[Feat] fix tasks and vllm to reproduce better results. (#774)

* modify gpt evaluation model name retrieval

* [Enhancement] Improve score processing and validation in mia_bench

* [Refactor] Update MathVerseEvaluator initialization and API key handling

* remove sampling-related parameters to avoid warnings

* fix llm judge with azure

* Refactor mathverse result aggregation by removing redundant file saving steps during evaluation

* Remove unused submission metric from mathverse_testmini.yaml configuration

* Remove unused generation parameters from mmmu_val_thinking.yaml to streamline configuration

* Refactor mathvista result aggregation by removing file saving steps from evaluation process

* [Refactor] Update VLLM model initialization and configuration parameters

* [Refactor] Update model argument naming in VLLM configuration

* [Refactor] Clean up whitespace and improve regex patterns in score processing

* feat(vllm): Add default temperature and top_p to generation params

---------

Co-authored-by: brian.li <[email protected]>

Author: Luodian

examples/models/vllm_qwen2vl.sh

@@ -10,7 +10,7 @@ export NCCL_DEBUG=DEBUG
 
 python3 -m lmms_eval \
     --model vllm \
-    --model_args model_version=Qwen/Qwen2-VL-7B-Instruct,tensor_parallel_size=4 \
+    --model_args model=Qwen/Qwen2-VL-7B-Instruct,tensor_parallel_size=4 \
     --tasks mme,gsm8k_cot_self_consistency,mmmu_val \
     --batch_size 64 \
     --log_samples \

lmms_eval/models/llava_onevision.py

@@ -518,18 +518,6 @@ def _collate(x):
                     prompt_question = conv.get_prompt()
                     question_input.append(prompt_question)
 
-            # preconfigure gen_kwargs with defaults
-            if "max_new_tokens" not in gen_kwargs:
-                gen_kwargs["max_new_tokens"] = 1024
-            if "temperature" not in gen_kwargs:
-                gen_kwargs["temperature"] = 0
-            if "do_sample" not in gen_kwargs:
-                gen_kwargs["do_sample"] = False
-            if "top_p" not in gen_kwargs:
-                gen_kwargs["top_p"] = None
-            if "num_beams" not in gen_kwargs:
-                gen_kwargs["num_beams"] = 1
-
             input_ids_list = [tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt") for prompt in question_input]
             pad_token_ids = self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id
             input_ids = self.pad_sequence(input_ids_list, batch_first=True, padding_value=pad_token_ids).to(self.device)
@@ -548,8 +536,18 @@ def _collate(x):
 
             # These steps are not in LLaVA's original code, but are necessary for generation to work
             # TODO: attention to this major generation step...
+            # preconfigure gen_kwargs with defaults
+            if "max_new_tokens" not in gen_kwargs:
+                gen_kwargs["max_new_tokens"] = 1024
+
             if "image_aspect_ratio" in gen_kwargs.keys():
                 gen_kwargs.pop("image_aspect_ratio")
+            # When do_sample=False, remove sampling-related parameters to avoid warnings
+            # These might be in gen_kwargs or in the model's generation_config
+            if not gen_kwargs.get("do_sample", False):
+                gen_kwargs.pop("temperature", None)
+                gen_kwargs.pop("top_p", None)
+                gen_kwargs.pop("top_k", None)
             try:
                 with torch.inference_mode():
                     cont = self.model.generate(input_ids, attention_mask=attention_masks, pad_token_id=pad_token_ids, images=image_tensor, use_cache=self.use_cache, **gen_kwargs)
@@ -732,12 +730,14 @@ def _collate(x):
                 # preconfigure gen_kwargs with defaults
                 if "max_new_tokens" not in gen_kwargs:
                     gen_kwargs["max_new_tokens"] = 1024
-                if "temperature" not in gen_kwargs:
-                    gen_kwargs["temperature"] = 0
                 if "do_sample" not in gen_kwargs:
                     gen_kwargs["do_sample"] = False
-                if "top_p" not in gen_kwargs:
-                    gen_kwargs["top_p"] = None
+                # Only set temperature and top_p if do_sample is True
+                if gen_kwargs.get("do_sample", False):
+                    if "temperature" not in gen_kwargs:
+                        gen_kwargs["temperature"] = 1.0  # Default temperature for sampling
+                    if "top_p" not in gen_kwargs:
+                        gen_kwargs["top_p"] = 1.0  # Default top_p for sampling
                 if "num_beams" not in gen_kwargs:
                     gen_kwargs["num_beams"] = 1
 
