Add model Qwen-VL-Chat to iGPU-perf (#11558)

* Add model Qwen-VL-Chat to iGPU-perf

* small fix

---------

Co-authored-by: ATMxsp01 <shou.xu@intel.com>

.github/workflows/llm_performance_tests.yml

@@ -554,7 +554,7 @@ jobs:
           pip install --upgrade pip
           pip install --upgrade wheel
           pip install --upgrade omegaconf pandas
-          pip install --upgrade tiktoken einops transformers_stream_generator
+          pip install --upgrade tiktoken einops transformers_stream_generator matplotlib
 
           cd python\llm
           python setup.py clean --all bdist_wheel --win
@@ -584,7 +584,7 @@ jobs:
           pip install --upgrade pip
           pip install --upgrade wheel
           pip install --upgrade omegaconf pandas
-          pip install --upgrade tiktoken einops transformers_stream_generator
+          pip install --upgrade tiktoken einops transformers_stream_generator matplotlib
 
           pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/
           pip show ipex-llm | findstr %TEST_VERSION_DATE%

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

@@ -44,6 +44,8 @@ LLAVA_IDS = ['liuhaotian/llava-v1.5-7b']
 
 PHI3VISION_IDS = ['microsoft/phi-3-vision-128k-instruct']
 
+QWENVL_IDS = ['Qwen/Qwen-VL-Chat']
+
 results = []
 excludes = []
 
@@ -923,6 +925,12 @@ def run_transformer_int4_gpu_win(repo_id,
                                                      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 QWENVL_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
+                                                     modules_to_not_convert=['c_fc', 'out_proj'],
+                                                     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).eval()
@@ -1038,6 +1046,13 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
                                                      torch_dtype=torch.float16).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
         model = model.to('xpu')
+    elif repo_id in QWENVL_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
+                                                     modules_to_not_convert=['c_fc', 'out_proj'],
+                                                     trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding,
+                                                     torch_dtype=torch.float16).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,
@@ -1149,6 +1164,12 @@ def run_transformer_int4_loadlowbit_gpu_win(repo_id,
                                                   use_cache=True, cpu_embedding=cpu_embedding).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
         model = model.to('xpu')
+    elif repo_id in QWENVL_IDS:
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                  modules_to_not_convert=['c_fc', 'out_proj'],
+                                                  use_cache=True, cpu_embedding=cpu_embedding).eval()
+        tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
+        model = model.to('xpu')
     else:
         model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
                                                   use_cache=True, cpu_embedding=cpu_embedding).eval()
@@ -1259,6 +1280,12 @@ def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
                                                   use_cache=True, cpu_embedding=cpu_embedding).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
         model = model.half().to('xpu')
+    elif repo_id in QWENVL_IDS:
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                     modules_to_not_convert=['c_fc', 'out_proj'],
+                                                  use_cache=True, cpu_embedding=cpu_embedding).eval()
+        tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
+        model = model.half().to('xpu')
     else:
         model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
                                                   use_cache=True, cpu_embedding=cpu_embedding).eval()

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

@@ -23,7 +23,7 @@ import os
 import sys
 import gc
 
-from run import LLAMA_IDS, CHATGLM_IDS, LLAVA_IDS, PHI3VISION_IDS, get_model_path
+from run import LLAMA_IDS, CHATGLM_IDS, LLAVA_IDS, PHI3VISION_IDS, QWENVL_IDS, get_model_path
 
 current_dir = os.path.dirname(os.path.realpath(__file__))
 
@@ -57,6 +57,11 @@ def save_model_in_low_bit(repo_id,
                                                      modules_to_not_convert=["vision_embed_tokens"],
                                                      trust_remote_code=True, use_cache=True).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in QWENVL_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
+                                                     modules_to_not_convert=['c_fc', 'out_proj'],
+                                                     trust_remote_code=True, use_cache=True).eval()
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
     else:
         model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
                                                      trust_remote_code=True, use_cache=True).eval()

python/llm/test/benchmark/igpu-perf/1024-128.yaml

@@ -12,6 +12,7 @@ repo_id:
   - 'deepseek-ai/deepseek-coder-7b-instruct-v1.5'
   - 'RWKV/v5-Eagle-7B-HF'
   - '01-ai/Yi-6B-Chat'
+  - 'Qwen/Qwen-VL-Chat'
 local_model_hub: 'path to your local model hub'
 warm_up: 1
 num_trials: 3

python/llm/test/benchmark/igpu-perf/1024-128_int4_fp16.yaml

@@ -11,6 +11,7 @@ repo_id:
   - 'mistralai/Mistral-7B-Instruct-v0.2'
   - 'deepseek-ai/deepseek-coder-7b-instruct-v1.5'
   - '01-ai/Yi-6B-Chat'
+  - 'Qwen/Qwen-VL-Chat'
 local_model_hub: 'path to your local model hub'
 warm_up: 1
 num_trials: 3

python/llm/test/benchmark/igpu-perf/1024-128_int4_fp16_loadlowbit.yaml

@@ -11,6 +11,7 @@ repo_id:
   - 'mistralai/Mistral-7B-Instruct-v0.2'
   - 'deepseek-ai/deepseek-coder-7b-instruct-v1.5'
   - '01-ai/Yi-6B-Chat'
+  - 'Qwen/Qwen-VL-Chat'
 local_model_hub: 'path to your local model hub'
 warm_up: 1
 num_trials: 3

python/llm/test/benchmark/igpu-perf/2048-256_int4_fp16.yaml

@@ -11,6 +11,7 @@ repo_id:
   - 'mistralai/Mistral-7B-Instruct-v0.2'
   - 'deepseek-ai/deepseek-coder-7b-instruct-v1.5'
   - '01-ai/Yi-6B-Chat'
+  - 'Qwen/Qwen-VL-Chat'
 local_model_hub: 'path to your local model hub'
 warm_up: 1
 num_trials: 3

python/llm/test/benchmark/igpu-perf/32-32_int4_fp16.yaml

@@ -11,6 +11,7 @@ repo_id:
   - 'mistralai/Mistral-7B-Instruct-v0.2'
   - 'deepseek-ai/deepseek-coder-7b-instruct-v1.5'
   - '01-ai/Yi-6B-Chat'
+  - 'Qwen/Qwen-VL-Chat'
 local_model_hub: 'path to your local model hub'
 warm_up: 3
 num_trials: 5