update readme

examples/grpo/cosyvoice2/Dockerfile

@@ -1,5 +1,5 @@
 FROM verlai/verl:app-verl0.4-vllm0.8.5-mcore0.12.2-te2.2
-COPY requirements-cosyvoice.txt /myworkspace/requirements.txt
+COPY requirements.txt /myworkspace/requirements.txt
 RUN pip install -r /myworkspace/requirements.txt
 RUN pip install -U nvidia-pytriton
 RUN git clone https://github.com/yuekaizhang/verl.git /myworkspace/verl -b thread && cd /myworkspace/verl && pip install --no-deps -e .

examples/grpo/cosyvoice2/README.md

@@ -18,6 +18,7 @@ We recommend using the pre-built Docker image below. Alternatively, you can manu
 ```bash
 docker pull soar97/verl:app-verl0.4-vllm0.8.5-mcore0.12.2-te2.2
 ```
+If Docker is not available, you can refer to `run.sh` `stage -2` to install the dependencies locally.
 
 ## Data Preparation
 
@@ -43,16 +44,16 @@ data/parquet_tiny/train.parquet
 data/parquet_tiny/test.parquet
 ```
 
-Each sample is automatically wrapped into a cosyvoice2-style prompt so that the LLM learns to output CosyVoice2 speech tokens.
+Each sample is automatically wrapped into a CosyVoice2-style prompt so that the LLM learns to output CosyVoice2 speech tokens.
 
 
 ## Reward Function & ASR Server
 
-To compute rewards we run a lightweight server that:
+To compute rewards, we run a lightweight server that:
 
 1. Converts generated speech tokens back to a 16 kHz waveform with the **CosyVoice2** pretrained U-Net model.
 2. Transcribes the waveform with **SenseVoice** ASR.
-3. Calculates the pinyin-level error rate relative to the ground-truth text and maps it to a score in the range \[0-1\].
+3. Calculates the pinyin-level error rate relative to the ground-truth text and maps it to a score between 0 and 1.
 
 Start the server (stage `1`) in a dedicated terminal or on a separate GPU:
 
@@ -61,7 +62,7 @@ bash run.sh 1 1
 # Triton server listens on ports 8000/8001/8002
 ```
 
-The custom reward implementation lives in [`reward_tts.py`](./reward_tts.py) and calls the server to obtain the reward score.
+The custom reward implementation is located in [`reward_tts.py`](./reward_tts.py) and calls the server to obtain the reward score.
 
 ## Training
 
@@ -78,10 +79,12 @@ Key CLI arguments passed to `verl.trainer.main_ppo`:
 * `custom_reward_function.path=reward_tts.py` – custom reward function described above.
 
 Adjust `CUDA_VISIBLE_DEVICES`, batch sizes, and other hyperparameters to match your hardware.
+> [!TIP]
+> Note: the lm_head bias is disabled during training to make the model compatible with VLLM and Transformers' Qwen model.
 
 ## Evaluation
 
-After training completes, collect the sharded FSDP weights and export a Hugging Face-style checkpoint (stage `3`):
+After training is complete, collect the sharded FSDP weights and export a Hugging Face-style checkpoint (stage `3`):
 
 ```bash
 bash run.sh 3 3   # merges weights into $llm_path/merged_hf_model
@@ -107,15 +110,16 @@ bash run.sh 5 5
 ```
 
 The script converts the Hugging Face checkpoint back into the format expected by the CosyVoice repository.
+> [!TIP]
+>  However, we observed a slight accuracy drop when using the RL-trained model after conversion, compared with the Hugging Face format. 
 
 ## Results
 
 | Model | Seed-TTS `test_zh` CER | CosyVoice3 `zero_shot_zh` CER | Comment |
 |-------|------------------------|------------------------------|---------|
 | CosyVoice2 LLM (official) | 1.45% | 4.08% | See the [paper](https://arxiv.org/abs/2412.10117) |
-| CosyVoice2 LLM + GRPO | 1.37 % | **3.36 %** | See the [decoding results](yuekai/official-cosyvoice-llm-grpo-aishell3) |
+| CosyVoice2 LLM + GRPO | 1.37% | **3.36%** | See the [decoding results](yuekai/official-cosyvoice-llm-grpo-aishell3), Hugging Face-format model |
 
 ## Acknowledgement
 
 This work was inspired by the implementation in [ch-tts-llasa-rl-grpo](https://github.com/channel-io/ch-tts-llasa-rl-grpo).
-

examples/grpo/cosyvoice2/pretrained_to_huggingface.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 
 # SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
@@ -94,7 +95,8 @@ if __name__ == "__main__":
     with torch.no_grad():
         # set the weight and bias of the new lm_head to 0
         new_lm_head.weight.data.zero_()
-        new_lm_head.bias.data.zero_()
+        # make bias value -inf
+        new_lm_head.bias.data.fill_(-float('inf'))
         new_lm_head.weight[original_tokenizer_vocab_size:original_tokenizer_vocab_size + cosyvoice2_token_size + 3] = llm_decoder.weight
         new_lm_head.bias[original_tokenizer_vocab_size:original_tokenizer_vocab_size + cosyvoice2_token_size + 3] = llm_decoder.bias
 
