add flow decoder cache

examples/libritts/cosyvoice2/conf/cosyvoice2.yaml

@@ -0,0 +1,232 @@
+# set random seed, so that you may reproduce your result.
+__set_seed1: !apply:random.seed [1986]
+__set_seed2: !apply:numpy.random.seed [1986]
+__set_seed3: !apply:torch.manual_seed [1986]
+__set_seed4: !apply:torch.cuda.manual_seed_all [1986]
+
+# fixed params
+sample_rate: 24000
+llm_input_size: 896
+llm_output_size: 896
+spk_embed_dim: 192
+qwen_pretrain_path: ''
+token_frame_rate: 25
+token_mel_ratio: 2
+
+# model params
+# for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
+# for system/third_party class/function, we do not require this.
+llm: !new:cosyvoice.llm.llm.Qwen2LM
+    llm_input_size: !ref <llm_input_size>
+    llm_output_size: !ref <llm_output_size>
+    speech_token_size: 6561
+    length_normalized_loss: True
+    lsm_weight: 0
+    mix_ratio: [5, 15]
+    llm: !new:cosyvoice.llm.llm.Qwen2Encoder
+        pretrain_path: !ref <qwen_pretrain_path>
+    sampling: !name:cosyvoice.utils.common.ras_sampling
+        top_p: 0.8
+        top_k: 25
+        win_size: 10
+        tau_r: 0.1
+
+flow: !new:cosyvoice.flow.flow.CausalMaskedDiffWithXvec
+    input_size: 512
+    output_size: 80
+    spk_embed_dim: !ref <spk_embed_dim>
+    output_type: 'mel'
+    vocab_size: 6561
+    input_frame_rate: !ref <token_frame_rate>
+    only_mask_loss: True
+    token_mel_ratio: !ref <token_mel_ratio>
+    pre_lookahead_len: 3
+    encoder: !new:cosyvoice.transformer.upsample_encoder.UpsampleConformerEncoder
+        output_size: 512
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 6
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.1
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        input_size: 512
+        use_cnn_module: False
+        macaron_style: False
+        use_dynamic_chunk: True
+    decoder: !new:cosyvoice.flow.flow_matching.CausalConditionalCFM
+        in_channels: 240
+        n_spks: 1
+        spk_emb_dim: 80
+        cfm_params: !new:omegaconf.DictConfig
+            content:
+                sigma_min: 1e-06
+                solver: 'euler'
+                t_scheduler: 'cosine'
+                training_cfg_rate: 0.2
+                inference_cfg_rate: 0.7
+                reg_loss_type: 'l1'
+        estimator: !new:cosyvoice.flow.decoder.CausalConditionalDecoder
+            in_channels: 320
+            out_channels: 80
+            channels: [256]
+            dropout: 0.0
+            attention_head_dim: 64
+            n_blocks: 4
+            num_mid_blocks: 12
+            num_heads: 8
+            act_fn: 'gelu'
+            static_chunk_size: !ref <token_frame_rate> * <token_mel_ratio> # here we use static_chunk_size because we want to fix kv cache size during inference
+            num_decoding_left_chunks: 2
+
+hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+    in_channels: 80
+    base_channels: 512
+    nb_harmonics: 8
+    sampling_rate: !ref <sample_rate>
+    nsf_alpha: 0.1
+    nsf_sigma: 0.003
+    nsf_voiced_threshold: 10
+    upsample_rates: [8, 5, 3]
+    upsample_kernel_sizes: [16, 11, 7]
+    istft_params:
+        n_fft: 16
+        hop_len: 4
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    source_resblock_kernel_sizes: [7, 7, 11]
+    source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    lrelu_slope: 0.1
+    audio_limit: 0.99
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+        num_class: 1
+        in_channels: 80
+        cond_channels: 512
+
+# gan related module
+mel_spec_transform1: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: null
+    center: False
+hifigan: !new:cosyvoice.hifigan.hifigan.HiFiGan
+    generator: !ref <hift>
+    discriminator: !new:cosyvoice.hifigan.discriminator.MultipleDiscriminator
+        mpd: !new:matcha.hifigan.models.MultiPeriodDiscriminator
+        mrd: !new:cosyvoice.hifigan.discriminator.MultiResolutionDiscriminator
+    mel_spec_transform: [
+        !ref <mel_spec_transform1>
+    ]
+
+# processor functions
+parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
+get_tokenizer: !name:cosyvoice.tokenizer.tokenizer.get_qwen_tokenizer
+    token_path: !ref <qwen_pretrain_path>
+    skip_special_tokens: True
+allowed_special: 'all'
+tokenize: !name:cosyvoice.dataset.processor.tokenize
+    get_tokenizer: !ref <get_tokenizer>
+    allowed_special: !ref <allowed_special>
+filter: !name:cosyvoice.dataset.processor.filter
+    max_length: 40960
+    min_length: 100
+    token_max_length: 200
+    token_min_length: 1
+resample: !name:cosyvoice.dataset.processor.resample
+    resample_rate: !ref <sample_rate>
+truncate: !name:cosyvoice.dataset.processor.truncate
+    truncate_length: 24480 # must be a multiplier of hop_size
+feat_extractor: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1920
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 480
+    win_size: 1920
+    fmin: 0
+    fmax: 8000
+    center: False
+compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
+    feat_extractor: !ref <feat_extractor>
+# pitch_extractor: !name:torchaudio.functional.compute_kaldi_pitch # TODO need to replace it
+#     sample_rate: !ref <sample_rate>
+#     frame_length: 46.4 # match feat_extractor win_size/sampling_rate
+#     frame_shift: 11.6 # match feat_extractor hop_size/sampling_rate
+# compute_f0: !name:cosyvoice.dataset.processor.compute_f0
+#     pitch_extractor: !ref <pitch_extractor>
+parse_embedding: !name:cosyvoice.dataset.processor.parse_embedding
+    normalize: True
+shuffle: !name:cosyvoice.dataset.processor.shuffle
+    shuffle_size: 1000
+sort: !name:cosyvoice.dataset.processor.sort
+    sort_size: 500  # sort_size should be less than shuffle_size
+batch: !name:cosyvoice.dataset.processor.batch
+    batch_type: 'dynamic'
+    max_frames_in_batch: 2500
+padding: !name:cosyvoice.dataset.processor.padding
+    use_spk_embedding: False # change to True during sft
+
+
+# dataset processor pipeline
+data_pipeline: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <compute_fbank>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+# data_pipeline_gan: [
+#     !ref <parquet_opener>,
+#     !ref <tokenize>,
+#     !ref <filter>,
+#     !ref <resample>,
+#     !ref <truncate>,
+#     !ref <compute_fbank>,
+#     !ref <compute_f0>,
+#     !ref <parse_embedding>,
+#     !ref <shuffle>,
+#     !ref <sort>,
+#     !ref <batch>,
+#     !ref <padding>,
+# ]
+
+# llm flow train conf
+train_conf:
+    optim: adam
+    optim_conf:
+        lr: 1e-5 # change to 1e-5 during sft
+    scheduler: constantlr # change to constantlr during sft
+    scheduler_conf:
+        warmup_steps: 2500
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 2
+    log_interval: 100
+    save_per_step: -1
+
+# gan train conf
+train_conf_gan:
+    optim: adam
+    optim_conf:
+        lr: 0.0002 # use small lr for gan training
+    scheduler: constantlr
+    optim_d: adam
+    optim_conf_d:
+        lr: 0.0002 # use small lr for gan training
+    scheduler_d: constantlr
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 1 # in gan training, accum_grad must be 1
+    log_interval: 100
+    save_per_step: -1

