Saving train state of step 10000

.gitignore

@@ -0,0 +1 @@
+wandb

accelerate_config.yaml

@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

checkpoint-10000-epoch-0/config.json

@@ -0,0 +1,278 @@
+{
+  "architectures": [
+    "ParlerTTSForConditionalGeneration"
+  ],
+  "audio_encoder": {
+    "_name_or_path": "parler-tts/dac_44khZ_8kbps",
+    "add_cross_attention": false,
+    "architectures": [
+      "DACModel"
+    ],
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "codebook_size": 1024,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "frame_rate": 86,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "latent_dim": 1024,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_bitrate": 8,
+    "model_type": "dac",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codebooks": 9,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sampling_rate": 44100,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false
+  },
+  "decoder": {
+    "_name_or_path": "./parler-tts-untrained-600M/decoder",
+    "activation_dropout": 0.0,
+    "activation_function": "gelu",
+    "add_cross_attention": true,
+    "architectures": [
+      "ParlerTTSForCausalLM"
+    ],
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 1025,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 1024,
+    "exponential_decay_length_penalty": null,
+    "ffn_dim": 4096,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_factor": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layerdrop": 0.0,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 4096,
+    "min_length": 0,
+    "model_type": "parler_tts_decoder",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codebooks": 9,
+    "num_hidden_layers": 24,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 1024,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rope_embeddings": false,
+    "rope_theta": 10000.0,
+    "scale_embedding": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 1088
+  },
+  "decoder_start_token_id": 1025,
+  "is_encoder_decoder": true,
+  "model_type": "parler_tts",
+  "pad_token_id": 1024,
+  "prompt_cross_attention": true,
+  "text_encoder": {
+    "_name_or_path": "google/flan-t5-base",
+    "add_cross_attention": false,
+    "architectures": [
+      "T5ForConditionalGeneration"
+    ],
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": null,
+    "d_ff": 2048,
+    "d_kv": 64,
+    "d_model": 768,
+    "decoder_start_token_id": 0,
+    "dense_act_fn": "gelu_new",
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout_rate": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 1,
+    "exponential_decay_length_penalty": null,
+    "feed_forward_proj": "gated-gelu",
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_factor": 1.0,
+    "is_decoder": false,
+    "is_encoder_decoder": true,
+    "is_gated_act": true,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_epsilon": 1e-06,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "t5",
+    "n_positions": 512,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_decoder_layers": 12,
+    "num_heads": 12,
+    "num_layers": 12,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_past": true,
+    "output_scores": false,
+    "pad_token_id": 0,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "relative_attention_max_distance": 128,
+    "relative_attention_num_buckets": 32,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": {
+      "summarization": {
+        "early_stopping": true,
+        "length_penalty": 2.0,
+        "max_length": 200,
+        "min_length": 30,
+        "no_repeat_ngram_size": 3,
+        "num_beams": 4,
+        "prefix": "summarize: "
+      },
+      "translation_en_to_de": {
+        "early_stopping": true,
+        "max_length": 300,
+        "num_beams": 4,
+        "prefix": "translate English to German: "
+      },
+      "translation_en_to_fr": {
+        "early_stopping": true,
+        "max_length": 300,
+        "num_beams": 4,
+        "prefix": "translate English to French: "
+      },
+      "translation_en_to_ro": {
+        "early_stopping": true,
+        "max_length": 300,
+        "num_beams": 4,
+        "prefix": "translate English to Romanian: "
+      }
+    },
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 32128
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.40.2",
+  "vocab_size": 32128
+}

checkpoint-10000-epoch-0/generation_config.json

@@ -0,0 +1,12 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1025,
+  "decoder_start_token_id": 1025,
+  "do_sample": true,
+  "eos_token_id": 1024,
+  "guidance_scale": 1,
+  "key": 10,
+  "max_length": 2580,
+  "pad_token_id": 1024,
+  "transformers_version": "4.40.2"
+}

checkpoint-10000-epoch-0/optimizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f593409051de29ef25a2113beb7cc0a9d2ba8e3990cdb2d3bec1eaa0884c7498
+size 3652763351

checkpoint-10000-epoch-0/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be131ce3d6964f98937dcfc065b75714e57c75d5efaa7c8801f6d9ce70949801
+size 2605239710

checkpoint-10000-epoch-0/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eb6348546ae6739037c597fbe0ef112363cfbe7960ba922bac0bf7286c1a843
+size 16036

checkpoint-10000-epoch-0/random_states_1.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c25798eb13128ebe68f7cb7e47010120927add7afcd7b129f8bc130d47e9f4d
+size 16036

checkpoint-10000-epoch-0/random_states_2.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88f5dca57b8bc684f334df9771d2b260e550ae95594b7c789f1c6fb9ea21830e
+size 16036

checkpoint-10000-epoch-0/random_states_3.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b681065b5ba34e6fc0336e31e2c0e351954153e6047d5c5451d59367de488486
+size 16036

checkpoint-10000-epoch-0/random_states_4.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f18f3555b7d8fe146fd5c9b755b9acd9a82584fd5f817792f08f018b1d50cebf
+size 16036

checkpoint-10000-epoch-0/random_states_5.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9661ec00571ae2af4064968895b8a8ea6e6bb4efde85fbb2e7140ce64b1251f6
+size 16036

checkpoint-10000-epoch-0/random_states_6.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:249678fbd844581d90ce2d0bdfd36361b09432dcd3e93305092b2441c58c202a
+size 16036

checkpoint-10000-epoch-0/random_states_7.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1c9f431217170b785ae8f1b56b146aaf06283b7938f562d07df8354863f0857
+size 16036

