fix decoupled mode

Author: yuekaiz

runtime/triton_trtllm/model_repo/cosyvoice2/1/model.py

@@ -295,11 +295,26 @@ def execute(self, requests):
             if self.decoupled:
                 response_sender = request.get_response_sender()
                 request_id = request.request_id()
-                for generated_ids in generated_ids_iter:
-                    raise NotImplementedError("Decoupled mode is not implemented")
+                generated_ids = []
+                for generated_id in generated_ids_iter:
+                    # convert the numpy array into a int32 tensor
+                    generated_id = generated_id.tolist()
+                    if len(generated_id) > 0:
+                        assert len(generated_id) == 1, "Generated ID is not a single integer"
+                        generated_ids.append(generated_id[0])
+                generated_ids = torch.tensor(generated_ids).unsqueeze(0).to(torch.int32).to(self.device)
+                prompt_spk_embedding = self._extract_spk_embedding(wav_tensor)
+                audio = self.forward_token2wav(prompt_speech_tokens, prompt_speech_feat, prompt_spk_embedding, generated_ids)
+                
+                # Prepare response
+                audio_tensor = pb_utils.Tensor.from_dlpack("waveform", to_dlpack(audio))
+                inference_response = pb_utils.InferenceResponse(output_tensors=[audio_tensor])
+                response_sender.send(inference_response)
+                response_sender.send(flags=pb_utils.TRITONSERVER_RESPONSE_COMPLETE_FINAL)
+                self.logger.log_info(f"send tritonserver_response_complete_final to end")
             else:
                 generated_ids = next(generated_ids_iter)
-                generated_ids = torch.tensor([generated_ids]).to(self.device)
+                generated_ids = torch.tensor(generated_ids).unsqueeze(0).to(self.device)
                 if generated_ids is None or len(generated_ids) == 0:
                     raise pb_utils.TritonModelException("Generated IDs is None or empty")
 
@@ -311,9 +326,5 @@ def execute(self, requests):
                 inference_response = pb_utils.InferenceResponse(output_tensors=[audio_tensor])
                 responses.append(inference_response)
 
-            if self.decoupled:
-                response_sender.send(flags=pb_utils.TRITONSERVER_RESPONSE_COMPLETE_FINAL)
-                self.logger.log_info(f"send tritonserver_response_complete_final to end")
-        
         if not self.decoupled:
             return responses

runtime/triton_trtllm/requirements.txt

@@ -10,4 +10,5 @@ wget
 librosa
 pyworld
 openai-whisper
-tritonclient
+tritonclient
+modelscope

runtime/triton_trtllm/run.sh

@@ -1,25 +1,31 @@
 
 export CUDA_VISIBLE_DEVICES=0
-export PYTHONPATH=/home/scratch.yuekaiz_wwfo_1/tts/cosyvoice/CosyVoice:$PYTHONPATH
-export PYTHONPATH=/home/scratch.yuekaiz_wwfo_1/tts/cosyvoice/CosyVoice/third_party/Matcha-TTS:$PYTHONPATH
+cosyvoice_path=/workspace/CosyVoice
+export PYTHONPATH=${cosyvoice_path}:$PYTHONPATH
+export PYTHONPATH=${cosyvoice_path}/third_party/Matcha-TTS:$PYTHONPATH
 stage=$1
 stop_stage=$2
 
-huggingface_model_local_dir=/home/scratch.yuekaiz_wwfo_1/tts/cosyvoice/cosyvoice2_llm
-model_scope_model_local_dir=/home/scratch.yuekaiz_wwfo_1/tts/cosyvoice/CosyVoice2-0.5B
+huggingface_model_local_dir=./cosyvoice2_llm
+model_scope_model_local_dir=./CosyVoice2-0.5B
 trt_dtype=bfloat16
-trt_weights_dir=/home/scratch.yuekaiz_wwfo_1/tts/cosyvoice/trt_weights_${trt_dtype}
-trt_engines_dir=/home/scratch.yuekaiz_wwfo_1/tts/cosyvoice/trt_engines_${trt_dtype}
+trt_weights_dir=./trt_weights_${trt_dtype}
+trt_engines_dir=./trt_engines_${trt_dtype}
 
 model_repo=./model_repo_cosyvoice2
 
-if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
-    echo " "
-    huggingface-cli download --local-dir cosyvoice2_llm yuekai/cosyvoice2_llm 
-    modelscope download --model iic/CosyVoice2-0.5B --local_dir ./CosyVoice2-0.5B/ 
-    git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
-    cd CosyVoice
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+    echo "Cloning CosyVoice"
+    git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git $cosyvoice_path
+    cd $cosyvoice_path
     git submodule update --init --recursive
+    cd runtime/triton_trtllm
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+    echo "Downloading CosyVoice2-0.5B"
+    huggingface-cli download --local-dir $huggingface_model_local_dir yuekai/cosyvoice2_llm 
+    modelscope download --model iic/CosyVoice2-0.5B --local_dir $model_scope_model_local_dir 
 fi
 
 
@@ -35,17 +41,15 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
                 --max_batch_size 16 \
                 --max_num_tokens 32768 \
                 --gemm_plugin $trt_dtype || exit 1
-fi
 
-if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
     echo "Testing TensorRT engines"
-    python3 ./test_llm.py --input_text "你好，请问你叫什么？" \
+    python3 ./scripts/test_llm.py --input_text "你好，请问你叫什么？" \
                     --tokenizer_dir $huggingface_model_local_dir \
                     --top_k 50 --top_p 0.95 --temperature 0.8 \
                     --engine_dir=$trt_engines_dir  || exit 1
 fi
 
-if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
     echo "Creating model repository"
     rm -rf $model_repo
     mkdir -p $model_repo
@@ -71,28 +75,31 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
 
 fi
 
-if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
-
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+   echo "Starting Triton server"
    tritonserver --model-repository $model_repo
 fi
 
-if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
-    echo "Testing TensorRT engines"
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+    echo "Single request test http"
     python3 client_http.py \
-        --reference-audio ./prompt_audio.wav \
+        --reference-audio ./assets/prompt_audio.wav \
         --reference-text "吃燕窝就选燕之屋，本节目由26年专注高品质燕窝的燕之屋冠名播出。豆奶牛奶换着喝，营养更均衡，本节目由豆本豆豆奶特约播出。" \
         --target-text "身临其境，换新体验。塑造开源语音合成新范式，让智能语音更自然。" \
         --model-name cosyvoice2
 fi
 
-if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
-    echo "Running benchmark client"
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+    echo "Running benchmark client grpc"
     num_task=4
+    # set mode=streaming, when decoupled=True
+    # set mode=offline, when decoupled=False
+    mode=offline
     python3 client_grpc.py \
         --server-addr localhost \
         --model-name cosyvoice2 \
         --num-tasks $num_task \
-        --mode offline \
+        --mode $mode \
         --huggingface-dataset yuekai/seed_tts_cosy2 \
-        --log-dir ./log_concurrent_tasks_${num_task}_offline_bls_4_${trt_dtype}
+        --log-dir ./log_concurrent_tasks_${num_task}_${mode}_bls_4_${trt_dtype}
 fi