@@ -761,6 +761,10 @@ def _collate(x):
                 # TODO: attention to this major generation step...
                 if "image_aspect_ratio" in gen_kwargs.keys():
                     gen_kwargs.pop("image_aspect_ratio")
+                # Remove temperature and top_p when do_sample=False to avoid warnings
+                if not gen_kwargs.get("do_sample", False):
+                    gen_kwargs.pop("temperature", None)
+                    gen_kwargs.pop("top_p", None)
                 try:
                     with torch.inference_mode():
                         cont = self.model.generate(input_ids, attention_mask=attention_masks, pad_token_id=pad_token_ids, images=image_tensor, use_cache=self.use_cache, **gen_kwargs)

lmms_eval/models/vllm.py

@@ -20,7 +20,8 @@
 from lmms_eval.api.model import lmms
 from lmms_eval.api.registry import register_model
 
-NUM_SECONDS_TO_SLEEP = 5
+NUM_SECONDS_TO_SLEEP = int(os.getenv("NUM_SECONDS_TO_SLEEP", "5"))
+WORKERS = int(os.getenv("WORKERS", "32"))
 
 try:
     from vllm import LLM, SamplingParams
@@ -50,7 +51,7 @@ class VLLM(lmms):
         - VLLM chat method: https://docs.vllm.ai/en/stable/models/generative_models.html#llmchat
 
     Args:
-        model_version (str): HuggingFace model identifier or path to the model.
+        model (str): HuggingFace model identifier or path to the model.
             Default: "Qwen/Qwen2.5-VL-3B-Instruct"
         tensor_parallel_size (int): Number of GPUs to use for tensor parallelism.
             Default: 1
@@ -81,7 +82,7 @@ class VLLM(lmms):
             "--model",
             "vllm",
             "--model_args",
-            "model_version=meta-llama/Llama-4-Scout-17B-16E-Instruct,"
+            "model=meta-llama/Llama-4-Scout-17B-16E-Instruct,"
             "tensor_parallel_size=4,"
             "dtype=bfloat16,"
             "max_model_len=10240,"
@@ -118,7 +119,7 @@ class VLLM(lmms):
         "--model",
         "vllm",
         "--model_args",
-        "model_version=deepseek-ai/deepseek-vl2,"
+        "model=deepseek-ai/deepseek-vl2,"
         'hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]},' # example of passing model specific arguments, JSON string will be parsed automatically
         f"chat_template={chat_template_file}," # chat template file path
         "tensor_parallel_size=2,"
@@ -145,12 +146,12 @@ class VLLM(lmms):
 
     def __init__(
         self,
-        model_version: str = "Qwen/Qwen2.5-VL-3B-Instruct",
+        model: str = "Qwen/Qwen2.5-VL-3B-Instruct",
         tensor_parallel_size: int = 1,
+        data_parallel_size: int = 1,
         gpu_memory_utilization: float = 0.8,
         batch_size: int = 1,
         max_frame_num: int = 32,
-        threads: int = 16,  # Threads to use for decoding visuals
         trust_remote_code: Optional[bool] = True,
         chat_template: Optional[str] = None,
         min_image_pixels: int = 28,  # minimum image dimension, required for Qwen 2/2.5-VL models
@@ -160,12 +161,13 @@ def __init__(
         # Manually set a image token for GPT4V so that we can search for it
         # and split the text and image
         # Here we just use the same token as llava for convenient
-        self.model_version = model_version
+        self.model = model
         self.max_frame_num = max_frame_num
-        self.threads = threads
+        self.chat_template = chat_template
         self.min_image_pixels = min_image_pixels
+        self.data_parallel_size = data_parallel_size
         # Qwen 2/2.5-VL models enforce minimum image dimensions
-        self._enforce_image_resize = self._is_qwen_vl_model(model_version)
+        self._enforce_image_resize = self._is_qwen_vl_model(model)
 