checkpoint-10000-epoch-0/scheduler.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed0f8986a3a4e57c81aefc0174bd43cad9587978e99fdf28b8483089a1eb2d3d
+size 1000

config.json

@@ -0,0 +1,278 @@
+{
+  "architectures": [
+    "ParlerTTSForConditionalGeneration"
+  ],
+  "audio_encoder": {
+    "_name_or_path": "parler-tts/dac_44khZ_8kbps",
+    "add_cross_attention": false,
+    "architectures": [
+      "DACModel"
+    ],
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "codebook_size": 1024,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "frame_rate": 86,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "latent_dim": 1024,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_bitrate": 8,
+    "model_type": "dac",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codebooks": 9,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sampling_rate": 44100,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false
+  },
+  "decoder": {
+    "_name_or_path": "./parler-tts-untrained-600M/decoder",
+    "activation_dropout": 0.0,
+    "activation_function": "gelu",
+    "add_cross_attention": true,
+    "architectures": [
+      "ParlerTTSForCausalLM"
+    ],
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 1025,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 1024,
+    "exponential_decay_length_penalty": null,
+    "ffn_dim": 4096,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_factor": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layerdrop": 0.0,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 4096,
+    "min_length": 0,
+    "model_type": "parler_tts_decoder",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codebooks": 9,
+    "num_hidden_layers": 24,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 1024,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rope_embeddings": false,
+    "rope_theta": 10000.0,
+    "scale_embedding": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 1088
+  },
+  "decoder_start_token_id": 1025,
+  "is_encoder_decoder": true,
+  "model_type": "parler_tts",
+  "pad_token_id": 1024,
+  "prompt_cross_attention": true,
+  "text_encoder": {
+    "_name_or_path": "google/flan-t5-base",
+    "add_cross_attention": false,
+    "architectures": [
+      "T5ForConditionalGeneration"
+    ],
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": null,
+    "d_ff": 2048,
+    "d_kv": 64,
+    "d_model": 768,
+    "decoder_start_token_id": 0,
+    "dense_act_fn": "gelu_new",
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout_rate": 0.1,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 1,
+    "exponential_decay_length_penalty": null,
+    "feed_forward_proj": "gated-gelu",
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_factor": 1.0,
+    "is_decoder": false,
+    "is_encoder_decoder": true,
+    "is_gated_act": true,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_epsilon": 1e-06,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "t5",
+    "n_positions": 512,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_decoder_layers": 12,
+    "num_heads": 12,
+    "num_layers": 12,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_past": true,
+    "output_scores": false,
+    "pad_token_id": 0,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "relative_attention_max_distance": 128,
+    "relative_attention_num_buckets": 32,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": {
+      "summarization": {
+        "early_stopping": true,
+        "length_penalty": 2.0,
+        "max_length": 200,
+        "min_length": 30,
+        "no_repeat_ngram_size": 3,
+        "num_beams": 4,
+        "prefix": "summarize: "
+      },
+      "translation_en_to_de": {
+        "early_stopping": true,
+        "max_length": 300,
+        "num_beams": 4,
+        "prefix": "translate English to German: "
+      },
+      "translation_en_to_fr": {
+        "early_stopping": true,
+        "max_length": 300,
+        "num_beams": 4,
+        "prefix": "translate English to French: "
+      },
+      "translation_en_to_ro": {
+        "early_stopping": true,
+        "max_length": 300,
+        "num_beams": 4,
+        "prefix": "translate English to Romanian: "
+      }
+    },
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 32128
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.40.2",
+  "vocab_size": 32128
+}

parler_tts/__init__.py

@@ -0,0 +1,16 @@
+__version__ = "0.1"
+
+
+from .configuration_parler_tts import ParlerTTSConfig, ParlerTTSDecoderConfig
+from .modeling_parler_tts import (
+    ParlerTTSForCausalLM,
+    ParlerTTSForConditionalGeneration,
+    apply_delay_pattern_mask,
+    build_delay_pattern_mask,
+)
+
+from .dac_wrapper import DACConfig, DACModel
+from transformers import AutoConfig, AutoModel
+
+AutoConfig.register("dac", DACConfig)
+AutoModel.register(DACConfig, DACModel)

