[RL][Frontend] Return weight-version spans on OpenAI completions and chat (#61347)

Signed-off-by: Shuolei Wang <shuoleiwang123@gmail.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
This commit is contained in:
Shuolei Wang
2026-10-17 02:08:07 +00:00
committed by GitHub
co-authored by aoshen02
parent 6b6a6c5c98
commit b3a7e2ff6e
7 changed files with 139 additions and 4 deletions
@@ -651,17 +651,21 @@ class OpenAIServingChat(GenerateBaseServing):
delta_text = output.text
if (
nothing_generated = (
not delta_text
and not output.token_ids
and not previous_num_tokens[i]
):
)
if nothing_generated and output.finish_reason is None:
# Chunked prefill case, don't return empty chunks
continue
delta_message: DeltaMessage | None
if parser is not None:
if nothing_generated:
# Finished without output: send the final chunk unparsed.
delta_message = DeltaMessage()
elif parser is not None:
delta_message = parser.parse_delta(
delta_text=delta_text,
delta_token_ids=as_list(output.token_ids),
@@ -773,6 +777,11 @@ class OpenAIServingChat(GenerateBaseServing):
token_ids=(
as_list(output.token_ids) if include_token_ids else None
),
weight_versions=(
output.weight_versions
if request.return_weight_versions
else None
),
)
finish_reason_sent[i] = True
@@ -1097,6 +1106,9 @@ class OpenAIServingChat(GenerateBaseServing):
else None
),
routed_experts=routed_experts_b64,
weight_versions=(
output.weight_versions if request.return_weight_versions else None
),
)
choice_data = maybe_filter_parallel_tool_calls(choice_data, request)