add online trt export

Author: aluminumbox

README.md

@@ -128,8 +128,6 @@ import torchaudio
 
 **CosyVoice2 Usage**
 ```python
-# NOTE if you want to use tensorRT to accerlate the flow matching inference, please set load_trt=True.
-# if you don't want to save tensorRT model on disk, please set environment variable `NOT_SAVE_TRT=1`.
 cosyvoice = CosyVoice2('pretrained_models/CosyVoice2-0.5B', load_jit=False, load_trt=False, fp16=False)
 
 # NOTE if you want to reproduce the results on https://funaudiollm.github.io/cosyvoice2, please add text_frontend=False during inference

cosyvoice/cli/cosyvoice.py

@@ -53,7 +53,9 @@ def __init__(self, model_dir, load_jit=False, load_trt=False, fp16=False):
                                 '{}/llm.llm.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
                                 '{}/flow.encoder.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'))
         if load_trt:
-            self.model.load_trt('{}/flow.decoder.estimator.{}.v100.plan'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'))
+            self.model.load_trt('{}/flow.decoder.estimator.{}.mygpu.plan'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
+                                '{}/flow.decoder.estimator.fp32.onnx'.format(model_dir),
+                                self.fp16)
         del configs
 
     def list_available_spks(self):
@@ -149,7 +151,9 @@ def __init__(self, model_dir, load_jit=False, load_trt=False, fp16=False):
         if load_jit:
             self.model.load_jit('{}/flow.encoder.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'))
         if load_trt:
-            self.model.load_trt('{}/flow.decoder.estimator'.format(model_dir), self.fp16)
+            self.model.load_trt('{}/flow.decoder.estimator.{}.mygpu.plan'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
+                                '{}/flow.decoder.estimator.fp32.onnx'.format(model_dir),
+                                self.fp16)
         del configs
 
     def inference_instruct(self, *args, **kwargs):

cosyvoice/cli/model.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import os
 import torch
 import numpy as np
 import threading
@@ -19,7 +20,7 @@
 from contextlib import nullcontext
 import uuid
 from cosyvoice.utils.common import fade_in_out
-from cosyvoice.trt.estimator_trt import EstimatorTRT
+from cosyvoice.utils.file_utils import convert_onnx_to_trt
 
 
 class CosyVoiceModel:
@@ -36,6 +37,9 @@ def __init__(self,
         self.fp16 = fp16
         self.llm.fp16 = fp16
         self.flow.fp16 = fp16
+        if self.fp16 is True:
+            self.llm.half()
+            self.flow.half()
         self.token_min_hop_len = 2 * self.flow.input_frame_rate
         self.token_max_hop_len = 4 * self.flow.input_frame_rate
         self.token_overlap_len = 20
@@ -70,9 +74,6 @@ def load(self, llm_model, flow_model, hift_model):
         hift_state_dict = {k.replace('generator.', ''): v for k, v in torch.load(hift_model, map_location=self.device).items()}
         self.hift.load_state_dict(hift_state_dict, strict=True)
         self.hift.to(self.device).eval()
-        if self.fp16 is True:
-            self.llm.half()
-            self.flow.half()
 
     def load_jit(self, llm_text_encoder_model, llm_llm_model, flow_encoder_model):
         llm_text_encoder = torch.jit.load(llm_text_encoder_model, map_location=self.device)
@@ -82,9 +83,17 @@ def load_jit(self, llm_text_encoder_model, llm_llm_model, flow_encoder_model):
         flow_encoder = torch.jit.load(flow_encoder_model, map_location=self.device)
         self.flow.encoder = flow_encoder
 
-    def load_trt(self, flow_decoder_estimator_model, fp16):
+    def load_trt(self, flow_decoder_estimator_model, flow_decoder_onnx_model, fp16):
+        assert torch.cuda.is_available(), 'tensorrt only supports gpu!'
+        if not os.path.exists(flow_decoder_estimator_model):
+            convert_onnx_to_trt(flow_decoder_estimator_model, flow_decoder_onnx_model, fp16)
         del self.flow.decoder.estimator
-        self.flow.decoder.estimator = EstimatorTRT(flow_decoder_estimator_model, self.device, fp16)
+        import tensorrt as trt
+        with open(flow_decoder_estimator_model, 'rb') as f:
+            self.flow.decoder.estimator_engine = trt.Runtime(trt.Logger(trt.Logger.INFO)).deserialize_cuda_engine(f.read())
+        if self.flow.decoder.estimator_engine is None:
+            raise ValueError('failed to load trt {}'.format(flow_decoder_estimator_model))
+        self.flow.decoder.estimator = self.flow.decoder.estimator_engine.create_execution_context()
 
     def llm_job(self, text, prompt_text, llm_prompt_speech_token, llm_embedding, uuid):
         with self.llm_context:
@@ -269,6 +278,9 @@ def __init__(self,
         self.fp16 = fp16
         self.llm.fp16 = fp16
         self.flow.fp16 = fp16
+        if self.fp16 is True:
+            self.llm.half()
+            self.flow.half()
         self.token_hop_len = 2 * self.flow.input_frame_rate
         # here we fix flow encoder/decoder decoding_chunk_size, in the future we will send it as arguments, or use cache