parler_tts/configuration_parler_tts.py

@@ -0,0 +1,249 @@
+# coding=utf-8
+# Copyright 2024 and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+""" Parler-TTS model configuration"""
+
+from transformers import AutoConfig, logging
+from transformers.configuration_utils import PretrainedConfig
+
+
+logger = logging.get_logger(__name__)
+
+MUSICGEN_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/parler_tts-small": "https://huggingface.co/facebook/parler_tts-small/resolve/main/config.json",
+    # See all ParlerTTS models at https://huggingface.co/models?filter=parler_tts
+}
+
+
+class ParlerTTSDecoderConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of an [`ParlerTTSDecoder`]. It is used to instantiate a
+    Parler-TTS decoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the Parler-TTS
+    [facebook/parler_tts-small](https://huggingface.co/facebook/parler_tts-small) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 2049):
+            Vocabulary size of the ParlerTTSDecoder model. Defines the number of different tokens that can be
+            represented by the `inputs_ids` passed when calling [`ParlerTTSDecoder`].
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of decoder layers.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer block.
+        ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer block.
+        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the decoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, text_encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically, set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        initializer_factor (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        scale_embedding (`bool`, *optional*, defaults to `False`):
+            Scale embeddings by diving by sqrt(hidden_size).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether the model should return the last key/values attentions (not used by all models)
+        num_codebooks (`int`, *optional*, defaults to 4):
+            The number of parallel codebooks forwarded to the model.
+        tie_word_embeddings(`bool`, *optional*, defaults to `False`):
+            Whether input and output word embeddings should be tied.
+        rope_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to use ROPE or absolute positional embeddings.
+        rope_theta (`float`, *optional*, defaults to 100000.0):
+            The base period of the RoPE embeddings.
+    """
+
+    model_type = "parler_tts_decoder"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=2049,  # vocab size = 2048 (encodec vocab size) + 1 (eos)
+        max_position_embeddings=2048,
+        num_hidden_layers=24,
+        ffn_dim=4096,
+        num_attention_heads=16,
+        layerdrop=0.0,
+        use_cache=True,
+        activation_function="gelu",
+        hidden_size=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        initializer_factor=0.02,
+        scale_embedding=False,
+        num_codebooks=4,
+        pad_token_id=2048,
+        bos_token_id=2049,
+        eos_token_id=2048,
+        tie_word_embeddings=False,
+        rope_embeddings=False,
+        rope_theta=10_000.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.ffn_dim = ffn_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.initializer_factor = initializer_factor
+        self.layerdrop = layerdrop
+        self.use_cache = use_cache
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+        self.num_codebooks = num_codebooks
+        self.rope_embeddings = rope_embeddings
+        self.rope_theta = rope_theta
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+class ParlerTTSConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ParlerTTSModel`]. It is used to instantiate a
+    Parler-TTS model according to the specified arguments, defining the text encoder, audio encoder and Parler-TTS decoder
+    configs.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 1024):
+            Vocabulary size of the prompt token ids. Defines the number of different tokens that can be
+            represented by the `prompt_inputs_ids`.
+        prompt_cross_attention (`bool`, *optional*, defaults to `False`):
+            Whether to use cross-attention conditioning for the prompt (as well as the description).
+        kwargs (*optional*):
+            Dictionary of keyword arguments. Notably:
+
+                - **text_encoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that
+                  defines the text encoder config.
+                - **audio_encoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that
+                  defines the audio encoder config.
+                - **decoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that defines
+                  the decoder config.
+
+    Example:
+
+    ```python
+    >>> from transformers import (
+    ...     ParlerTTSConfig,
+    ...     ParlerTTSDecoderConfig,
+    ...     T5Config,
+    ...     EncodecConfig,
+    ...     ParlerTTSForConditionalGeneration,
+    ... )
+
+    >>> # Initializing text encoder, audio encoder, and decoder model configurations
+    >>> text_encoder_config = T5Config()
+    >>> audio_encoder_config = EncodecConfig()
+    >>> decoder_config = ParlerTTSDecoderConfig()
+
+    >>> configuration = ParlerTTSConfig.from_sub_models_config(
+    ...     text_encoder_config, audio_encoder_config, decoder_config
+    ... )
+
+    >>> # Initializing a ParlerTTSForConditionalGeneration (with random weights) from the facebook/parler_tts-small style configuration
+    >>> model = ParlerTTSForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    >>> config_text_encoder = model.config.text_encoder
+    >>> config_audio_encoder = model.config.audio_encoder
+    >>> config_decoder = model.config.decoder
+
+    >>> # Saving the model, including its configuration
+    >>> model.save_pretrained("parler_tts-model")
+
+    >>> # loading model and config from pretrained folder
+    >>> parler_tts_config = ParlerTTSConfig.from_pretrained("parler_tts-model")
+    >>> model = ParlerTTSForConditionalGeneration.from_pretrained("parler_tts-model", config=parler_tts_config)
+    ```"""
+
+    model_type = "parler_tts"
+    is_composition = True
+
+    def __init__(self, vocab_size=1024, prompt_cross_attention=False, **kwargs):
+        super().__init__(**kwargs)
+        if "text_encoder" not in kwargs or "audio_encoder" not in kwargs or "decoder" not in kwargs:
+            raise ValueError("Config has to be initialized with text_encoder, audio_encoder and decoder config")
+
+        text_encoder_config = kwargs.pop("text_encoder")
+        text_encoder_model_type = text_encoder_config.pop("model_type")
+
+        audio_encoder_config = kwargs.pop("audio_encoder")
+        audio_encoder_model_type = audio_encoder_config.pop("model_type")
+
+        decoder_config = kwargs.pop("decoder")
+
+        self.vocab_size = vocab_size
+        self.prompt_cross_attention = prompt_cross_attention
+        self.text_encoder = AutoConfig.for_model(text_encoder_model_type, **text_encoder_config)
+        self.audio_encoder = AutoConfig.for_model(audio_encoder_model_type, **audio_encoder_config)
+        self.decoder = ParlerTTSDecoderConfig(**decoder_config)
+        self.is_encoder_decoder = True
+
+    @classmethod
+    def from_sub_models_config(
+        cls,
+        text_encoder_config: PretrainedConfig,
+        audio_encoder_config: PretrainedConfig,
+        decoder_config: ParlerTTSDecoderConfig,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`ParlerTTSConfig`] (or a derived class) from text encoder, audio encoder and decoder
+        configurations.
+
+        Returns:
+            [`ParlerTTSConfig`]: An instance of a configuration object
+        """
+
+        return cls(
+            text_encoder=text_encoder_config.to_dict(),
+            audio_encoder=audio_encoder_config.to_dict(),
+            decoder=decoder_config.to_dict(),
+            **kwargs,
+        )
+
+    @property
+    # This is a property because you might want to change the codec model on the fly
+    def sampling_rate(self):
+        return self.audio_encoder.sampling_rate

parler_tts/dac_wrapper/__init__.py

@@ -0,0 +1,2 @@
+from .configuration_dac import DACConfig
+from .modeling_dac import DACModel