         # Load chat template during initialization
         self.chat_template = None
@@ -191,13 +193,6 @@ def __init__(
 
         # Set up vllm client
         os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-        self.client = LLM(
-            model=self.model_version,
-            tensor_parallel_size=tensor_parallel_size,
-            gpu_memory_utilization=gpu_memory_utilization,
-            trust_remote_code=trust_remote_code,
-            **kwargs,
-        )
 
         accelerator = Accelerator()
         if accelerator.num_processes > 1:
@@ -211,13 +206,27 @@ def __init__(
             self.accelerator = accelerator
             self._rank = self.accelerator.local_process_index
             self._world_size = self.accelerator.num_processes
+        # TODO: Support tensor parallelism in the future for flexible vllm parallel
+        if data_parallel_size > 1:
+            assert tensor_parallel_size == 1, "Data parallelism is not supported with tensor parallelism. For current vllm version"
+        if accelerator.num_processes > 1:
+            kwargs["distributed_executor_backend"] = "external_launcher"
+        self.client = LLM(
+            model=self.model,
+            tensor_parallel_size=tensor_parallel_size,
+            gpu_memory_utilization=gpu_memory_utilization,
+            trust_remote_code=trust_remote_code,
+            disable_log_stats=False,
+            seed=1,
+            **kwargs,
+        )
 
         self.device = self.accelerator.device
         self.batch_size_per_gpu = int(batch_size)
 
-    def _is_qwen_vl_model(self, model_version: str) -> bool:
+    def _is_qwen_vl_model(self, model: str) -> bool:
         qwen_vl_patterns = ["qwen2-vl", "qwen2.5-vl"]
-        return any(pattern in model_version.lower() for pattern in qwen_vl_patterns)
+        return any(pattern in model.lower() for pattern in qwen_vl_patterns)
 
     def _maybe_resize_image(self, img: Image.Image) -> Image.Image:
         # edge‐case validation
@@ -294,16 +303,14 @@ def generate_until(self, requests) -> List[str]:
                 contexts, gen_kwargs, doc_to_visual, doc_id, task, split = batch_requests[idx].arguments
                 if "max_new_tokens" not in gen_kwargs:
                     gen_kwargs["max_new_tokens"] = 1024
-                if gen_kwargs["max_new_tokens"] > 4096:
-                    gen_kwargs["max_new_tokens"] = 4096
                 if "temperature" not in gen_kwargs:
                     gen_kwargs["temperature"] = 0
                 if "top_p" not in gen_kwargs:
                     gen_kwargs["top_p"] = 0.95
 
                 params = {
-                    "temperature": gen_kwargs["temperature"],
                     "max_tokens": gen_kwargs["max_new_tokens"],
+                    "temperature": gen_kwargs["temperature"],
                     "top_p": gen_kwargs["top_p"],
                 }
                 sampling_params = SamplingParams(**params)
@@ -316,7 +323,7 @@ def generate_until(self, requests) -> List[str]:
                     visuals = self.flatten(visuals)
                     imgs = []  # multiple images or frames for video
                     all_tasks = []
-                    with ThreadPoolExecutor(max_workers=self.threads) as executor:
+                    with ThreadPoolExecutor(max_workers=WORKERS) as executor:
                         for visual in visuals:
                             if isinstance(visual, str) and (".mp4" in visual or ".avi" in visual or ".mov" in visual or ".flv" in visual or ".wmv" in visual):
                                 all_tasks.append(executor.submit(self.encode_video, visual))
@@ -329,10 +336,10 @@ def generate_until(self, requests) -> List[str]:
                             imgs.append(task.result())
 
