File size: 12,209 Bytes
d6cd530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
#!/usr/bin/env python3
# coding: utf‑8
"""
CosyVoice gRPC back‑end – updated to mirror the FastAPI logic
*   loads CosyVoice2 with TRT / FP16 first (falls back to CosyVoice)
*   inference_zero_shot  ➜  adds   stream=False   +   speed
*   inference_instruct   ➜  keeps original β€œspeaker‑ID” path
*   inference_instruct2  ➜  new:  prompt‑audio + speed (no speaker‑ID)
"""

import io, tempfile, requests, soundfile as sf, torchaudio
import os
import sys
from concurrent import futures
import argparse
import logging
import grpc
import numpy as np
import torch

import cosyvoice_pb2
import cosyvoice_pb2_grpc

# ────────────────────────────────────────────────────────────────────────────────
# set‑up
# ────────────────────────────────────────────────────────────────────────────────
logging.getLogger("matplotlib").setLevel(logging.WARNING)
logging.basicConfig(level=logging.INFO,
                    format="%(asctime)s %(levelname)s %(message)s")

ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.extend([
    f"{ROOT_DIR}/../../..",
    f"{ROOT_DIR}/../../../third_party/Matcha-TTS",
])

from cosyvoice.cli.cosyvoice import CosyVoice2          # noqa: E402


# ────────────────────────────────────────────────────────────────────────────────
# helpers
# ────────────────────────────────────────────────────────────────────────────────
def _bytes_to_tensor(wav_bytes: bytes) -> torch.Tensor:
    """
    Convert int16 little‑endian PCM bytes β†’ torch.FloatTensor in range [‑1,1]
    """
    speech = torch.from_numpy(
        np.frombuffer(wav_bytes, dtype=np.int16)
    ).unsqueeze(0).float() / (2 ** 15)
    return speech                                                      # [1,β€―T]


def _yield_audio(model_output):
    """
    Generator that converts CosyVoice output β†’ protobuf Response messages.
    """
    for seg in model_output:
        pcm16 = (seg["tts_speech"].numpy() * (2 ** 15)).astype(np.int16)
        resp = cosyvoice_pb2.Response(tts_audio=pcm16.tobytes())
        yield resp

import urllib.parse

def _load_prompt_from_url(url: str, target_sr: int = 16_000) -> torch.Tensor:
    """
    Download *url* (wav / mp3 / flac) ➜ mono torch.FloatTensor [1,β€―T] @ target_sr.
    The temp file is removed before return.
    """
    resp = requests.get(url, timeout=10)
    resp.raise_for_status()

    # keep the original extension so torchaudio picks the right decoder
    ext = os.path.splitext(urllib.parse.urlparse(url).path)[1] or ".tmp"
    with tempfile.NamedTemporaryFile(delete=False, suffix=ext) as f:
        f.write(resp.content)
        tmp_path = f.name

    try:
        wav, sr = torchaudio.load(tmp_path)        # handles wav / mp3 / flac
        if wav.ndim > 1:
            wav = wav.mean(dim=0, keepdim=True)
        if sr != target_sr:
            wav = torchaudio.functional.resample(wav, sr, target_sr)
        return wav                                  # [1,β€―T] float32 in [-1,1]
    finally:
        try:
            os.remove(tmp_path)
        except Exception as e:
            logging.warning("Could not delete temp file %s: %s", tmp_path, e)

# ────────────────────────────────────────────────────────────────────────────────
# gRPC service
# ────────────────────────────────────────────────────────────────────────────────
class CosyVoiceServiceImpl(cosyvoice_pb2_grpc.CosyVoiceServicer):
    def __init__(self, args):
        # try CosyVoice2 first (preferred runtime: TRT / FP16)
        try:
            self.cosyvoice = CosyVoice2(args.model_dir,
                                        load_jit=False,
                                        load_trt=True,
                                        fp16=True)
            logging.info("Loaded CosyVoice2 (TRT / FP16).")
        except Exception:
            raise TypeError("No valid CosyVoice model found!")

