Mirrors/bark-with-voice-clone
1
0
Fork 0
mirror of https://github.com/serp-ai/bark-with-voice-clone.git synced 2025-12-15 03:07:58 +01:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #27 from fiq/main

Browse Source 
Fixes MPS device errors from Tensor.type() when using generate_text_semantic and generate_coarse
This commit is contained in:
Francis LaBounty
2023-05-14 13:27:46 -06:00
committed by GitHub
parent 03eafe7937 c84f71e187
commit 43cbc43fb2
1 changed files with 4 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
bark/generation.py
View File

@@ -339,7 +339,6 @@ def preload_models(
# Generation Functionality
####
def _tokenize(tokenizer, text):
return tokenizer.encode(text, add_special_tokens=False)
@@ -351,7 +350,6 @@ def _detokenize(tokenizer, enc_text):
def _normalize_whitespace(text):
return re.sub(r"\s+", " ", text).strip()
TEXT_ENCODING_OFFSET = 10_048
SEMANTIC_PAD_TOKEN = 10_000
TEXT_PAD_TOKEN = 129_595
@@ -453,8 +451,7 @@ def generate_text_semantic(
)
if top_p is not None:
# faster to convert to numpy
logits_device = relevant_logits.device
logits_dtype = relevant_logits.type()
original_device = relevant_logits.device
relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
sorted_indices = np.argsort(relevant_logits)[::-1]
sorted_logits = relevant_logits[sorted_indices]
@@ -464,7 +461,7 @@ def generate_text_semantic(
sorted_indices_to_remove[0] = False
relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
relevant_logits = torch.from_numpy(relevant_logits)
relevant_logits = relevant_logits.to(logits_device).type(logits_dtype)
relevant_logits = relevant_logits.to(original_device)
if top_k is not None:
v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
relevant_logits[relevant_logits < v[-1]] = -float("Inf")
@@ -649,8 +646,7 @@ def generate_coarse(
relevant_logits = logits[0, 0, logit_start_idx:logit_end_idx]
if top_p is not None:
# faster to convert to numpy
logits_device = relevant_logits.device
logits_dtype = relevant_logits.type()
original_device = relevant_logits.device
relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
sorted_indices = np.argsort(relevant_logits)[::-1]
sorted_logits = relevant_logits[sorted_indices]
@@ -660,7 +656,7 @@ def generate_coarse(
sorted_indices_to_remove[0] = False
relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
relevant_logits = torch.from_numpy(relevant_logits)
relevant_logits = relevant_logits.to(logits_device).type(logits_dtype)
relevant_logits = relevant_logits.to(original_device)
if top_k is not None:
v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
relevant_logits[relevant_logits < v[-1]] = -float("Inf")