[Core] Replace routing replay with device cache and async D2H pipeline (#39917)

Signed-off-by: Tomer Barnatan <tbarnatan@nvidia.com>
This commit is contained in:
TomerBN-Nvidia
2026-05-07 11:24:56 -07:00
committed by GitHub
parent 8eb401134e
commit 8189a15914
15 changed files with 1398 additions and 642 deletions
@@ -542,6 +542,11 @@ class OpenAIServingCompletion(OpenAIServing):
token_ids=(
as_list(output.token_ids) if request.return_token_ids else None
),
routed_experts=(
output.routed_experts.tolist()
if output.routed_experts is not None
else None
),
)
choices.append(choice_data)
@@ -565,8 +570,13 @@ class OpenAIServingCompletion(OpenAIServing):
)
request_metadata.final_usage_info = usage
prompt_routed_experts = None
if final_res_batch:
kv_transfer_params = final_res_batch[0].kv_transfer_params
pre = final_res_batch[0].prompt_routed_experts
if pre is not None:
prompt_routed_experts = pre.tolist()
return CompletionResponse(
id=request_id,
created=created_time,
@@ -575,6 +585,7 @@ class OpenAIServingCompletion(OpenAIServing):
usage=usage,
system_fingerprint=self.system_fingerprint,
kv_transfer_params=kv_transfer_params,
prompt_routed_experts=prompt_routed_experts,
)
def _create_completion_logprobs(