examples/libritts/cosyvoice2/conf/ds_stage2.json

@@ -0,0 +1,42 @@
+{
+  "train_micro_batch_size_per_gpu": 1,
+  "gradient_accumulation_steps": 1,
+  "steps_per_print": 100,
+  "gradient_clipping": 5,
+  "fp16": {
+    "enabled": false,
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 16,
+    "loss_scale_window": 256,
+    "hysteresis": 2,
+    "consecutive_hysteresis": false,
+    "min_loss_scale": 1
+  },
+  "bf16": {
+    "enabled": false
+  },
+  "zero_force_ds_cpu_optimizer": false,
+  "zero_optimization": {
+    "stage": 2,
+    "offload_optimizer": {
+      "device": "none",
+      "pin_memory": true
+    },
+    "allgather_partitions": true,
+    "allgather_bucket_size": 5e8,
+    "overlap_comm": false,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 5e8,
+    "contiguous_gradients" : true
+  },
+  "optimizer": {
+    "type": "AdamW",
+    "params": {
+        "lr": 0.001,
+        "weight_decay": 0.0001,
+        "torch_adam": true,
+        "adam_w_mode": true
+    }
+  }
+}

examples/libritts/cosyvoice2/path.sh

@@ -0,0 +1,3 @@
+# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
+export PYTHONIOENCODING=UTF-8
+export PYTHONPATH=../../../:../../../third_party/Matcha-TTS:$PYTHONPATH

