gitea-mirror/F5-TTS
1
0
Fork 0
mirror of https://github.com/SWivid/F5-TTS.git synced 2025-12-12 15:50:07 -08:00
Code Issues Packages Projects Releases Wiki Activity

VRAM usage on long texts gradually uses up memory.

Browse Source
This commit is contained in:
niknah
2025-03-14 16:53:58 +11:00
parent 128f4e4bf3
commit 48d1a9312e
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/f5_tts/infer/utils_infer.py
View File

@@ -479,6 +479,7 @@ def infer_batch_process(
cfg_strength=cfg_strength,
sway_sampling_coef=sway_sampling_coef,
)
del _
generated = generated.to(torch.float32)
generated = generated[:, ref_audio_len:, :]
@@ -490,6 +491,8 @@ def infer_batch_process(
if rms < target_rms:
generated_wave = generated_wave * rms / target_rms
del generated
# wav -> numpy
generated_wave = generated_wave.squeeze().cpu().numpy()
@@ -497,7 +500,9 @@ def infer_batch_process(
for j in range(0, len(generated_wave), chunk_size):
yield generated_wave[j : j + chunk_size], target_sample_rate
else:
yield generated_wave, generated_mel_spec[0].cpu().numpy()
generated_cpu = generated_mel_spec[0].cpu().numpy()
del generated_mel_spec
yield generated_wave, generated_cpu
if streaming:
for gen_text in progress.tqdm(gen_text_batches) if progress is not None else gen_text_batches: