chore: improve mmmu benchmark (#7000)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

benchmark/mmmu/bench_sglang.py

@@ -125,7 +125,6 @@ async def eval_mmmu(args) -> None:
     client = openai.AsyncOpenAI(
         api_key="sk", base_url=f"http://127.0.0.1:{args.port}/v1"
     )
-    semaphore = asyncio.Semaphore(args.concurrency)
     start = time.perf_counter()
     base_url = f"http://127.0.0.1:{args.port}"
 
@@ -139,16 +138,26 @@ async def eval_mmmu(args) -> None:
 
         samples = samples[: args.profile_number]
 
-    tasks = [
-        process_sample_with_semaphore(
-            semaphore, client, sample, sampling_params, lora_path
-        )
-        for sample in samples
-    ]
-
-    for coro in tqdm(asyncio.as_completed(tasks), total=len(tasks)):
-        sample, response = await coro
-        process_result(response, sample, answer_dict, out_samples)
+    if args.concurrency == 1:
+        # For concurrency == 1, run in sequential mode to ensure consistent order
+        # this is mainly for profiling
+        for sample in tqdm(samples):
+            _, response = await process_sample(
+                client, sample, sampling_params, lora_path
+            )
+            process_result(response, sample, answer_dict, out_samples)
+    else:
+        semaphore = asyncio.Semaphore(args.concurrency)
+        tasks = [
+            process_sample_with_semaphore(
+                semaphore, client, sample, sampling_params, lora_path
+            )
+            for sample in samples
+        ]
+
+        for coro in tqdm(asyncio.as_completed(tasks), total=len(tasks)):
+            sample, response = await coro
+            process_result(response, sample, answer_dict, out_samples)
 
     if args.profile:
         print("Stopping profiler...")

benchmark/mmmu/eval_utils.py

@@ -27,8 +27,7 @@ from tqdm import tqdm
 class EvalArgs:
     seed: int = 42
     split: str = "validation"
-    # Default setting to make the benchmark available on A100 for most 7B models
-    image_pixels_limit: int = 4300000
+    image_pixels_limit: int = -1
     result_filename: str = ""
     prompt_format_file: str = "prompt_format.yaml"
     dataset_path: str = "MMMU/MMMU"
@@ -190,7 +189,7 @@ def prepare_samples(eval_args: EvalArgs):
         sample = construct_prompt(sample, eval_args.config)
         image = sample["image"]
         width, height = image.size
-        if width * height >= eval_args.image_pixels_limit:
+        if 0 < eval_args.image_pixels_limit <= width * height:
             return None, True
         # Use a unique identifier for the image path to avoid potential collisions if indices reset
         image_path = f"{images_path}/image_{sample['id']}.png"
@@ -217,6 +216,8 @@ def prepare_samples(eval_args: EvalArgs):
             elif sample:
                 samples.append(sample)
 
+    samples.sort(key=lambda x: x["final_input_prompt"])
+
     print(
         f"Skipping {skip_count} samples with large images, {round((float(skip_count) / len(dataset)) * 100, 2)}% of dataset"
     )