update model inference

2026-02-05 01:49:25 +08:00 · 2024-07-24 19:18:09 +08:00
parent a13411c561
commit 02f941d348
5 changed files with 85 additions and 64 deletions
--- a/cosyvoice/cli/cosyvoice.py
+++ b/cosyvoice/cli/cosyvoice.py
@@ -46,9 +46,9 @@ class CosyVoice:
        return spks

    def inference_sft(self, tts_text, spk_id, stream=False):
-        start_time = time.time()
        for i in self.frontend.text_normalize(tts_text, split=True):
            model_input = self.frontend.frontend_sft(i, spk_id)
+            start_time = time.time()
            for model_output in self.model.inference(**model_input, stream=stream):
                speech_len = model_output['tts_speech'].shape[1] / 22050
                logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
@@ -56,10 +56,10 @@ class CosyVoice:
                start_time = time.time()

    def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False):
-        start_time = time.time()
        prompt_text = self.frontend.text_normalize(prompt_text, split=False)
        for i in self.frontend.text_normalize(tts_text, split=True):
            model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
+            start_time = time.time()
            for model_output in self.model.inference(**model_input, stream=stream):
                speech_len = model_output['tts_speech'].shape[1] / 22050
                logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
@@ -69,9 +69,9 @@ class CosyVoice:
    def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False):
        if self.frontend.instruct is True:
            raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
-        start_time = time.time()
        for i in self.frontend.text_normalize(tts_text, split=True):
            model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
+            start_time = time.time()
            for model_output in self.model.inference(**model_input, stream=stream):
                speech_len = model_output['tts_speech'].shape[1] / 22050
                logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
@@ -81,10 +81,10 @@ class CosyVoice:
    def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False):
        if self.frontend.instruct is False:
            raise ValueError('{} do not support instruct inference'.format(self.model_dir))
-        start_time = time.time()
        instruct_text = self.frontend.text_normalize(instruct_text, split=False)
        for i in self.frontend.text_normalize(tts_text, split=True):
            model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text)
+            start_time = time.time()
            for model_output in self.model.inference(**model_input, stream=stream):
                speech_len = model_output['tts_speech'].shape[1] / 22050
                logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
--- a/cosyvoice/cli/model.py
+++ b/cosyvoice/cli/model.py
@@ -13,6 +13,9 @@
 # limitations under the License.
 import torch
 import numpy as np
+import threading
+import time
+from contextlib import nullcontext


 class CosyVoiceModel:
@@ -25,10 +28,13 @@ class CosyVoiceModel:
        self.llm = llm
        self.flow = flow
        self.hift = hift
-        self.stream_win_len = 60
-        self.stream_hop_len = 50
-        self.overlap = 4395 # 10 token equals 4395 sample point
+        self.stream_win_len = 60 * 4
+        self.stream_hop_len = 50 * 4
+        self.overlap = 4395 * 4 # 10 token equals 4395 sample point
        self.window = np.hamming(2 * self.overlap)
+        self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
+        self.flow_hift_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
+        self.lock = threading.Lock()

    def load(self, llm_model, flow_model, hift_model):
        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device))
@@ -38,13 +44,8 @@ class CosyVoiceModel:
        self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
        self.hift.to(self.device).eval()

