Merge remote-tracking branch 'origin/main' into fix_image_api

Author: sangchengmeng

lightllm/models/deepseek2/model.py

@@ -74,6 +74,7 @@ def _init_some_value(self):
 
     def _init_custom(self):
         self._init_to_get_yarn_rotary()
+        dist_group_manager.new_deepep_group(self.config["n_routed_experts"])
 
     def _verify_params(self):
         return super()._verify_params()

lightllm/server/multimodal_params.py

@@ -29,7 +29,8 @@ async def preload(self, request: Request):
         try:
             if self._type == "url":
                 timeout = int(os.getenv("REQUEST_TIMEOUT", "5"))
-                img_data = await fetch_image(self._data, request, timeout=timeout)
+                proxy = os.getenv("REQUEST_PROXY", None)
+                img_data = await fetch_image(self._data, request, timeout=timeout, proxy=proxy)
             elif self._type == "base64":
                 img_data = base64.b64decode(self._data)
             elif self._type == "image_size":

lightllm/server/router/model_infer/mode_backend/base_backend.py

@@ -192,7 +192,6 @@ def init_model(self, kvargs):
             elif self.model_type == "phi3":
                 self.model = Phi3TpPartModel(model_kvargs)
             elif self.model_type in ["deepseek_v2", "deepseek_v3"]:
-                dist_group_manager.new_deepep_group(model_cfg["n_routed_experts"])
                 self.model = Deepseek2TpPartModel(model_kvargs)
             elif self.model_type == "internvl_chat":
                 llm_model_type = model_cfg.get("llm_config").get("model_type")

lightllm/server/router/model_infer/mode_backend/dp_backend/impl.py

@@ -31,7 +31,7 @@ def init_custom(self):
         # nan 值，避免后续构建的fake请求在计算的过程中出现计算错误。
         from .pre_process import padded_prepare_prefill_inputs
 
-        kwargs, run_reqs, padded_req_num = padded_prepare_prefill_inputs([], 1, is_multimodal=False)
+        kwargs, run_reqs, padded_req_num = padded_prepare_prefill_inputs([], 1, is_multimodal=self.is_multimodal)
         self.model.forward(**kwargs)
         assert len(run_reqs) == 0 and padded_req_num == 1
         return

lightllm/utils/image_utils.py

@@ -19,10 +19,10 @@ def image2base64(img_str: str):
     return base64.b64encode(buffer.getvalue()).decode("utf-8")
 
 
-async def fetch_image(url, request: Request, timeout):
+async def fetch_image(url, request: Request, timeout, proxy=None):
     logger.info(f"Begin to download image from url: {url}")
     start_time = time.time()
-    async with httpx.AsyncClient() as client:
+    async with httpx.AsyncClient(proxy=proxy) as client:
         async with client.stream("GET", url, timeout=timeout) as response:
             response.raise_for_status()
             ans_bytes = []