feat: pp 增加输出每个micro batch调度的请求数功能，VLLM_PP_DEBUG打开，默认关闭

vllm/envs.py

@@ -189,6 +189,7 @@ if TYPE_CHECKING:
     VLLM_USE_FUSE_SILU_AND_MUL: bool = False
     VLLM_USE_OPT_RESHAPE_AND_CACHE: bool = False
     VLLM_USE_TOPK_RENORM: bool = False
+    VLLM_PP_DEBUG: bool = False
 
 def get_default_cache_root():
     return os.getenv(
@@ -1237,6 +1238,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
         lambda:
         (os.environ.get("VLLM_USE_TOPK_RENORM", "True").lower() in
                 ("true", "1")),
+    "VLLM_PP_DEBUG":
+        lambda:
+        (os.environ.get("VLLM_PP_DEBUG", "False").lower() in
+         ("true", "1")),
         
 }
 

vllm/v1/engine/core.py

@@ -275,6 +275,12 @@ class EngineCore:
                 pass
             scheduler_output = self.scheduler.schedule()
             if scheduler_output.total_num_scheduled_tokens > 0:
+                if envs.VLLM_PP_DEBUG:
+                    import sys,os
+                    num_run_reqs = len(scheduler_output.scheduled_new_reqs) + scheduler_output.scheduled_cached_reqs.num_reqs
+                    sys.stderr.write(f"[pid- {os.getpid()}]running requests in micro batch is:{num_run_reqs}, "
+                                     f"total_num_scheduled_tokens is {scheduler_output.total_num_scheduled_tokens}\n")
+                    sys.stderr.flush()
                 future = self.model_executor.execute_model(scheduler_output)
                 self.batch_queue.put_nowait(
                     (future, scheduler_output))  # type: ignore