Skip to content

[Perf][Fish Speech] Free unused DAC codec components to save VRAM #2430

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
linyueqian merged 2 commits into from
Apr 5, 2026
Merged

[Perf][Fish Speech] Free unused DAC codec components to save VRAM #2430

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7bd0dd7
[Perf][Fish Speech] Free unused DAC codec components to save ~1.2 GiB…
Sy0307 Apr 1, 2026
e39bcd7
Merge branch 'main' into perf/fish-speech-dac-free-unused-components
Sy0307 Apr 4, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions vllm_omni/model_executor/models/fish_speech/dac_encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,9 @@ def _load_dac_codec(
if "generator" in state_dict:
state_dict = state_dict["generator"]
codec.load_state_dict(state_dict, strict=False)
# Encoder path only uses encoder + quantizer.forward(); prune the
# decoder before moving to device to avoid unnecessary GPU allocation.
codec.decoder = None
codec = codec.to(device=device, dtype=dtype)
codec.eval()

Expand Down
7 changes: 7 additions & 0 deletions vllm_omni/model_executor/models/fish_speech/fish_speech_dac_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,6 +141,13 @@ def _ensure_codec_loaded(self) -> None:
self._bake_weight_norm(codec)
self._cache_attention_masks(codec)

# Decode path only uses quantizer.decode() + decoder; prune
# encode-only components before moving to device to avoid
# unnecessary GPU allocation.
codec.encoder = None
codec.quantizer.pre_module = None
codec.quantizer.downsample = None

device = self.vllm_config.device_config.device
codec = codec.to(device=device, dtype=torch.float32)
codec.eval()
Expand Down
Loading