Skip to content

SOTA Open Source TTS

Fix top-1 guard in logits_to_probs for batched logits (#1326)

Rishabh Jainc164eb5a4417
diff --git a/fish_speech/models/text2semantic/inference.py b/fish_speech/models/text2semantic/inference.py
index 59653d5..63b977c 100644
--- a/fish_speech/models/text2semantic/inference.py
+++ b/fish_speech/models/text2semantic/inference.py
@@ -63,7 +63,7 @@ def logits_to_probs(
indices = torch.arange(sorted_logits.shape[-1], device=sorted_logits.device)
top_k_mask = indices >= top_k
sorted_indices_to_remove = (cum_probs > top_p) | top_k_mask
- sorted_indices_to_remove[0] = False # 单元素修改问题不大,或者写成 | (indices != 0)
+ sorted_indices_to_remove[..., 0] = False # keep the top-1 token, per row
indices_to_remove = sorted_indices_to_remove.scatter(
dim=-1, index=sorted_indices, src=sorted_indices_to_remove