parler_tts/dac_wrapper/configuration_dac.py

@@ -0,0 +1,25 @@
+from transformers import PretrainedConfig
+from typing import List
+
+
+class DACConfig(PretrainedConfig):
+    model_type = "dac"
+
+    def __init__(
+        self,
+        num_codebooks: int = 9,
+        model_bitrate: int = 8,  # kbps
+        codebook_size: int = 1024,
+        latent_dim: int = 1024,
+        frame_rate: int = 86,
+        sampling_rate: int = 44100,
+        **kwargs,
+    ):
+        self.codebook_size = codebook_size
+        self.model_bitrate = model_bitrate
+        self.latent_dim = latent_dim
+        self.num_codebooks = num_codebooks
+        self.frame_rate = frame_rate
+        self.sampling_rate = sampling_rate
+
+        super().__init__(**kwargs)

parler_tts/dac_wrapper/modeling_dac.py

@@ -0,0 +1,137 @@
+import torch
+
+from transformers import PreTrainedModel
+from transformers.models.encodec.modeling_encodec import EncodecEncoderOutput, EncodecDecoderOutput
+from .configuration_dac import DACConfig
+
+from dac.model import DAC
+
+
+# model doesn't support batching yet
+
+
+class DACModel(PreTrainedModel):
+    config_class = DACConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.model = DAC(
+            n_codebooks=config.num_codebooks,
+            latent_dim=config.latent_dim,
+            codebook_size=config.codebook_size,
+        )
+
+    def encode(
+        self, input_values, padding_mask=None, bandwidth=None, return_dict=None, n_quantizers=None, sample_rate=None
+    ):
+        """
+        Encodes the input audio waveform into discrete codes.
+
+        Args:
+            input_values (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`):
+                Float values of the input audio waveform.
+            padding_mask (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`):
+                Padding mask used to pad the `input_values`.
+            bandwidth (`float`, *optional*):
+                Not used, kept to have the same inferface as HF encodec.
+            n_quantizers (`int`, *optional*) :
+                Number of quantizers to use, by default None
+                If None, all quantizers are used.
+            sample_rate (`int`, *optional*) :
+                Signal sampling_rate
+
+        Returns:
+            A list of frames containing the discrete encoded codes for the input audio waveform, along with rescaling
+            factors for each chunk when `normalize` is True. Each frames is a tuple `(codebook, scale)`, with
+            `codebook` of shape `[batch_size, num_codebooks, frames]`.
+            Scale is not used here.
+
+        """
+        _, channels, input_length = input_values.shape
+
+        if channels < 1 or channels > 2:
+            raise ValueError(f"Number of audio channels must be 1 or 2, but got {channels}")
+
+        audio_data = self.model.preprocess(input_values, sample_rate)
+
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # TODO: for now, no chunk length
+
+        chunk_length = None  # self.config.chunk_length
+        if chunk_length is None:
+            chunk_length = input_length
+            stride = input_length
+        else:
+            stride = self.config.chunk_stride
+
+        if padding_mask is None:
+            padding_mask = torch.ones_like(input_values).bool()
+
+        encoded_frames = []
+        scales = []
+
+        step = chunk_length - stride
+        if (input_length % stride) - step != 0:
+            raise ValueError(
+                "The input length is not properly padded for batched chunked decoding. Make sure to pad the input correctly."
+            )
+
+        for offset in range(0, input_length - step, stride):
+            mask = padding_mask[..., offset : offset + chunk_length].bool()
+            frame = audio_data[:, :, offset : offset + chunk_length]
+
+            scale = None
+
+            _, encoded_frame, _, _, _ = self.model.encode(frame, n_quantizers=n_quantizers)
+            encoded_frames.append(encoded_frame)
+            scales.append(scale)
+
+        encoded_frames = torch.stack(encoded_frames)
+
+        if not return_dict:
+            return (encoded_frames, scales)
+
+        return EncodecEncoderOutput(encoded_frames, scales)
+
+    def decode(
+        self,
+        audio_codes,
+        audio_scales,
+        padding_mask=None,
+        return_dict=None,
+    ):
+        """
+        Decodes the given frames into an output audio waveform.
+
+        Note that the output might be a bit bigger than the input. In that case, any extra steps at the end can be
+        trimmed.
+
+        Args:
+            audio_codes (`torch.FloatTensor`  of shape `(batch_size, nb_chunks, chunk_length)`, *optional*):
+                Discret code embeddings computed using `model.encode`.
+            audio_scales (`torch.Tensor` of shape `(batch_size, nb_chunks)`, *optional*):
+                Not used, kept to have the same inferface as HF encodec.
+            padding_mask (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`):
+                Padding mask used to pad the `input_values`.
+                Not used yet, kept to have the same inferface as HF encodec.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+
+        """
+        return_dict = return_dict or self.config.return_dict
+
+        # TODO: for now, no chunk length
+
+        if len(audio_codes) != 1:
+            raise ValueError(f"Expected one frame, got {len(audio_codes)}")
+
+        audio_values = self.model.quantizer.from_codes(audio_codes.squeeze(0))[0]
+        audio_values = self.model.decode(audio_values)
+        if not return_dict:
+            return (audio_values,)
+        return EncodecDecoderOutput(audio_values)
+
+    def forward(self, tensor):
+        raise ValueError(f"`DACModel.forward` not implemented yet")