    # ---------------------------------------------------------------------
    # single bi‑di streaming RPC
    # ---------------------------------------------------------------------
    def Inference(self, request, context):
        """Route to the correct model call based on the oneof field present."""
        # 1. Supervised fine‑tuning
        if request.HasField("sft_request"):
            logging.info("Received SFT inference request")
            mo = self.cosyvoice.inference_sft(
                request.sft_request.tts_text,
                request.sft_request.spk_id
            )
            yield from _yield_audio(mo)
            return

        # 2. Zero‑shot speaker cloning  (bytes OR S3 URL)
        if request.HasField("zero_shot_request"):
            logging.info("Received zero‑shot inference request")
            zr = request.zero_shot_request
            tmp_path = None  # initialise so we can delete later
        
            try:
                # ───── determine payload type ──────────────────────────────────────
                if zr.prompt_audio.startswith(b'http'):
                    prompt = _load_prompt_from_url(zr.prompt_audio.decode('utf‑8'))
                else:
                    # β€”β€” legacy raw PCM bytes β€”β€” -----------------------------------
                    prompt = _bytes_to_tensor(zr.prompt_audio)
        
                # ───── call the model ──────────────────────────────────────────────
                speed = getattr(zr, "speed", 1.0)
                mo = self.cosyvoice.inference_zero_shot(
                    zr.tts_text,
                    zr.prompt_text,
                    prompt,
                    stream=False,
                    speed=speed,
                )
          
            finally:
                # clean up any temporary file we created
                if tmp_path and os.path.exists(tmp_path):
                    try:
                        os.remove(tmp_path)
                    except Exception as e:
                        logging.warning("Could not remove temp file %s: %s", tmp_path, e)

            yield from _yield_audio(mo)
            return
      
        # 3. Cross‑lingual
        if request.HasField("cross_lingual_request"):
            logging.info("Received cross‑lingual inference request")
            cr = request.cross_lingual_request
            tmp_path = None
        
            try:
                if cr.prompt_audio.startswith(b'http'):          # S3 URL case
                    prompt = _load_prompt_from_url(cr.prompt_audio.decode('utf‑8'))        
                else:                                           # legacy raw bytes
                    prompt = _bytes_to_tensor(cr.prompt_audio)
        
                mo = self.cosyvoice.inference_cross_lingual(
                    cr.tts_text,
                    prompt
                )
        
            finally:
                if tmp_path and os.path.exists(tmp_path):
                    try:
                        os.remove(tmp_path)
                    except Exception as e:
                        logging.warning("Could not remove temp file %s: %s",
                                        tmp_path, e)
        
            yield from _yield_audio(mo)
            return


        # 4. Instruction‑TTS (two flavours)
        if request.HasField("instruct_request"):
            ir = request.instruct_request
        
            # -- instruct‑2 (prompt‑audio) ------------------------------------------
            if ('prompt_audio' in ir.DESCRIPTOR.fields_by_name   # <- descriptor guard
                    and ir.HasField("prompt_audio")):
                logging.info("Received instruct‑2 inference request")

                tmp_path = None
                try:
                    if ir.prompt_audio.startswith(b'http'):
                        prompt = _load_prompt_from_url(ir.prompt_audio.decode('utf‑8'))        
                    else:
                        # legacy raw‑bytes payload
                        prompt = _bytes_to_tensor(ir.prompt_audio)
        
                    speed = getattr(ir, "speed", 1.0)
                    mo = self.cosyvoice.inference_instruct2(
                        ir.tts_text,
                        ir.instruct_text,
                        prompt,
                        stream=False,
                        speed=speed
                    )
        
                finally:
                    if tmp_path and os.path.exists(tmp_path):
                        try:
                            os.remove(tmp_path)
                        except Exception as e:
                            logging.warning("Could not remove temp file %s: %s",
                                            tmp_path, e)
        
            # ──────────────────────────────────────────────────────────────────
            # 4‑b) classic instruct (speaker‑ID, no prompt audio)
            # ──────────────────────────────────────────────────────────────────
            else:
                logging.info("Received instruct inference request")
                mo = self.cosyvoice.inference_instruct(
                    ir.tts_text,
                    ir.spk_id,
                    ir.instruct_text
                )
        
            yield from _yield_audio(mo)
            return

        # unknown request type
        context.abort(grpc.StatusCode.INVALID_ARGUMENT,
                      "Unsupported request type in oneof field.")


# ────────────────────────────────────────────────────────────────────────────────
# entry‑point
# ────────────────────────────────────────────────────────────────────────────────
def serve(args):
    server = grpc.server(
        futures.ThreadPoolExecutor(max_workers=args.max_conc),
        maximum_concurrent_rpcs=args.max_conc
    )
    cosyvoice_pb2_grpc.add_CosyVoiceServicer_to_server(
        CosyVoiceServiceImpl(args), server
    )
    server.add_insecure_port(f"0.0.0.0:{args.port}")
    server.start()
    logging.info("CosyVoice gRPC server listening on 0.0.0.0:%d", args.port)
    server.wait_for_termination()


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--port", type=int, default=8000)
    parser.add_argument("--max_conc", type=int, default=4,
                        help="maximum concurrent requests / threads")
    parser.add_argument("--model_dir", type=str,
                        default="pretrained_models/CosyVoice2-0.5B",
                        help="local path or ModelScope repo id")
    serve(parser.parse_args())