[perf] Reuse a pooled HTTP session for multimodal URL downloads (#28056)

Signed-off-by: Abatom <abzhonghua@gmail.com>

python/sglang/srt/multimodal/processors/mimo_audio.py

@@ -10,9 +10,9 @@ from typing import Optional
 
 import numpy as np
 import pybase64
-import requests
 import torch
 
+from sglang.srt.utils import common
 from sglang.utils import logger
 
 try:
@@ -138,8 +138,6 @@ class MiMoAudioPipeline:
         )
         self._resamplers_max = max_resamplers
 
-        self.http_session = requests.Session()
-
     @property
     def audio_token_per_second(self) -> float:
         return self.audio_input_id_per_second / self.audio_group_size
@@ -188,18 +186,18 @@ class MiMoAudioPipeline:
                     dl_start = time.perf_counter()
                     timeout = int(os.getenv("REQUEST_TIMEOUT", "5"))
                     try:
-                        response = self.http_session.get(
+                        with common.get_mm_http_session().get(
                             audio, stream=True, timeout=timeout
-                        )
+                        ) as response:
-                        dl_elapsed_ms = (time.perf_counter() - dl_start) * 1000
+                            response.raise_for_status()
-                        if dl_elapsed_ms > 1000.0:
+                            dl_elapsed_ms = (time.perf_counter() - dl_start) * 1000
-                            content_len = len(response.content)
+                            if dl_elapsed_ms > 1000.0:
-                            logger.warning(
+                                content_len = len(response.content)
-                                f"Slow audio download: {dl_elapsed_ms:.2f}ms, "
+                                logger.warning(
-                                f"size={content_len / 1024:.1f}KB, url={audio}"
+                                    f"Slow audio download: {dl_elapsed_ms:.2f}ms, "
-                            )
+                                    f"size={content_len / 1024:.1f}KB, url={audio}"
-                        file = io.BytesIO(response.content)
+                                )
-                        response.close()
+                            file = io.BytesIO(response.content)
                     except Exception as e:
                         dl_elapsed_ms = (time.perf_counter() - dl_start) * 1000
                         logger.error(

python/sglang/srt/utils/common.py

@@ -782,6 +782,22 @@ def set_random_seed(seed: int) -> None:
         torch.xpu.manual_seed_all(seed)
 
 
+_mm_http_session = threading.local()
+
+
+def get_mm_http_session() -> requests.Session:
+    """Per-thread HTTP session for multimodal downloads, to pool/reuse TCP
+    connections. Pid-checked so a forked worker rebuilds its own, not the parent's.
+    """
+    pid = os.getpid()
+    session = getattr(_mm_http_session, "session", None)
+    if session is None or getattr(_mm_http_session, "pid", None) != pid:
+        session = requests.Session()
+        _mm_http_session.session = session
+        _mm_http_session.pid = pid
+    return session
+
+
 def load_audio(
     audio_file: str, sr: Optional[int] = None, mono: bool = True
 ) -> np.ndarray:
@@ -797,7 +813,7 @@ def load_audio(
         audio_file.startswith("http://") or audio_file.startswith("https://")
     ):
         timeout = int(os.getenv("REQUEST_TIMEOUT", "5"))
-        with requests.get(audio_file, timeout=timeout) as response:
+        with get_mm_http_session().get(audio_file, timeout=timeout) as response:
             response.raise_for_status()
             source = response.content
     elif isinstance(audio_file, str) and audio_file.startswith("file://"):
@@ -958,7 +974,7 @@ def get_image_bytes(image_file: Union[str, bytes]) -> bytes:
         return image_file
     if image_file.startswith(("http://", "https://")):
         timeout = int(os.getenv("REQUEST_TIMEOUT", "3"))
-        response = requests.get(image_file, timeout=timeout)
+        response = get_mm_http_session().get(image_file, timeout=timeout)
         try:
             response.raise_for_status()
             result = response.content
@@ -990,9 +1006,11 @@ def _normalize_video_input(
     elif isinstance(video_file, str):
         if video_file.startswith(("http://", "https://")):
             timeout = int(os.getenv("REQUEST_TIMEOUT", "10"))
-            response = requests.get(video_file, stream=True, timeout=timeout)
+            with get_mm_http_session().get(
-            response.raise_for_status()
+                video_file, stream=True, timeout=timeout
-            return response.content
+            ) as response:
+                response.raise_for_status()
+                return response.content
         elif video_file.startswith("data:"):
             _, encoded = video_file.split(",", 1)
             return pybase64.b64decode(encoded, validate=True)