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "chunk_length_s": null,
+  "feature_extractor_type": "EncodecFeatureExtractor",
+  "feature_size": 1,
+  "overlap": null,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 44100
+}

slurm_job.slurm

@@ -0,0 +1,74 @@
+#!/bin/bash
+#SBATCH --job-name=parler-tts
+#SBATCH --nodes=1
+# set 48h for job wall time limit
+#SBATCH --time=48:00:00
+#SBATCH --ntasks-per-node=1          # crucial - only 1 task per dist per node!
+#SBATCH --cpus-per-task=32
+#SBATCH --gres=gpu:8
+#SBATCH --partition=hopper-prod
+#SBATCH --output=/fsx/sanchit/logs/%x-%j.out
+
+set -x -e
+
+# START EDIT
+source ~/.bashrc
+source /fsx/sanchit/miniconda3/bin/activate venv
+
+LOG_PATH="/fsx/sanchit/logs/main_log.txt"
+SAVE_DIR="/fsx/sanchit"
+# END EDIT
+
+echo "START TIME: $(date)"
+
+GPUS_PER_NODE=8
+NNODES=$SLURM_NNODES
+
+# so processes know who to talk to
+MASTER_ADDR=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1`
+
+# From https://i.hsfzxjy.site/2021-03-10-obtain-a-random-unused-tcp-port-with-bash/
+function unused_port() {
+    N=${1:-1}
+    comm -23 \
+        <(seq "1025" "65535" | sort) \
+        <(ss -Htan |
+            awk '{print $4}' |
+            cut -d':' -f2 |
+            sort -u) |
+        shuf |
+        head -n "$N"
+}
+MASTER_PORT=$(unused_port)
+
+# export TORCH_CPP_LOG_LEVEL=INFO
+# export TORCH_DISTRIBUTED_DEBUG=DETAIL
+
+export LAUNCHER="python -u -m accelerate.commands.launch --config_file ./accelerate_config.yaml"
+
+export PROGRAM="./training/run_parler_tts_training.py ./starting_point_0.01.json"
+export CMD="$LAUNCHER $PROGRAM"
+echo $CMD
+
+SRUN_ARGS=" \
+    --wait=60 \
+    --kill-on-bad-exit=1 \
+    "
+
+# py-spy top -s -i -n -- $LAUNCHER --node_rank $SLURM_PROCID --role $SLURMD_NODENAME: $CMD
+clear; srun $SRUN_ARGS --jobid $SLURM_JOB_ID bash -c "$CMD" 2>&1 | tee -a $SAVE_DIR/logs/main_log.txt
+
+
+# srun error handling:
+# --wait=60: wait 60 sec after the first task terminates before terminating all remaining tasks
+# --kill-on-bad-exit=1: terminate a step if any task exits with a non-zero exit code
+
+# SRUN_ARGS=" \
+#     --wait=60 \
+#     --kill-on-bad-exit=1 \
+#     "
+#
+# # py-spy top -s -i -n -- $LAUNCHER --node_rank $SLURM_PROCID --role $SLURMD_NODENAME: $CMD
+# clear; srun $SRUN_ARGS --jobid $SLURM_JOBID bash -c "$CMD" 2>&1 | tee -a $SAVE_DIR/logs/main_log.txt
+
+echo "END TIME: $(date)"

special_tokens_map.json

@@ -0,0 +1,125 @@
+{
+  "additional_special_tokens": [
+    "<extra_id_0>",
+    "<extra_id_1>",
+    "<extra_id_2>",
+    "<extra_id_3>",
+    "<extra_id_4>",
+    "<extra_id_5>",
+    "<extra_id_6>",
+    "<extra_id_7>",
+    "<extra_id_8>",
+    "<extra_id_9>",
+    "<extra_id_10>",
+    "<extra_id_11>",
+    "<extra_id_12>",
+    "<extra_id_13>",
+    "<extra_id_14>",
+    "<extra_id_15>",
+    "<extra_id_16>",
+    "<extra_id_17>",
+    "<extra_id_18>",
+    "<extra_id_19>",
+    "<extra_id_20>",
+    "<extra_id_21>",
+    "<extra_id_22>",
+    "<extra_id_23>",
+    "<extra_id_24>",
+    "<extra_id_25>",
+    "<extra_id_26>",
+    "<extra_id_27>",
+    "<extra_id_28>",
+    "<extra_id_29>",
+    "<extra_id_30>",
+    "<extra_id_31>",
+    "<extra_id_32>",
+    "<extra_id_33>",
+    "<extra_id_34>",
+    "<extra_id_35>",
+    "<extra_id_36>",
+    "<extra_id_37>",
+    "<extra_id_38>",
+    "<extra_id_39>",
+    "<extra_id_40>",
+    "<extra_id_41>",
+    "<extra_id_42>",
+    "<extra_id_43>",
+    "<extra_id_44>",
+    "<extra_id_45>",
+    "<extra_id_46>",
+    "<extra_id_47>",
+    "<extra_id_48>",
+    "<extra_id_49>",
+    "<extra_id_50>",
+    "<extra_id_51>",
+    "<extra_id_52>",
+    "<extra_id_53>",
+    "<extra_id_54>",
+    "<extra_id_55>",
+    "<extra_id_56>",
+    "<extra_id_57>",
+    "<extra_id_58>",
+    "<extra_id_59>",
+    "<extra_id_60>",
+    "<extra_id_61>",
+    "<extra_id_62>",
+    "<extra_id_63>",
+    "<extra_id_64>",
+    "<extra_id_65>",
+    "<extra_id_66>",
+    "<extra_id_67>",
+    "<extra_id_68>",
+    "<extra_id_69>",
+    "<extra_id_70>",
+    "<extra_id_71>",
+    "<extra_id_72>",
+    "<extra_id_73>",
+    "<extra_id_74>",
+    "<extra_id_75>",
+    "<extra_id_76>",
+    "<extra_id_77>",
+    "<extra_id_78>",
+    "<extra_id_79>",
+    "<extra_id_80>",
+    "<extra_id_81>",
+    "<extra_id_82>",
+    "<extra_id_83>",
+    "<extra_id_84>",
+    "<extra_id_85>",
+    "<extra_id_86>",
+    "<extra_id_87>",
+    "<extra_id_88>",
+    "<extra_id_89>",
+    "<extra_id_90>",
+    "<extra_id_91>",
+    "<extra_id_92>",
+    "<extra_id_93>",
+    "<extra_id_94>",
+    "<extra_id_95>",
+    "<extra_id_96>",
+    "<extra_id_97>",
+    "<extra_id_98>",
+    "<extra_id_99>"
+  ],
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

