updated

Signed-off-by: Robert Shaw <rshaw@neuralmagic.com>

vllm/v1/engine/core.py

@@ -1050,6 +1050,7 @@ class DPEngineCoreProc(EngineCoreProc):
             self._maybe_publish_request_counts()
 
             local_unfinished_reqs = self.scheduler.has_unfinished_requests()
+            logger.info(f"{local_unfinished_reqs=}")
             if not executed:
                 if not local_unfinished_reqs and not self.engines_running:
                     # All engines are idle.
@@ -1057,6 +1058,7 @@ class DPEngineCoreProc(EngineCoreProc):
 
                 # We are in a running state and so must execute a dummy pass
                 # if the model didn't execute any ready requests.
+                logger.info("===EXECUTE_DUMMY_BATCH===")
                 self.execute_dummy_batch()
 
             # 3) All-reduce operation to determine global unfinished reqs.

vllm/v1/engine/core_client.py

@@ -1077,7 +1077,7 @@ class DPAsyncMPClient(AsyncMPClient):
                     if counts is not None:
                         sliced_counts = counts[count_slice]
                         self.lb_engines = sliced_counts
-                        logger.debug("Received counts: %s (%s)", sliced_counts,
+                        logger.debug("Received counts: %s (%s)", counts,
                                      count_slice)
 
         resources.stats_update_task = asyncio.create_task(

vllm/v1/worker/gpu_model_runner.py

@@ -2400,6 +2400,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
 
         # Run the model.
         # Use persistent buffers for CUDA graphs.
+        logger.info(f"====== EXECUTE {ubatch_slices=}, {num_input_tokens=}, {num_tokens_across_dp=}")
         with (set_forward_context(
                 attn_metadata,
                 self.vllm_config,
@@ -3046,6 +3047,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                 (1 token) and prefill (multiple tokens) requests.
             remove_lora: If False, dummy LoRAs are not destroyed after the run
         """
+        logger.info("====== DUMMY RUN")
         assert cudagraph_runtime_mode is None or \
             cudagraph_runtime_mode.valid_runtime_modes()
 

vllm/v1/worker/ubatch_splitting.py

@@ -167,6 +167,7 @@ def ubatch_split(
         num_tokens_unpadded,
         uniform_decode=uniform_decode,
     )
+    logger.info(f"==== {should_attempt_ubatching=}, {num_tokens_unpadded=}")
 
     # Don't microbatch unless every other DP worker is also microbatching
     should_ubatch, num_tokens_after_padding = get_dp_padding_ubatch(
@@ -175,6 +176,7 @@ def ubatch_split(
         should_attempt_ubatching,
         vllm_config,
     )
+    logger.info(f"==== {should_ubatch=}, {num_tokens_after_padding=}")
 
     if not should_ubatch:
         return (None, None)