doc: add vLLM instruction

ChatTTS/model/gpt.py

@@ -5,7 +5,6 @@
 import gc
 from pathlib import Path
 
-from safetensors.torch import save_file
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -92,6 +91,8 @@ def __init__(
 
     def from_pretrained(self, file_path: str):
         if self.is_vllm and platform.system().lower() == "linux":
+            from safetensors.torch import save_file
+
             from .velocity.llm import LLM
             from .velocity.post_model import PostModel
 
@@ -104,7 +105,7 @@ def from_pretrained(self, file_path: str):
                 gpt.gpt.save_pretrained(vllm_folder / "gpt")
                 post_model = (
                     PostModel(
-                        int(self.gpt.config.hidden_size),
+                        int(gpt.gpt.config.hidden_size),
                         self.num_audio_tokens,
                         self.num_text_tokens,
                     )

README.md

@@ -101,7 +101,12 @@ conda activate chattts
 pip install -r requirements.txt
 ```
 
-#### Optional: Install TransformerEngine if using NVIDIA GPU (Linux only)
+#### Optional: Install vLLM (Linux only)
+```bash
+pip install safetensors vllm==0.2.7 torchaudio
+```
+
+#### Unrecommended Optional: Install TransformerEngine if using NVIDIA GPU (Linux only)
 > [!Note]
 > The installation process is very slow.
 
@@ -113,7 +118,7 @@ pip install -r requirements.txt
 pip install git+https://github.com/NVIDIA/TransformerEngine.git@stable
 ```
 
-#### Optional: Install FlashAttention-2 (mainly NVIDIA GPU)
+#### Unrecommended Optional: Install FlashAttention-2 (mainly NVIDIA GPU)
 > [!Note]
 > See supported devices at the [Hugging Face Doc](https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2).
 

requirements.txt

@@ -14,5 +14,3 @@ WeTextProcessing; sys_platform == 'linux'
 nemo_text_processing; sys_platform == 'linux'
 av
 pydub
-safetensors
-vllm>=0.2.7; sys_platform == 'linux'

setup.py

@@ -28,7 +28,6 @@
         "transformers>=4.41.1",
         "vector_quantize_pytorch",
         "vocos",
-        "safetensors",
     ],
     platforms="any",
     classifiers=[