                 messages = [{"role": "user", "content": []}]
-                # When there is no image token in the context, append the image to the text
-                messages[0]["content"].append({"type": "text", "text": contexts})
+                # Add images first, then text
                 for img in self.flatten(imgs):
                     messages[0]["content"].append({"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img}"}})
+                messages[0]["content"].append({"type": "text", "text": contexts})
 
                 batched_messages.append(messages)
 

lmms_eval/tasks/k12/k12.yaml

@@ -12,11 +12,6 @@ generation_kwargs:
   until:
     - "ASSISTANT:"
   max_new_tokens: 16384
-  temperature: 0
-  top_p: 0.95
-  num_beams: 1
-  do_sample: false
-  repetition_penalty: 1.2
 process_results: !function utils.k12_process_results
 metric_list:
   - metric: llm_as_judge_eval

lmms_eval/tasks/mathverse/mathverse_evals.py

@@ -4,6 +4,7 @@
 import pandas as pd
 import requests
 from loguru import logger as eval_logger
+from openai import AzureOpenAI, OpenAI
 from tqdm import tqdm
 
 DEMO_PROMPT_EXTRACT = """
@@ -74,67 +75,45 @@
 
 class MathVerseEvaluator:
     API_TYPE = os.getenv("API_TYPE", "openai")
-
     if API_TYPE == "openai":
         API_URL = os.getenv("OPENAI_API_URL", "https://api.openai.com/v1/chat/completions")
         API_KEY = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY")
         headers = {
             "Authorization": f"Bearer {API_KEY}",
             "Content-Type": "application/json",
         }
+        client = OpenAI(api_key=API_KEY, base_url=API_URL.rstrip("chat/completions"))
+
     elif API_TYPE == "azure":
         API_URL = os.getenv("AZURE_ENDPOINT", "https://api.cognitive.microsoft.com/sts/v1.0/issueToken")
         API_KEY = os.getenv("AZURE_API_KEY", "YOUR_API_KEY")
-        headers = {
-            "api-key": API_KEY,
-            "Content-Type": "application/json",
-        }
+        API_VERSION = os.getenv("AZURE_API_VERSION", "2023-07-01-preview")
+        client = AzureOpenAI(azure_endpoint=API_URL, api_version=API_VERSION, api_key=API_KEY)
+        gpt_model = os.getenv("MODEL_VERSION", "gpt-4o-2024-11-20")
 
-    def __init__(self, api_key, gpt_model="gpt-3.5-turbo", quick_extract=False):
-        self.api_key = api_key
-        self.gpt_model = gpt_model
+    def __init__(self, quick_extract=False):
         self.quick_extract = quick_extract
 
-    def _post_request(self, payload):
-        headers = {
-            "Authorization": f"Bearer {self.api_key}",
-            "Content-Type": "application/json",
-        }
-        response = requests.post(self.API_URL, headers=headers, json=payload, timeout=30)
-        response.raise_for_status()
-        return response.json()
-
-    def get_chat_response(self, prompt, temperature=0, max_tokens=256, n=1, patience=10000000, sleep_time=0):
+    def get_chat_response(self, prompt, temperature=0, max_tokens=256, n=1, patience=5, sleep_time=0):
         messages = [
             {"role": "user", "content": prompt},
         ]
-        payload = {"model": self.gpt_model, "messages": messages, "temperature": temperature, "max_tokens": max_tokens, "n": n}
+        payload = {"model": self.gpt_model, "messages": messages, "temperature": temperature, "max_tokens": max_tokens}
 
