fix(gpt): compile failed

ChatTTS/core.py

@@ -331,7 +331,6 @@ class Chat:
 
         return self.has_loaded()
 
-    @torch.inference_mode()
     def _infer(
         self,
         text,
@@ -506,7 +505,7 @@ class Chat:
         torch.where(cond, n, emb, out=emb)
         del cond, n
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def _infer_code(
         self,
         text: Tuple[List[str], str],
@@ -549,9 +548,7 @@ class Chat:
 
         input_ids, attention_mask, text_mask = self._text_to_token(text, gpt.device_gpt)
 
-        with torch.inference_mode(not self.compile):
-            with torch.no_grad():
-                emb = gpt(input_ids, text_mask)
+        emb = gpt(input_ids, text_mask)
 
         del text_mask
 
@@ -592,7 +589,7 @@ class Chat:
 
         return result
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def _refine_text(
         self,
         text: str,
@@ -616,9 +613,7 @@ class Chat:
             repetition_penalty=params.repetition_penalty,
         )
 
-        with torch.inference_mode(not self.compile):
-            with torch.no_grad():
-                emb = gpt(input_ids, text_mask)
+        emb = gpt(input_ids, text_mask)
 
         del text_mask
 

ChatTTS/model/gpt.py

@@ -347,7 +347,7 @@ class GPT(nn.Module):
             hiddens=hiddens,
         )
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def generate(
         self,
         emb: torch.Tensor,
@@ -366,7 +366,6 @@ class GPT(nn.Module):
         show_tqdm=True,
         ensure_non_empty=True,
         stream_batch=24,
-        compile=False,
         context=Context(),
     ):
 
@@ -436,17 +435,15 @@ class GPT(nn.Module):
 
             model_input.to(self.device_gpt, self.gpt.dtype)
 
-            with torch.inference_mode(not compile):
-                with torch.no_grad():
-                    outputs: BaseModelOutputWithPast = self.gpt(
-                        attention_mask=model_input.attention_mask,
-                        position_ids=model_input.position_ids,
-                        past_key_values=model_input.past_key_values,
-                        inputs_embeds=model_input.inputs_embeds,
-                        use_cache=model_input.use_cache,
-                        output_attentions=return_attn,
-                        cache_position=model_input.cache_position,
-                    )
+            outputs: BaseModelOutputWithPast = self.gpt(
+                attention_mask=model_input.attention_mask,
+                position_ids=model_input.position_ids,
+                past_key_values=model_input.past_key_values,
+                inputs_embeds=model_input.inputs_embeds,
+                use_cache=model_input.use_cache,
+                output_attentions=return_attn,
+                cache_position=model_input.cache_position,
+            )
             del_all(model_input)
             attentions.append(outputs.attentions)
             hidden_states = outputs.last_hidden_state.to(

examples/cmd/stream.py

@@ -1,37 +1,13 @@
-import sys
-import torch
-import numpy as np
-import ChatTTS
-from IPython.display import Audio
-
-
 import io
 import threading
 import time
-import pyaudio
-
 import random
 
+import pyaudio
+import numpy as np
+import ChatTTS
 
-# 如果不对batch进行统一量纲，同batch的多段话之间量纲会有较大差异，以致于无法在后处理中获取正常的历史语音结果？
-def batch_unsafe_float_to_int16(audios: list[np.ndarray], am=None) -> list[np.ndarray]:
-    """
-    This function will destroy audio, use only once.
-    """
-
-    valid_audios = [i for i in audios if i is not None]
-    if len(valid_audios) > 1:
-        am = np.abs(np.concatenate(valid_audios, axis=1)).max() * 32768
-    else:
-        am = np.abs(valid_audios[0]).max() * 32768
-    am = 32767 * 32768 / am
-
-    for i in range(len(audios)):
-        if audios[i] is not None:
-            np.multiply(audios[i], am, audios[i])
-            audios[i] = audios[i].astype(np.int16)
-    return audios
-
+from tools.audio import batch_unsafe_float_to_int16
 
 # 流式声音处理器
 class AudioStreamer:

tools/audio/__init__.py

@@ -1,3 +1,3 @@
 from .mp3 import wav_arr_to_mp3_view
 from .ffmpeg import has_ffmpeg_installed
-from .np import unsafe_float_to_int16
+from .np import unsafe_float_to_int16, batch_unsafe_float_to_int16

tools/audio/np.py

@@ -12,3 +12,22 @@ def unsafe_float_to_int16(audio: np.ndarray) -> np.ndarray:
     np.multiply(audio, am, audio)
     audio16 = audio.astype(np.int16)
     return audio16
+
+@jit
+def batch_unsafe_float_to_int16(audios: list[np.ndarray]) -> list[np.ndarray]:
+    """
+    This function will destroy audio, use only once.
+    """
+
+    valid_audios = [i for i in audios if i is not None]
+    if len(valid_audios) > 1:
+        am = np.abs(np.concatenate(valid_audios, axis=1)).max() * 32768
+    else:
+        am = np.abs(valid_audios[0]).max() * 32768
+    am = 32767 * 32768 / am
+
+    for i in range(len(audios)):
+        if audios[i] is not None:
+            np.multiply(audios[i], am, audios[i])
+            audios[i] = audios[i].astype(np.int16)
+    return audios