[Model][Core] Enable async_chunk streaming pipeline for CosyVoice3#1703

Merged

linyueqian merged 17 commits into

vllm-project:mainfrom

indevn:feat/cosyvoice3-async-batch-benchmark-signedoff

Apr 5, 2026

Commits on Apr 5, 2026

[CosyVoice3][async_chunk] Enable async pipeline and batch-safe code2wav hardening

indevn
authored and
linyueqian
committed
[CosyVoice3][async_chunk] Fix AR cleanup path and add code2wav trim diagnostics

indevn
authored and
linyueqian
committed
[CosyVoice3][async_chunk] Fix cleanup-save race by moving reclaim to post-send

indevn
authored and
linyueqian
committed
[CosyVoice3][async_chunk] Fix duplicate terminal emission and clamp chunk audio span

indevn
authored and
linyueqian
committed
[CosyVoice3][async_chunk] Honor single-request seq_token_counts in code2wav slicing

indevn
authored and
linyueqian
committed
refactor: clarify multimodal output fallback chain

indevn
authored and
linyueqian
committed
fix: align cosyvoice3 async streaming and sampler behavior

indevn
authored and
linyueqian
committed
test: add cosyvoice3 unit coverage for streaming helpers

indevn
authored and
linyueqian
committed
test: add cosyvoice3 offline e2e coverage

indevn
authored and
linyueqian
committed
fix: clarify chunk transfer external request id handling

indevn
authored and
linyueqian
committed
fix: align cosyvoice3 async_chunk e2e with current runtime

indevn
authored and
linyueqian
committed
test: rename cosyvoice3 e2e around reference zero-shot semantics

indevn
authored and
linyueqian
committed
fix: align async_chunk config with vLLM 0.19.0 and enable streaming test
linyueqian
committed
fix: correct CosyVoice3 stop token handling and enforce min_tokens
linyueqian
committed
test: disable CUDA graphs in async_chunk e2e test for faster CI
linyueqian
committed
fix: update talker mock model_stage and config in unit tests
linyueqian
committed
fix: set enforce_eager=true in async_chunk YAML to avoid CI timeout
linyueqian
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Model][Core] Enable async_chunk streaming pipeline for CosyVoice3#1703

[Model][Core] Enable async_chunk streaming pipeline for CosyVoice3#1703
linyueqian merged 17 commits into
vllm-project:mainfrom
indevn:feat/cosyvoice3-async-batch-benchmark-signedoff

Commits on Apr 5, 2026

[CosyVoice3][async_chunk] Enable async pipeline and batch-safe code2wav hardening

[CosyVoice3][async_chunk] Fix AR cleanup path and add code2wav trim diagnostics

[CosyVoice3][async_chunk] Fix cleanup-save race by moving reclaim to post-send

[CosyVoice3][async_chunk] Fix duplicate terminal emission and clamp chunk audio span

[CosyVoice3][async_chunk] Honor single-request seq_token_counts in code2wav slicing

refactor: clarify multimodal output fallback chain

fix: align cosyvoice3 async streaming and sampler behavior

test: add cosyvoice3 unit coverage for streaming helpers

test: add cosyvoice3 offline e2e coverage

fix: clarify chunk transfer external request id handling

fix: align cosyvoice3 async_chunk e2e with current runtime

test: rename cosyvoice3 e2e around reference zero-shot semantics

fix: align async_chunk config with vLLM 0.19.0 and enable streaming test

fix: correct CosyVoice3 stop token handling and enforce min_tokens

test: disable CUDA graphs in async_chunk e2e test for faster CI

fix: update talker mock model_stage and config in unit tests

fix: set enforce_eager=true in async_chunk YAML to avoid CI timeout