A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 21:00 UTC
[Bugfix] - Add Trace Headers to Beam Search Path (#29100)
Signed-off-by: dsuhinin <suhinin.dmitriy@gmail.com>
Software Developer committed
Nov 20, 2025 at 21:00 UTC
4d01b6428448225807e6605d04e37e29fe729b44
3 files changed
+4
vllm/entrypoints/openai/serving_chat.py
+1
@@ -319,6 +319,7 @@ class OpenAIServingChat(OpenAIServing):
319
request_id=request_id,
320
params=sampling_params,
321
lora_request=lora_request,
322
+ trace_headers=trace_headers,
323
)
324
else:
325
engine_request, tokenization_kwargs = await self._process_inputs(
vllm/entrypoints/openai/serving_completion.py
+1
@@ -216,6 +216,7 @@ class OpenAIServingCompletion(OpenAIServing):
216
request_id=request_id,
217
params=sampling_params,
218
lora_request=lora_request,
219
+ trace_headers=trace_headers,
220
)
221
else:
222
engine_request, tokenization_kwargs = await self._process_inputs(
vllm/entrypoints/openai/serving_engine.py
+2
@@ -343,6 +343,7 @@ class OpenAIServing:
343
request_id: str,
344
params: BeamSearchParams,
345
lora_request: LoRARequest | None = None,
346
+ trace_headers: Mapping[str, str] | None = None,
347
) -> AsyncGenerator[RequestOutput, None]:
348
beam_width = params.beam_width
349
max_tokens = params.max_tokens
@@ -437,6 +438,7 @@ class OpenAIServing:
438
beam_search_params,
439
request_id_item,
440
lora_request=lora_req,
441
+ trace_headers=trace_headers,
442
)
443
)
444
)