Skip to content

SOTA Open Source TTS

Optimize inifinity generation

Lengyuee7e7a8c2e39b
diff --git a/tools/llama/generate.py b/tools/llama/generate.py
index 3a02bd6..ec9a549 100644
--- a/tools/llama/generate.py
+++ b/tools/llama/generate.py
@@ -109,7 +109,7 @@ def decode_one_token(
# Disable <s> and </s> tokens for codebooks
if model.config.num_codebooks != 0:
- logits.codebook_logits[:, :, :, :2] = -float("Inf")
+ logits.codebook_logits[:, :, :, :1] = -float("Inf")
for i in range(model.config.num_codebooks):
codebooks.append(
@@ -194,7 +194,7 @@ def decode_n_tokens(
)
# TODO: use tokenizer's eos
- if (cur_token[0, 0, -1] == eos_token_id).any():
+ if cur_token[0, 0, -1] == eos_token_id or (cur_token[0, 1:, -1] == 1).any():
break
return previous_tokens[:, : i + 1]