examples/libritts/cosyvoice2/run.sh

@@ -0,0 +1,130 @@
+#!/bin/bash
+# Copyright 2024 Alibaba Inc. All Rights Reserved.
+. ./path.sh || exit 1;
+
+stage=-1
+stop_stage=3
+
+data_url=www.openslr.org/resources/60
+data_dir=/mnt/lyuxiang.lx/data/tts/openslr/libritts
+pretrained_model_dir=/mnt/lyuxiang.lx/data/tts/models/IIC/CosyVoice2-0.5B/
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+  echo "Data Download"
+  for part in dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500; do
+    local/download_and_untar.sh ${data_dir} ${data_url} ${part}
+  done
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+  echo "Data preparation, prepare wav.scp/text/utt2spk/spk2utt"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    mkdir -p data/$x
+    python local/prepare_data.py --src_dir $data_dir/LibriTTS/$x --des_dir data/$x
+  done
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+  echo "Extract campplus speaker embedding, you will get spk2embedding.pt and utt2embedding.pt in data/$x dir"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    tools/extract_embedding.py --dir data/$x \
+      --onnx_path $pretrained_model_dir/campplus.onnx
+  done
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+  echo "Extract discrete speech token, you will get utt2speech_token.pt in data/$x dir"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    tools/extract_speech_token.py --dir data/$x \
+      --onnx_path $pretrained_model_dir/speech_tokenizer_v2.onnx
+  done
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+  echo "Prepare required parquet format data, you should have prepared wav.scp/text/utt2spk/spk2utt/utt2embedding.pt/spk2embedding.pt/utt2speech_token.pt"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    mkdir -p data/$x/parquet
+    tools/make_parquet_list.py --num_utts_per_parquet 1000 \
+      --num_processes 10 \
+      --src_dir data/$x \
+      --des_dir data/$x/parquet
+  done
+fi
+
+# inference
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+  echo "Run inference. Please make sure utt in tts_text is in prompt_data"
+  # TODO consider remove bin/inference.py, or use similar initilization method as in readme
+  for mode in sft zero_shot; do
+    python cosyvoice/bin/inference.py --mode $mode \
+      --gpu 0 \
+      --config conf/cosyvoice2.yaml \
+      --prompt_data data/test-clean/parquet/data.list \
+      --prompt_utt2data data/test-clean/parquet/utt2data.list \
+      --tts_text `pwd`/tts_text.json \
+      --qwen_pretrain_path $pretrained_model_dir/CosyVoice-BlankEN \
+      --llm_model $pretrained_model_dir/llm.pt \
+      --flow_model $pretrained_model_dir/flow.pt \
+      --hifigan_model $pretrained_model_dir/hift.pt \
+      --result_dir `pwd`/exp/cosyvoice/test-clean/$mode
+  done
+fi
+
+# train llm
+export CUDA_VISIBLE_DEVICES="2,3,4,5,6,7"
+num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+job_id=1986
+dist_backend="nccl"
+num_workers=2
+prefetch=100
+train_engine=torch_ddp
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+  echo "Run train. We only support llm traning for now. If your want to train from scratch, please use conf/cosyvoice.fromscratch.yaml"
+  if [ $train_engine == 'deepspeed' ]; then
+    echo "Notice deepspeed has its own optimizer config. Modify conf/ds_stage2.json if necessary"
+  fi
+  cat data/{train-clean-100,train-clean-360,train-other-500}/parquet/data.list > data/train.data.list
+  cat data/{dev-clean,dev-other}/parquet/data.list > data/dev.data.list
+  # NOTE will update llm/hift training later
+  for model in flow; do
+    torchrun --nnodes=1 --nproc_per_node=$num_gpus \
+        --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:1234" \
+      cosyvoice/bin/train.py \
+      --train_engine $train_engine \
+      --config conf/cosyvoice2.yaml \
+      --train_data data/train.data.list \
+      --cv_data data/dev.data.list \
+      --qwen_pretrain_path $pretrained_model_dir/CosyVoice-BlankEN \
+      --model $model \
+      --checkpoint $pretrained_model_dir/$model.pt \
+      --model_dir `pwd`/exp/cosyvoice2/$model/$train_engine \
+      --tensorboard_dir `pwd`/tensorboard/cosyvoice2/$model/$train_engine \
+      --ddp.dist_backend $dist_backend \
+      --num_workers ${num_workers} \
+      --prefetch ${prefetch} \
+      --pin_memory \
+      --use_amp \
+      --deepspeed_config ./conf/ds_stage2.json \
+      --deepspeed.save_states model+optimizer
+  done
+fi
+
+# average model
+average_num=5
+if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
+  for model in llm flow hifigan; do
+    decode_checkpoint=`pwd`/exp/cosyvoice/$model/$train_engine/${model}.pt
+    echo "do model average and final checkpoint is $decode_checkpoint"
+    python cosyvoice/bin/average_model.py \
+      --dst_model $decode_checkpoint \
+      --src_path `pwd`/exp/cosyvoice/$model/$train_engine  \
+      --num ${average_num} \
+      --val_best
+  done
+fi
+
+if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
+  echo "Export your model for inference speedup. Remember copy your llm or flow model to model_dir"
+  python cosyvoice/bin/export_jit.py --model_dir $pretrained_model_dir
+  python cosyvoice/bin/export_onnx.py --model_dir $pretrained_model_dir
+fi

examples/libritts/cosyvoice2/tts_text.json

@@ -0,0 +1,5 @@
+{
+  "1089_134686_000002_000000": [
+    "hello, my name is Jack. What is your name?"
+  ]
+}