[2/N] Add a minimal temporal chunk callback for MiniMax-H3 (#7017)

Signed-off-by: specture724 <specture724@gmail.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Anjie HouandClaude Sonnet 5 authored and GitHub committed 2026-09-09 16:40:32 +08:00
1 parent ffa991b24f
commit ecdd762ff5
6 files changed
+641 -27

No files matched your search

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
from contextlib import nullcontext
from typing import Any
from typing import Any, ClassVar
import torch
import torch.distributed as dist
@@ -62,6 +62,10 @@ class OmniAutoencoderKLWan(AutoencoderKLWan):
return self._decode_temporal(z, return_dict=return_dict)
# Chunks are published straight from the decoder, clamped to the
# checkpoint's output range.
chunk_value_range: ClassVar[tuple[float, float]] = (-1.0, 1.0)
@apply_forward_hook
def decode_with_chunks(self, z: torch.Tensor, *, on_chunk: DecodedChunkConsumer) -> None:
"""Decode ``z`` while synchronously delivering temporal chunks.