spiece.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d60acb128cf7b7f2536e8f38a5b18a05535c9e14c7a355904270e15b0945ea86
+size 791656

starting_point_0.01.json

@@ -0,0 +1,79 @@
+{
+    "model_name_or_path": "parler-tts/parler-tts-untrained-600M-cross-attention",
+    "save_to_disk":  "/fsx/sanchit/10k_hours_processed_punctuated",
+    "temporary_save_to_disk": "/scratch/tmp_dataset_audio/",
+    "push_to_hub": true,
+
+
+    "feature_extractor_name":"ylacombe/dac_44khZ_8kbps",
+    "description_tokenizer_name":"google/flan-t5-base",
+    "prompt_tokenizer_name":"google/flan-t5-base",
+
+    "report_to": ["wandb"],
+    "wandb_run_name": "parler-tts-600M-cross-attention-right-pad",
+    "overwrite_output_dir": false,
+    "output_dir": "./",
+    "save_total_limit": 2,
+
+    "train_dataset_name": "blabble-io/libritts_r+blabble-io/libritts_r+blabble-io/libritts_r+parler-tts/mls_eng_10k",
+    "train_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated",
+    "train_dataset_config_name": "clean+clean+other+default",
+    "train_split_name": "train.clean.360+train.clean.100+train.other.500+train",
+
+    "eval_dataset_name": "blabble-io/libritts_r+parler-tts/mls_eng_10k",
+    "eval_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated",
+    "eval_dataset_config_name": "other+default",
+    "eval_split_name": "test.other+test",
+
+    "target_audio_column_name": "audio", 
+    "description_column_name": "text_description",
+    "prompt_column_name": "text",
+
+    "max_eval_samples": 96,
+    
+    "max_duration_in_seconds": 30,
+    "min_duration_in_seconds": 2.0,
+    "max_text_length": 400,
+
+    "group_by_length": true,
+
+    "add_audio_samples_to_wandb": true,
+    "id_column_name": "id",
+
+    "preprocessing_num_workers": 8,
+
+    "do_train": true,
+    "num_train_epochs": 40,
+    "gradient_accumulation_steps": 8,
+    "gradient_checkpointing": false,
+    "per_device_train_batch_size": 3,
+    "learning_rate": 0.00095,
+    "adam_beta1": 0.9,
+    "adam_beta2": 0.99,
+    "weight_decay": 0.01,
+
+    "lr_scheduler_type": "constant_with_warmup",
+    "warmup_steps":  20000,
+
+
+    "logging_steps": 1000,
+    "freeze_text_encoder": true,
+
+
+    "do_eval": true, 
+    "predict_with_generate": true,
+    "include_inputs_for_metrics": true,
+    "evaluation_strategy": "steps",
+    "eval_steps": 10000,
+    "save_steps": 10000,
+
+    "per_device_eval_batch_size": 12,
+
+    "audio_encoder_per_device_batch_size":20,
+    "dtype": "bfloat16",
+    "seed": 456,
+    "ddp_timeout": 7200,
+
+    "dataloader_num_workers":8
+}
+

