Taichang ZhouandClaude Opus 4.8
9678ae58cf
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 Collapses the 46-commit dev/migrate-MR-v2 history (10 merge commits + superseded pre-rebase work) into a single commit, so the diff reflects only the genuine MR-V2 feature (worker_v2/ runners + model_states, omni schedulers, stage engine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>
34 lines
1.0 KiB
Python
34 lines
1.0 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
"""Forward-call compatibility helpers for Omni MR V2 runners."""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any
|
|
|
|
|
|
def add_forward_compat_kwargs(
|
|
model_inputs: dict[str, Any],
|
|
input_batch: Any,
|
|
sampler: Any,
|
|
) -> None:
|
|
"""Add Omni model forward kwargs that V1 model runners provided.
|
|
|
|
Some Omni model stages inspect these kwargs inside ``forward`` rather
|
|
than only in the runner. Qwen2.5-Omni's talker, for example, rewrites
|
|
text-model pad ids in ``sampling_metadata`` to a talker-vocab pad id
|
|
before sampling codec tokens.
|
|
"""
|
|
model_inputs.setdefault(
|
|
"sampling_metadata",
|
|
getattr(input_batch, "sampling_metadata", None),
|
|
)
|
|
logits_index = getattr(
|
|
input_batch,
|
|
"logits_indices",
|
|
getattr(input_batch, "logits_index", None),
|
|
)
|
|
model_inputs.setdefault("logits_index", logits_index)
|
|
model_inputs.setdefault("sampler", sampler)
|