Debug stats

run_stats.py

@@ -0,0 +1,38 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# hf download Qwen/Qwen3-0.6B --local-dir /data/local/models/qwen3_06b
+# rm /data/local/models/qwen3_06b/tokenizer*
+
+from vllm import LLM, SamplingParams, TokensPrompt
+from vllm.sampling_params import RequestOutputKind
+
+prompts = [
+    "Hello, my name is",
+    "The president of the United States is",
+    # "The capital of France is",
+    # "The future of AI is",
+]
+sampling_params = SamplingParams(
+    temperature=0.8,
+    top_p=0.95,
+    max_tokens=500,
+    stop="is",
+    n=2,
+    output_kind=RequestOutputKind.CUMULATIVE,
+    # stop_token_ids=[5],
+)
+tokens_prompt = TokensPrompt(prompt_token_ids=[2, 3, 4], )
+
+if __name__ == "__main__":
+    llm = LLM(
+        model="Qwen/Qwen3-0.6B",
+        enforce_eager=True,
+        # skip_tokenizer_init=True,
+        gpu_memory_utilization=0.8,
+        disable_log_stats=False,
+    )
+    outputs = llm.generate(prompts=prompts, sampling_params=sampling_params)
+    for output in outputs:
+        prompt = output.prompt_token_ids
+        generated_token_ids = output.outputs[0].token_ids
+        print(f"Prompt: {prompt!r}, Generated tokens: {generated_token_ids!r}")

vllm/entrypoints/llm.py

@@ -46,8 +46,7 @@
                           ScoringRequestOutput)
 from vllm.plugins.io_processors import get_io_processor
 from vllm.pooling_params import PoolingParams
-from vllm.sampling_params import (BeamSearchParams, RequestOutputKind,
-                                  SamplingParams)
+from vllm.sampling_params import BeamSearchParams, SamplingParams
 from vllm.tasks import PoolingTask
 from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer,
                                                get_cached_tokenizer,
@@ -90,7 +89,7 @@ class LLM:
             or videos from directories specified by the server file system.
             This is a security risk. Should only be enabled in trusted
             environments.
-        allowed_media_domains: If set, only media URLs that belong to this 
+        allowed_media_domains: If set, only media URLs that belong to this
             domain can be used for multi-modal inputs.
         tensor_parallel_size: The number of GPUs to use for distributed
             execution with tensor parallelism.
@@ -1504,10 +1503,10 @@ def _validate_and_add_requests(
             raise ValueError("The lengths of prompts and lora_request "
                              "must be the same.")
 
-        for sp in params if isinstance(params, Sequence) else (params, ):
-            if isinstance(sp, SamplingParams):
-                # We only care about the final output
-                sp.output_kind = RequestOutputKind.FINAL_ONLY
+        # for sp in params if isinstance(params, Sequence) else (params, ):
+        #     if isinstance(sp, SamplingParams):
+        #         # We only care about the final output
+        #         sp.output_kind = RequestOutputKind.FINAL_ONLY
 
         # Add requests to the engine.
         it = prompts

vllm/v1/engine/llm_engine.py

@@ -274,6 +274,7 @@
 
         # 1) Get EngineCoreOutput from the EngineCore.
         outputs = self.engine_core.get_output()
+        logger.info(f"{len(outputs.outputs)=}")
 
         # 2) Process EngineCoreOutputs.
         iteration_stats = IterationStats() if self.log_stats else None
@@ -294,6 +295,9 @@
             )
             self.do_log_stats_with_interval()
 
+        logger.info(
+            f"EngineCore: {outputs.scheduler_stats=}, {iteration_stats=}, {processed_outputs.request_outputs=}"
+        )
         return processed_outputs.request_outputs
 
     def get_vllm_config(self):

vllm/v1/engine/output_processor.py

@@ -8,6 +8,7 @@
 
 import torch
 
+from vllm.logger import init_logger
 from vllm.outputs import (CompletionOutput, PoolingOutput,
                           PoolingRequestOutput, RequestOutput)
 from vllm.sampling_params import RequestOutputKind
@@ -22,6 +23,8 @@
 from vllm.v1.metrics.stats import (IterationStats, LoRARequestStates,
                                    RequestStateStats)
 
+logger = init_logger(__name__)
+
 
 class RequestOutputCollector:
     """
@@ -190,6 +193,9 @@
         kv_transfer_params: Optional[dict[str, Any]] = None,
     ) -> Optional[Union[RequestOutput, PoolingRequestOutput]]:
 
+        logger.info(
+            f"{self.request_id=}, {self.output_kind=}, {finish_reason=}, {stop_reason=}"
+        )
         finished = finish_reason is not None
         final_only = self.output_kind == RequestOutputKind.FINAL_ONLY
 
@@ -446,6 +452,9 @@
             if request_output := req_state.make_request_output(
                     new_token_ids, pooling_output, finish_reason, stop_reason,
                     kv_transfer_params):
+                assert isinstance(request_output, RequestOutput)
+                logger.info(
+                    f"Request {req_id} metrics: {request_output.metrics}")
                 if req_state.queue is not None:
                     # AsyncLLM: put into queue for handling by generate().
                     req_state.queue.put(request_output)