@@ -107,8 +109,7 @@ if __name__ == "__main__":
 
     eos_token_ids = [original_tokenizer_vocab_size + cosyvoice2_token_size,
                      original_tokenizer_vocab_size + cosyvoice2_token_size + 1,
-                     original_tokenizer_vocab_size + cosyvoice2_token_size + 2,
+                     original_tokenizer_vocab_size + cosyvoice2_token_size + 2]
-                     original_tokenizer_vocab_size + cosyvoice2_token_size + 3]
     llm.generation_config.eos_token_id = eos_token_ids
     llm.generation_config.temperature = 1.0
     llm.generation_config.top_p = 0.8
@@ -121,6 +122,14 @@ if __name__ == "__main__":
     llm.to(torch.bfloat16)
     llm.save_pretrained(args.save_path)
 
-    TEMPLATE = "{%- for message in messages %}{%- if message['role'] == 'user' %}{{- '<|sos|>' + message['content'] + '<|task_id|>' }}{%- elif message['role'] == 'assistant' %}{{- message['content']}}{%- endif %}{%- endfor %}"
+    TEMPLATE = (
+        "{%- for message in messages %}"
+        "{%- if message['role'] == 'user' %}"
+        "{{- '<|sos|>' + message['content'] + '<|task_id|>' }}"
+        "{%- elif message['role'] == 'assistant' %}"
+        "{{- message['content']}}"
+        "{%- endif %}"
+        "{%- endfor %}"
+    )
     tokenizer.chat_template = TEMPLATE
     tokenizer.save_pretrained(args.save_path)

examples/grpo/cosyvoice2/run.sh

@@ -3,7 +3,7 @@
 set -eou pipefail
 
 stage=-1
-stop_stage=5
+stop_stage=4
 
 log() {
   # This function is from espnet
@@ -15,6 +15,22 @@ export PYTHONPATH=/workspace/CosyVoice
 model_scope_model_path=./CosyVoice2-0.5B
 sft_model_path=./transformers_cosyvoice2_llm
 
+if [ $stage -le -2 ] && [ $stop_stage -ge -2 ]; then
+  log "stage -2: install dependencies locally if pre-built docker image is not available"
+  conda create -n cosyvoice2 python=3.10 -y
+  conda activate cosyvoice2
+    # install verl
+  git clone https://github.com/yuekaizhang/verl.git -b thread
+  cd verl
+  USE_MEGATRON=0 bash scripts/install_vllm_sglang_mcore.sh
+  pip install --no-deps -e .
+  cd -
+  # install requirements
+  pip install -r requirements.txt
+  pip install -U nvidia-pytriton
+  git clone https://github.com/yuekaizhang/PytritonSenseVoice.git && cd PytritonSenseVoice && pip install -e .
+fi
+
 if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
   log "stage -1: download official CosyVoice2-0.5B LLM model and convert to huggingface compatible checkpoint"
   modelscope download --model iic/CosyVoice2-0.5B --local_dir $model_scope_model_path 
@@ -24,13 +40,15 @@ if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
 
   # Or, you could use the following command to download the huggingface compatible checkpoint
   # huggingface-cli download --local-dir $sft_model_path yuekai/cosyvoice2_llm
+
+  # Note: we remove the lm_head's bias to make it compatible with the Qwen2.5-0.5B model in Transformers.
 fi
 
 data_dir=data/parquet_aishell3
 if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
   log "stage 0: prepare data into verl format"
   mkdir -p $data_dir
-  wget https://huggingface.co/datasets/SparkAudio/voxbox/resolve/main/metadata/aishell-3.jsonl -O data/aishell-3.jsonl
+  wget -O data/aishell-3.jsonl https://huggingface.co/datasets/SparkAudio/voxbox/resolve/main/metadata/aishell-3.jsonl
   # total 88035 samples
   head -n 80000 data/aishell-3.jsonl > data/train.jsonl
   tail -n 100 data/aishell-3.jsonl > data/test.jsonl
@@ -98,7 +116,8 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
       trainer.val_before_train=False
 fi
 
-step=400
+steps=(100 200 300 400 500)
+for step in ${steps[@]}; do
 llm_path=./checkpoints/cosyvoice2_grpo/$exp_name/global_step_${step}
 if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   log "stage 3: merge the model"
@@ -111,7 +130,7 @@ fi
 if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
   log "stage 4: Test the model"
   dataset=zero_shot_zh
-  # dataset=test_zh
+  # dataset=test_zh seed_tts test_zh
   output_dir=./outputs_${exp_name}_${step}_${dataset}
 
   token2wav_path=/workspace/CosyVoice2-0.5B
@@ -127,12 +146,14 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
 
   bash scripts/compute_wer.sh $output_dir ${dataset}
 fi
+done
 
 if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
   log "stage 5: Convert the RL trained model to CosyVoice repo format"
   python3 huggingface_to_pretrained.py \
     --hf-cosyvoice2-llm-path $llm_path/merged_hf_model \
-    --pretrained-cosyvoice2-path /workspace/CosyVoice2-0.5B \
     --output-path /workspace/CosyVoice2-0.5B/llm-new.pt
   # You need to manually move the llm-new.pt to overwrite /workspace/CosyVoice2-0.5B/llm.pt
+  # However, we found that the RL trained model accuracy would slightly drop after this conversion.
+  # Please be careful or use the huggingface format inference code.
 fi

examples/grpo/cosyvoice2/scripts/compute_wer.sh

@@ -10,6 +10,7 @@ model_path=models/sherpa-onnx-paraformer-zh-2023-09-14
 if [ ! -d $model_path ]; then
     pip install sherpa-onnx
     wget -nc https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-paraformer-zh-2023-09-14.tar.bz2
+    mkdir models
     tar xvf sherpa-onnx-paraformer-zh-2023-09-14.tar.bz2 -C models
 fi