tokenizer_config.json

@@ -0,0 +1,938 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32000": {
+      "content": "<extra_id_99>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32001": {
+      "content": "<extra_id_98>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32002": {
+      "content": "<extra_id_97>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32003": {
+      "content": "<extra_id_96>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32004": {
+      "content": "<extra_id_95>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32005": {
+      "content": "<extra_id_94>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32006": {
+      "content": "<extra_id_93>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32007": {
+      "content": "<extra_id_92>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32008": {
+      "content": "<extra_id_91>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32009": {
+      "content": "<extra_id_90>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32010": {
+      "content": "<extra_id_89>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32011": {
+      "content": "<extra_id_88>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32012": {
+      "content": "<extra_id_87>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32013": {
+      "content": "<extra_id_86>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32014": {
+      "content": "<extra_id_85>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32015": {
+      "content": "<extra_id_84>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32016": {
+      "content": "<extra_id_83>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32017": {
+      "content": "<extra_id_82>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32018": {
+      "content": "<extra_id_81>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32019": {
+      "content": "<extra_id_80>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32020": {
+      "content": "<extra_id_79>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32021": {
+      "content": "<extra_id_78>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32022": {
+      "content": "<extra_id_77>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32023": {
+      "content": "<extra_id_76>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32024": {
+      "content": "<extra_id_75>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32025": {
+      "content": "<extra_id_74>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32026": {
+      "content": "<extra_id_73>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32027": {
+      "content": "<extra_id_72>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32028": {
+      "content": "<extra_id_71>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32029": {
+      "content": "<extra_id_70>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32030": {
+      "content": "<extra_id_69>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32031": {
+      "content": "<extra_id_68>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32032": {
+      "content": "<extra_id_67>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32033": {
+      "content": "<extra_id_66>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32034": {
+      "content": "<extra_id_65>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32035": {
+      "content": "<extra_id_64>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32036": {
+      "content": "<extra_id_63>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32037": {
+      "content": "<extra_id_62>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32038": {
+      "content": "<extra_id_61>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32039": {
+      "content": "<extra_id_60>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32040": {
+      "content": "<extra_id_59>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32041": {
+      "content": "<extra_id_58>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32042": {
+      "content": "<extra_id_57>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32043": {
+      "content": "<extra_id_56>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32044": {
+      "content": "<extra_id_55>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32045": {
+      "content": "<extra_id_54>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32046": {
+      "content": "<extra_id_53>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32047": {
+      "content": "<extra_id_52>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32048": {
+      "content": "<extra_id_51>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32049": {
+      "content": "<extra_id_50>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32050": {
+      "content": "<extra_id_49>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32051": {
+      "content": "<extra_id_48>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32052": {
+      "content": "<extra_id_47>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32053": {
+      "content": "<extra_id_46>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32054": {
+      "content": "<extra_id_45>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32055": {
+      "content": "<extra_id_44>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32056": {
+      "content": "<extra_id_43>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32057": {
+      "content": "<extra_id_42>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32058": {
+      "content": "<extra_id_41>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32059": {
+      "content": "<extra_id_40>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32060": {
+      "content": "<extra_id_39>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32061": {
+      "content": "<extra_id_38>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32062": {
+      "content": "<extra_id_37>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32063": {
+      "content": "<extra_id_36>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32064": {
+      "content": "<extra_id_35>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32065": {
+      "content": "<extra_id_34>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32066": {
+      "content": "<extra_id_33>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32067": {
+      "content": "<extra_id_32>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32068": {
+      "content": "<extra_id_31>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32069": {
+      "content": "<extra_id_30>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32070": {
+      "content": "<extra_id_29>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32071": {
+      "content": "<extra_id_28>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32072": {
+      "content": "<extra_id_27>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32073": {
+      "content": "<extra_id_26>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32074": {
+      "content": "<extra_id_25>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32075": {
+      "content": "<extra_id_24>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32076": {
+      "content": "<extra_id_23>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32077": {
+      "content": "<extra_id_22>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32078": {
+      "content": "<extra_id_21>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32079": {
+      "content": "<extra_id_20>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32080": {
+      "content": "<extra_id_19>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32081": {
+      "content": "<extra_id_18>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32082": {
+      "content": "<extra_id_17>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32083": {
+      "content": "<extra_id_16>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32084": {
+      "content": "<extra_id_15>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32085": {
+      "content": "<extra_id_14>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32086": {
+      "content": "<extra_id_13>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32087": {
+      "content": "<extra_id_12>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32088": {
+      "content": "<extra_id_11>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32089": {
+      "content": "<extra_id_10>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32090": {
+      "content": "<extra_id_9>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32091": {
+      "content": "<extra_id_8>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32092": {
+      "content": "<extra_id_7>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32093": {
+      "content": "<extra_id_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32094": {
+      "content": "<extra_id_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32095": {
+      "content": "<extra_id_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32096": {
+      "content": "<extra_id_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32097": {
+      "content": "<extra_id_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32098": {
+      "content": "<extra_id_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32099": {
+      "content": "<extra_id_0>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<extra_id_0>",
+    "<extra_id_1>",
+    "<extra_id_2>",
+    "<extra_id_3>",
+    "<extra_id_4>",
+    "<extra_id_5>",
+    "<extra_id_6>",
+    "<extra_id_7>",
+    "<extra_id_8>",
+    "<extra_id_9>",
+    "<extra_id_10>",
+    "<extra_id_11>",
+    "<extra_id_12>",
+    "<extra_id_13>",
+    "<extra_id_14>",
+    "<extra_id_15>",
+    "<extra_id_16>",
+    "<extra_id_17>",
+    "<extra_id_18>",
+    "<extra_id_19>",
+    "<extra_id_20>",
+    "<extra_id_21>",
+    "<extra_id_22>",
+    "<extra_id_23>",
+    "<extra_id_24>",
+    "<extra_id_25>",
+    "<extra_id_26>",
+    "<extra_id_27>",
+    "<extra_id_28>",
+    "<extra_id_29>",
+    "<extra_id_30>",
+    "<extra_id_31>",
+    "<extra_id_32>",
+    "<extra_id_33>",
+    "<extra_id_34>",
+    "<extra_id_35>",
+    "<extra_id_36>",
+    "<extra_id_37>",
+    "<extra_id_38>",
+    "<extra_id_39>",
+    "<extra_id_40>",
+    "<extra_id_41>",
+    "<extra_id_42>",
+    "<extra_id_43>",
+    "<extra_id_44>",
+    "<extra_id_45>",
+    "<extra_id_46>",
+    "<extra_id_47>",
+    "<extra_id_48>",
+    "<extra_id_49>",
+    "<extra_id_50>",
+    "<extra_id_51>",
+    "<extra_id_52>",
+    "<extra_id_53>",
+    "<extra_id_54>",
+    "<extra_id_55>",
+    "<extra_id_56>",
+    "<extra_id_57>",
+    "<extra_id_58>",
+    "<extra_id_59>",
+    "<extra_id_60>",
+    "<extra_id_61>",
+    "<extra_id_62>",
+    "<extra_id_63>",
+    "<extra_id_64>",
+    "<extra_id_65>",
+    "<extra_id_66>",
+    "<extra_id_67>",
+    "<extra_id_68>",
+    "<extra_id_69>",
+    "<extra_id_70>",
+    "<extra_id_71>",
+    "<extra_id_72>",
+    "<extra_id_73>",
+    "<extra_id_74>",
+    "<extra_id_75>",
+    "<extra_id_76>",
+    "<extra_id_77>",
+    "<extra_id_78>",
+    "<extra_id_79>",
+    "<extra_id_80>",
+    "<extra_id_81>",
+    "<extra_id_82>",
+    "<extra_id_83>",
+    "<extra_id_84>",
+    "<extra_id_85>",
+    "<extra_id_86>",
+    "<extra_id_87>",
+    "<extra_id_88>",
+    "<extra_id_89>",
+    "<extra_id_90>",
+    "<extra_id_91>",
+    "<extra_id_92>",
+    "<extra_id_93>",
+    "<extra_id_94>",
+    "<extra_id_95>",
+    "<extra_id_96>",
+    "<extra_id_97>",
+    "<extra_id_98>",
+    "<extra_id_99>"
+  ],
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "</s>",
+  "extra_ids": 100,
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "T5Tokenizer",
+  "unk_token": "<unk>"
+}

