Skip to content

SOTA Open Source TTS

Implement delay pattern generate

Lengyue37f4b517c2c9
diff --git a/tools/llama/generate.py b/tools/llama/generate.py
index b475b16..14ba931 100644
--- a/tools/llama/generate.py
+++ b/tools/llama/generate.py
@@ -109,8 +109,6 @@ def decode_one_token(
# Disable <s> and </s> tokens for codebooks
if model.config.num_codebooks != 0:
- logits.codebook_logits[:, :, :, :1] = -float("Inf")
-
for i in range(model.config.num_codebooks):
codebooks.append(
sample(
@@ -122,12 +120,7 @@ def decode_one_token(
)[0]
)
- codebooks = torch.stack(codebooks, dim=0)
- codebooks[1] = torch.where(
- codebooks[0] <= 32311, codebooks[1], codebooks[0] - 32311 + 2
- )
-
- return codebooks
+ return torch.stack(codebooks, dim=0)
def prefill(
@@ -143,10 +136,7 @@ def prefill(
)[0]
]
- # Disable <s> and </s> tokens for codebooks
if model.config.num_codebooks != 0:
- logits.codebook_logits[:, :, :, :2] = -float("Inf")
-
for i in range(model.config.num_codebooks):
codebooks.append(
sample(
@@ -330,9 +320,9 @@ def encode_tokens(
data = data[:num_codebooks]
# Since 1.0, we use <s:xxx> to replace <semantic>
- main_tokens = [f"<s:{i}>" for i in data[0]]
- main_token_ids = tokenizer.convert_tokens_to_ids(main_tokens)
- main_token_ids = torch.tensor([main_token_ids], dtype=torch.int, device=device)
+ main_token_ids = torch.tensor(
+ [[tokenizer.pad_token_id] * data.size(1)], dtype=torch.int, device=device
+ )
data = torch.cat((main_token_ids, data), dim=0)
prompt = torch.cat((prompt, data), dim=1)
@@ -502,7 +492,7 @@ def main(
decode_one_token, mode="reduce-overhead", fullgraph=True
)
- for i in range(num_samples):
+ for idx in range(num_samples):
torch.cuda.synchronize()
t0 = time.perf_counter()
@@ -518,7 +508,7 @@ def main(
repetition_penalty=repetition_penalty,
)
- if i == 0 and compile:
+ if idx == 0 and compile:
logger.info(f"Compilation time: {time.perf_counter() - t0:.2f} seconds")
torch.cuda.synchronize()
@@ -535,11 +525,18 @@ def main(
)
codes = y[1:, prompt_length:-1]
+ new_codes = []
+ for j, code in enumerate(codes):
+ new_codes.append(
+ code[j : codes.shape[1] - (model.config.num_codebooks - j - 1)]
+ )
+
+ codes = torch.stack(new_codes, dim=0)
codes = codes - 2
assert (codes >= 0).all(), "Codes should be >= 0"
- np.save(f"codes_{i}.npy", codes.cpu().numpy())
- logger.info(f"Saved codes to codes_{i}.npy")
+ np.save(f"codes_{idx}.npy", codes.cpu().numpy())
+ logger.info(f"Saved codes to codes_{idx}.npy")
if __name__ == "__main__":