-    def inference(self, text, text_len, flow_embedding, llm_embedding=torch.zeros(0, 192),
-                  prompt_text=torch.zeros(1, 0, dtype=torch.int32), prompt_text_len=torch.zeros(1, dtype=torch.int32),
-                  llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), llm_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
-                  flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), flow_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
-                  prompt_speech_feat=torch.zeros(1, 0, 80), prompt_speech_feat_len=torch.zeros(1, dtype=torch.int32), stream=False):
-        if stream is True:
-            tts_speech_token, cache_speech = [], None
+    def llm_job(self, text, text_len, prompt_text, prompt_text_len, llm_prompt_speech_token, llm_prompt_speech_token_len, llm_embedding):
+        with self.llm_context:
            for i in self.llm.inference(text=text.to(self.device),
                                                text_len=text_len.to(self.device),
                                                prompt_text=prompt_text.to(self.device),
@@ -56,10 +57,56 @@ class CosyVoiceModel:
                                                sampling=25,
                                                max_token_text_ratio=30,
                                                min_token_text_ratio=3,
-                                                stream=stream):
-                tts_speech_token.append(i)
-                if len(tts_speech_token) == self.stream_win_len:
-                    this_tts_speech_token = torch.concat(tts_speech_token, dim=1)
+                                                stream=True):
+                self.tts_speech_token.append(i)
+        self.llm_end = True
+
+    def token2wav(self, token, prompt_token, prompt_token_len, prompt_feat, prompt_feat_len, embedding):
+        with self.flow_hift_context:
+            tts_mel = self.flow.inference(token=token.to(self.device),
+                                        token_len=torch.tensor([token.size(1)], dtype=torch.int32).to(self.device),
+                                        prompt_token=prompt_token.to(self.device),
+                                        prompt_token_len=prompt_token_len.to(self.device),
+                                        prompt_feat=prompt_feat.to(self.device),
+                                        prompt_feat_len=prompt_feat_len.to(self.device),
+                                        embedding=embedding.to(self.device))
+            tts_speech = self.hift.inference(mel=tts_mel).cpu()
+        return tts_speech
+
+    def inference(self, text, text_len, flow_embedding, llm_embedding=torch.zeros(0, 192),
+                  prompt_text=torch.zeros(1, 0, dtype=torch.int32), prompt_text_len=torch.zeros(1, dtype=torch.int32),
+                  llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), llm_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
+                  flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), flow_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
+                  prompt_speech_feat=torch.zeros(1, 0, 80), prompt_speech_feat_len=torch.zeros(1, dtype=torch.int32), stream=False):
+        if stream is True:
+            self.tts_speech_token, self.llm_end, cache_speech = [], False, None
+            p = threading.Thread(target=self.llm_job, args=(text.to(self.device), text_len.to(self.device), prompt_text.to(self.device), prompt_text_len.to(self.device),
+                                                     llm_prompt_speech_token.to(self.device), llm_prompt_speech_token_len.to(self.device), llm_embedding.to(self.device)))
+            p.start()
+            while True:
+                time.sleep(0.1)
+                if len(self.tts_speech_token) >= self.stream_win_len:
+                    this_tts_speech_token = torch.concat(self.tts_speech_token[:self.stream_win_len], dim=1)
+                    with self.flow_hift_context:
+                        this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                                    prompt_token=flow_prompt_speech_token.to(self.device),
+                                                    prompt_token_len=flow_prompt_speech_token_len.to(self.device),
+                                                    prompt_feat=prompt_speech_feat.to(self.device),
+                                                    prompt_feat_len=prompt_speech_feat_len.to(self.device),
+                                                    embedding=flow_embedding.to(self.device))
+                    # fade in/out if necessary
+                    if cache_speech is not None:
+                        this_tts_speech[:, :self.overlap] = this_tts_speech[:, :self.overlap] * self.window[:self.overlap] + cache_speech * self.window[-self.overlap:]
+                    yield  {'tts_speech': this_tts_speech[:, :-self.overlap]}
+                    cache_speech = this_tts_speech[:, -self.overlap:]
+                    with self.lock:
+                        self.tts_speech_token = self.tts_speech_token[self.stream_hop_len:]
+                if self.llm_end is True:
+                    break
+            # deal with remain tokens
+            if cache_speech is None or len(self.tts_speech_token) > self.stream_win_len - self.stream_hop_len:
+                this_tts_speech_token = torch.concat(self.tts_speech_token, dim=1)
+                with self.flow_hift_context:
                    this_tts_mel = self.flow.inference(token=this_tts_speech_token,
                                                token_len=torch.tensor([this_tts_speech_token.size(1)], dtype=torch.int32).to(self.device),
                                                prompt_token=flow_prompt_speech_token.to(self.device),
@@ -68,29 +115,14 @@ class CosyVoiceModel:
                                                prompt_feat_len=prompt_speech_feat_len.to(self.device),
                                                embedding=flow_embedding.to(self.device))
                    this_tts_speech = self.hift.inference(mel=this_tts_mel).cpu()
-                    # fade in/out if necessary
-                    if cache_speech is not None:
-                        this_tts_speech[:, :self.overlap] = this_tts_speech[:, :self.overlap] * self.window[:self.overlap] + cache_speech * self.window[-self.overlap:]
-                    yield  {'tts_speech': this_tts_speech[:, :-self.overlap]}
-                    cache_speech = this_tts_speech[:, -self.overlap:]
-                    tts_speech_token = tts_speech_token[-(self.stream_win_len - self.stream_hop_len):]
-            # deal with remain tokens
-            if cache_speech is None or len(tts_speech_token) > self.stream_win_len - self.stream_hop_len:
-                this_tts_speech_token = torch.concat(tts_speech_token, dim=1)
-                this_tts_mel = self.flow.inference(token=this_tts_speech_token,
-                                            token_len=torch.tensor([this_tts_speech_token.size(1)], dtype=torch.int32).to(self.device),
-                                            prompt_token=flow_prompt_speech_token.to(self.device),
-                                            prompt_token_len=flow_prompt_speech_token_len.to(self.device),
-                                            prompt_feat=prompt_speech_feat.to(self.device),
-                                            prompt_feat_len=prompt_speech_feat_len.to(self.device),
-                                            embedding=flow_embedding.to(self.device))
-                this_tts_speech = self.hift.inference(mel=this_tts_mel).cpu()
                if cache_speech is not None:
                    this_tts_speech[:, :self.overlap] = this_tts_speech[:, :self.overlap] * self.window[:self.overlap] + cache_speech * self.window[-self.overlap:]
                yield {'tts_speech': this_tts_speech}
            else:
-                assert len(tts_speech_token) == self.stream_win_len - self.stream_hop_len, 'tts_speech_token not equal to {}'.format(self.stream_win_len - self.stream_hop_len)
+                assert len(self.tts_speech_token) == self.stream_win_len - self.stream_hop_len, 'tts_speech_token not equal to {}'.format(self.stream_win_len - self.stream_hop_len)
                yield {'tts_speech': cache_speech}
+            p.join()
+            torch.cuda.synchronize()
        else:
            tts_speech_token = []
            for i in self.llm.inference(text=text.to(self.device),