.to('xpu')
         if isinstance(model, GPTJForCausalLM):

python/llm/test/benchmark/igpu-perf/32-256.yaml

@@ -3,7 +3,7 @@ repo_id:
   - 'THUDM/chatglm3-6b'
   - 'baichuan-inc/Baichuan2-7B-Chat'
   - 'internlm/internlm-chat-7b-8k'
-  - 'Qwen/Qwen-7B-Chat-10-12'
+  - 'Qwen/Qwen-7B-Chat'
   - 'BAAI/AquilaChat2-7B'
   - '01-ai/Yi-6B'
   - 'meta-llama/Llama-2-7b-chat-hf'

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

@@ -3,7 +3,7 @@ repo_id:
   - 'THUDM/chatglm3-6b'
   - 'baichuan-inc/Baichuan2-7B-Chat'
   - 'internlm/internlm-chat-7b-8k'
-  - 'Qwen/Qwen-7B-Chat-10-12'
+  - 'Qwen/Qwen-7B-Chat'
   - 'BAAI/AquilaChat2-7B'
   - '01-ai/Yi-6B'
   - 'meta-llama/Llama-2-7b-chat-hf'

python/llm/test/benchmark/igpu-perf/32-512.yaml

@@ -3,7 +3,7 @@ repo_id:
   - 'THUDM/chatglm3-6b'
   - 'baichuan-inc/Baichuan2-7B-Chat'
   - 'internlm/internlm-chat-7b-8k'
-  - 'Qwen/Qwen-7B-Chat-10-12'
+  - 'Qwen/Qwen-7B-Chat'
   - 'BAAI/AquilaChat2-7B'
   - '01-ai/Yi-6B'
   - 'meta-llama/Llama-2-7b-chat-hf'

python/llm/test/benchmark/igpu-perf/512-64.yaml

@@ -1,16 +1,16 @@
 repo_id:
   - 'THUDM/chatglm2-6b'
-  # - 'THUDM/chatglm3-6b'
+  - 'THUDM/chatglm3-6b'
   # - 'baichuan-inc/Baichuan2-7B-Chat'
-  # - 'internlm/internlm-chat-7b-8k'
+  - 'internlm/internlm-chat-7b-8k'
-  # - 'Qwen/Qwen-7B-Chat-10-12'
+  # - 'Qwen/Qwen-7B-Chat'
-  # - 'BAAI/AquilaChat2-7B'
+  - 'BAAI/AquilaChat2-7B'
   - '01-ai/Yi-6B'
-  # - 'meta-llama/Llama-2-7b-chat-hf'
+  - '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'
+  - 'liuhaotian/llava-v1.5-7b'
   - 'RWKV/rwkv-4-world-7b'
 local_model_hub: 'path to your local model hub'
 warm_up: 1