Add model phi-3-vision-128k-instruct to iGPU-perf benchmark (#11554)

* try to improve MIniCPM performance

* Add model phi-3-vision-128k-instruct to iGPU-perf benchmark

---------

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

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

@@ -42,6 +42,8 @@ CHATGLM_IDS = ['THUDM/chatglm-6b', 'THUDM/chatglm2-6b', 'THUDM/chatglm3-6b']
 
 LLAVA_IDS = ['liuhaotian/llava-v1.5-7b']
 
+PHI3VISION_IDS = ['microsoft/phi-3-vision-128k-instruct']
+
 results = []
 excludes = []
 
@@ -914,6 +916,13 @@ 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 PHI3VISION_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
+                                                     _attn_implementation="eager",
+                                                     modules_to_not_convert=["vision_embed_tokens"],
+                                                     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()
@@ -1021,6 +1030,14 @@ 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 PHI3VISION_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
+                                                     _attn_implementation="eager",
+                                                     modules_to_not_convert=["vision_embed_tokens"],
+                                                     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,
@@ -1125,6 +1142,13 @@ 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 PHI3VISION_IDS:
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                  _attn_implementation="eager",
+                                                  modules_to_not_convert=["vision_embed_tokens"],
+                                                  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()
@@ -1228,6 +1252,13 @@ 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 PHI3VISION_IDS:
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                  _attn_implementation="eager",
+                                                  modules_to_not_convert=["vision_embed_tokens"],
+                                                  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, get_model_path
+from run import LLAMA_IDS, CHATGLM_IDS, LLAVA_IDS, PHI3VISION_IDS, get_model_path
 
 current_dir = os.path.dirname(os.path.realpath(__file__))
 
@@ -51,6 +51,12 @@ def save_model_in_low_bit(repo_id,
         model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
                                           trust_remote_code=True, use_cache=True).eval()
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in PHI3VISION_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
+                                                     _attn_implementation="eager",
+                                                     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)
     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_437.yaml

@@ -3,6 +3,7 @@ repo_id:
   - 'Qwen/Qwen2-7B-Instruct'
   - 'microsoft/Phi-3-mini-4k-instruct'
   - 'microsoft/Phi-3-mini-128k-instruct'
+  - 'microsoft/phi-3-vision-128k-instruct'
 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_437.yaml

@@ -3,6 +3,7 @@ repo_id:
   - 'Qwen/Qwen2-7B-Instruct'
   - 'microsoft/Phi-3-mini-4k-instruct'
   - 'microsoft/Phi-3-mini-128k-instruct'
+  - 'microsoft/phi-3-vision-128k-instruct'
 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_437.yaml

@@ -3,6 +3,7 @@ repo_id:
   - 'Qwen/Qwen2-7B-Instruct'
   - 'microsoft/Phi-3-mini-4k-instruct'
   - 'microsoft/Phi-3-mini-128k-instruct'
+  - 'microsoft/phi-3-vision-128k-instruct'
 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_437.yaml

@@ -3,6 +3,7 @@ repo_id:
   - 'Qwen/Qwen2-7B-Instruct'
   - 'microsoft/Phi-3-mini-4k-instruct'
   - 'microsoft/Phi-3-mini-128k-instruct'
+  - 'microsoft/phi-3-vision-128k-instruct'
 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_437.yaml

@@ -3,6 +3,7 @@ repo_id:
   - 'Qwen/Qwen2-7B-Instruct'
   - 'microsoft/Phi-3-mini-4k-instruct'
   - 'microsoft/Phi-3-mini-128k-instruct'
+  - 'microsoft/phi-3-vision-128k-instruct'
 local_model_hub: 'path to your local model hub'
 warm_up: 3
 num_trials: 5