         self.flow.encoder.static_chunk_size = 2 * self.flow.input_frame_rate

cosyvoice/dataset/processor.py

@@ -21,7 +21,6 @@
 from torch.nn.utils.rnn import pad_sequence
 import torch.nn.functional as F
 
-torchaudio.set_audio_backend('soundfile')
 
 AUDIO_FORMAT_SETS = {'flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma'}
 

cosyvoice/flow/flow_matching.py

@@ -134,12 +134,12 @@ def forward_estimator(self, x, mask, mu, t, spks, cond):
                 self.estimator.set_input_shape('cond', (2, 80, x.size(2)))
                 # run trt engine
                 self.estimator.execute_v2([x.contiguous().data_ptr(),
-                                        mask.contiguous().data_ptr(),
-                                        mu.contiguous().data_ptr(),
-                                        t.contiguous().data_ptr(),
-                                        spks.contiguous().data_ptr(),
-                                        cond.contiguous().data_ptr(),
-                                        x.data_ptr()])
+                                           mask.contiguous().data_ptr(),
+                                           mu.contiguous().data_ptr(),
+                                           t.contiguous().data_ptr(),
+                                           spks.contiguous().data_ptr(),
+                                           cond.contiguous().data_ptr(),
+                                           x.data_ptr()])
             return x
 
     def compute_loss(self, x1, mask, mu, spks=None, cond=None):

cosyvoice/hifigan/discriminator.py

@@ -1,6 +1,6 @@
 import torch
 import torch.nn as nn
-from torch.nn.utils import weight_norm
+from torch.nn.utils.parametrizations import weight_norm
 from typing import List, Optional, Tuple
 from einops import rearrange
 from torchaudio.transforms import Spectrogram

cosyvoice/hifigan/f0_predictor.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 import torch
 import torch.nn as nn
-from torch.nn.utils import weight_norm
+from torch.nn.utils.parametrizations import weight_norm
 
 
 class ConvRNNF0Predictor(nn.Module):

cosyvoice/hifigan/generator.py

@@ -23,7 +23,7 @@
 from torch.nn import Conv1d
 from torch.nn import ConvTranspose1d
 from torch.nn.utils import remove_weight_norm
-from torch.nn.utils import weight_norm
+from torch.nn.utils.parametrizations import weight_norm
 from torch.distributions.uniform import Uniform
 
 from cosyvoice.transformer.activation import Snake

cosyvoice/trt/estimator_trt.py

@@ -1,5 +1,5 @@
 # Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang)
-#               2024 Alibaba Inc (authors: Xiang Lyu)
+#               2024 Alibaba Inc (authors: Xiang Lyu, Zetao Hu)
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,6 +14,7 @@
 # limitations under the License.
 
 import json
+import tensorrt as trt
 import torchaudio
 import logging
 logging.getLogger('matplotlib').setLevel(logging.WARNING)
@@ -45,3 +46,44 @@ def load_wav(wav, target_sr):
         assert sample_rate > target_sr, 'wav sample rate {} must be greater than {}'.format(sample_rate, target_sr)
         speech = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sr)(speech)
     return speech
+
+
+def convert_onnx_to_trt(trt_model, onnx_model, fp16):
+    _min_shape = [(2, 80, 4), (2, 1, 4), (2, 80, 4), (2,), (2, 80), (2, 80, 4)]
+    _opt_shape = [(2, 80, 193), (2, 1, 193), (2, 80, 193), (2,), (2, 80), (2, 80, 193)]
+    _max_shape = [(2, 80, 6800), (2, 1, 6800), (2, 80, 6800), (2,), (2, 80), (2, 80, 6800)]
+    input_names = ["x", "mask", "mu", "t", "spks", "cond"]
+
+    logging.info("Converting onnx to trt...")
+    network_flags = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
+    logger = trt.Logger(trt.Logger.INFO)
+    builder = trt.Builder(logger)
+    network = builder.create_network(network_flags)
+    parser = trt.OnnxParser(network, logger)
+    config = builder.create_builder_config()
+    config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 33)  # 8GB
+    if fp16:
+        config.set_flag(trt.BuilderFlag.FP16)
+    profile = builder.create_optimization_profile()
+    # load onnx model
+    with open(onnx_model, "rb") as f:
+        if not parser.parse(f.read()):
+            for error in range(parser.num_errors):
+                print(parser.get_error(error))
+            raise ValueError('failed to parse {}'.format(onnx_model))
+    # set input shapes
+    for i in range(len(input_names)):
+        profile.set_shape(input_names[i], _min_shape[i], _opt_shape[i], _max_shape[i])
+    tensor_dtype = trt.DataType.HALF if fp16 else trt.DataType.FLOAT
+    # set input and output data type
+    for i in range(network.num_inputs):
+        input_tensor = network.get_input(i)
+        input_tensor.dtype = tensor_dtype
+    for i in range(network.num_outputs):
+        output_tensor = network.get_output(i)
+        output_tensor.dtype = tensor_dtype
+    config.add_optimization_profile(profile)
+    engine_bytes = builder.build_serialized_network(network, config)
+    # save trt engine
+    with open(trt_model, "wb") as f:
+        f.write(engine_bytes)