         while patience > 0:
             patience -= 1
             try:
-                response = self._post_request(payload)
+                response = self.client.chat.completions.create(**payload)
                 if n == 1:
-                    prediction = response["choices"][0]["message"]["content"].strip()
+                    prediction = response.choices[0].message.content.strip()
                     if prediction and prediction != "":
                         return prediction
                 else:
-                    prediction = [choice["message"]["content"].strip() for choice in response["choices"]]
+                    prediction = [choice.message.content.strip() for choice in response.choices]
                     if prediction and prediction[0] != "":
                         return prediction
 
             except Exception as e:
-                # some model may output repetitive answer, which ChatGPT will throw an error.
-                if "repetitive patterns" in str(e):
-                    print(str(e))
-                    print("Continue with empty answer")
-                    return ""
-                # some answer may contain some sensitive words, like 'test'
-                if "sensitive" in str(e) or "400" in str(e):
-                    print(str(e))
-                    print("Continue with empty answer")
-                    return "0"
-
                 if "Rate limit" not in str(e):
                     eval_logger.error(e)
 

lmms_eval/tasks/mathverse/mathverse_testmini.yaml

@@ -11,11 +11,6 @@ doc_to_text: !function utils.mathverse_doc_to_text
 doc_to_target: "answer"
 generation_kwargs:
   max_new_tokens: 16384
-  temperature: 0.0
-  do_sample: true
-  top_p: 0.95
-  top_k: 50
-  repetition_penalty: 1.2
   until:
     - "</s>"
     - "Q:"
@@ -25,14 +20,11 @@ metric_list:
   - metric: gpt_eval_score
     aggregation: !function utils.mathverse_aggregate_results_eval
     higher_is_better: true
-  - metric: submission
-    aggregation: !function utils.mathverse_aggregate_results_submission
-    higher_is_better: true
-  
+
 lmms_eval_specific_kwargs:
   default:
     shot_type: "format-prompt" # can also be "custom-prompt"
     query_type: "query_wo" # now only support query_wo
 model_specific_generation_kwargs:
   llava:
-    image_aspect_ratio: original
+    image_aspect_ratio: original

lmms_eval/tasks/mathverse/utils.py

@@ -19,7 +19,7 @@
 
     config = yaml.safe_load("".join(safe_data))
 
-mathverse_evaluator = MathVerseEvaluator(api_key=os.getenv("OPENAI_API_KEY", "YOUR_API_KEY"), gpt_model=config["metadata"]["gpt_eval_model_name"])
+mathverse_evaluator = MathVerseEvaluator()
 
 
 def mathverse_doc_to_visual(doc):
@@ -82,21 +82,7 @@ def mathverse_aggregate_results_submission(results, args, *, calculate_gain=Fals
 def mathverse_aggregate_results_eval(results, args, *, calculate_gain=False, random_scores=None):
     split_flag = results[0]["metadata"]["split"]
     problem_version = results[0]["problem_version"].lower().replace(" ", "_")
-    # save the result first, in case the gpt evaluation fails
-    path = generate_submission_file(f"mathverse_{split_flag}_{problem_version}_results.json", args)
-    with open(path, "w") as f:
-        json.dump(results, f, indent=4)
-    # gpt evaluation
     results_dict, scores = mathverse_evaluator.eval_results(results, config)
-    # save results
-    path = generate_submission_file(f"mathverse_{split_flag}_{problem_version}_results.json", args)
-    with open(path, "w") as f:
-        json.dump(results_dict, f, indent=4)
-    # save scores
-    path = generate_submission_file(f"mathverse_{split_flag}_{problem_version}_scores.json", args)
-    with open(path, "w") as f:
-        json.dump(scores, f, indent=4)
-    eval_logger.info(f"Saved scores to {path}")
     if scores["average"]["accuracy"] == 0:
         return None
     return scores["average"]["accuracy"]