[MR V2] Migrate to Model Runner V2

Squashed migration rebased onto current main. Collapses the 46-commit
dev/migrate-MR-v2 history (10 merge commits + superseded pre-rebase work)
into a single commit, so the diff reflects only the genuine MR-V2 feature
(worker_v2/ runners + model_states, omni schedulers, stage engine pipeline,
pooling_output serialization bridge). Content already merged to main
separately (step_audio2, tts_adapters, npu/_310p, etc.) drops out.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Taichang ZhouandClaude Opus 4.8 committed 2026-06-20 07:19:40 +00:00
1 parent cead615288
commit 9678ae58cf
74 files changed
+8244 -95

No files matched your search

+7
View File
@@ -121,6 +121,13 @@ class OmniEngineCoreOutput(EngineCoreOutput):
# pooling_output is inherited from EngineCoreOutput as torch.Tensor | None
# and retains its original vLLM semantics for pooling/embedding tasks.
multimodal_output: dict[str, torch.Tensor] | None = None
# Bytes-serialized carrier for a dict-shaped pooling_output (MR V2 cross-stage
# hidden/codec handoff). vLLM 0.23 strictly decodes ``pooling_output`` as a
# ``torch.Tensor``, so a per-request dict payload cannot ride that field. The
# producer (OmniARScheduler) sets ``pooling_output=None`` and serializes the
# dict here; the consumer (StagePool._poll_stage_raw) rehydrates it back into
# ``pooling_output`` after decode. Tensor pooling outputs are unaffected.
pooling_output_payload: AdditionalInformationPayload | None = None
# Finished flag for streaming input segment
is_segment_finished: bool | None = False
# Streaming update prompt length