training/README.md

@@ -0,0 +1,211 @@
+# Training Parler-TTS
+
+<a target="_blank" href="https://colab.research.google.com/github/ylacombe/scripts_and_notebooks/blob/main/Finetuning_Parler_TTS_on_a_single_speaker_dataset.ipynb"> 
+  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> 
+</a>
+
+**TL;DR:** After having followed the [installation steps](#requirements), you can reproduce the [Parler-TTS Mini v0.1](https://huggingface.co/parler-tts/parler_tts_mini_v0.1) training recipe with the following command line:
+
+```sh
+accelerate launch ./training/run_parler_tts_training.py ./helpers/training_configs/starting_point_0.01.json
+```
+
+-------------
+
+This sub-folder contains all the information to train or fine-tune your own Parler-TTS model. It consists of:
+- [1. An introduction to the Parler-TTS architecture](#a-architecture)
+- [2. First steps to get started](#b-getting-started)
+- [3. Training guide](#c-training)
+
+> [!IMPORTANT]
+> You can also follow [this fine-tuning guide](https://colab.research.google.com/github/ylacombe/scripts_and_notebooks/blob/main/Finetuning_Parler_TTS_on_a_single_speaker_dataset.ipynb) on a mono-speaker dataset example.
+
+## 1. Architecture
+
+At the moment, Parler-TTS architecture is a carbon copy of the [MusicGen architecture](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/musicgen#model-structure) and can be decomposed into three distinct stages:
+1. Text encoder: maps the text descriptions to a sequence of hidden-state representations. Parler-TTS uses a frozen text encoder initialised entirely from Flan-T5
+2. Parler-TTS decoder: a language model (LM) that auto-regressively generates audio tokens (or codes) conditional on the encoder hidden-state representations
+3. Audio codec: used to recover the audio waveform from the audio tokens predicted by the decoder. We use the [DAC model](https://github.com/descriptinc/descript-audio-codec) from Descript, although other codec models, such as [EnCodec](https://huggingface.co/facebook/encodec_48khz), can also be used
+
+Parler-TTS however introduces some small tweaks:
+- The text **description** is passed through the text encoder and used in the cross-attention layers of the decoder.
+- The text **prompt** is simply passed through an embedding layer and concatenated to the decoder input hidden states.
+- The audio encoder used is [**DAC**](https://descript.notion.site/Descript-Audio-Codec-11389fce0ce2419891d6591a68f814d5) instead of [Encodec](https://github.com/facebookresearch/encodec), as it exhibits better quality.
+
+
+## 2. Getting started
+
+To get started, you need to follow a few steps:
+1. Install the requirements.
+2. Find or initialize the model you'll train on. 
+3. Find and/or annotate the dataset you'll train your model on.
+
+### Requirements
+
+The Parler-TTS code is written in [PyTorch](https://pytorch.org) and [Accelerate](https://huggingface.co/docs/accelerate/index). It uses some additional requirements, like [wandb](https://wandb.ai/), especially for logging and evaluation.
+
+To install the package for training, you need to clone the repository from source...
+
+```bash
+git clone https://github.com/huggingface/parler-tts.git
+cd parler-tts
+```
+
+... And then install the requirements:
+
+```bash
+pip install -e .[train]
+```
+
+Optionally, you can create a wandb account and login to it by following [this guide](https://docs.wandb.ai/quickstart). [`wandb`](https://docs.wandb.ai/) allows for better tracking of the experiments metrics and losses.
+
+You also have the option to configure Accelerate by running the following command. Note that you should set the number of GPUs you wish to use for training, and also the data type (dtype) to your preferred dtype for training/inference (e.g. `bfloat16` on A100 GPUs, `float16` on V100 GPUs, etc.):
+
+```bash
+accelerate config
+```
+
+Lastly, you can link you Hugging Face account so that you can push model repositories on the Hub. This will allow you to save your trained models on the Hub so that you can share them with the community. Run the command:
+
+```bash
+git config --global credential.helper store
+huggingface-cli login
+```
+And then enter an authentication token from https://huggingface.co/settings/tokens. Create a new token if you do not have one already. You should make sure that this token has "write" privileges.
+
+### Initialize a model from scratch or use a pre-trained one.
+
+Depending on your compute resources and your dataset, you need to choose between fine-tuning a pre-trained model and training a new model from scratch.
+
+In that sense, we released a 600M checkpoint trained on 10.5K hours of annotated data under the repository id: [`parler-tts/parler_tts_mini_v0.1`](https://huggingface.co/parler-tts/parler_tts_mini_v0.1), that you can fine-tune for your own use-case.
+
+You can also train you own model from scratch. You can find [here](/helpers/model_init_scripts/) examples on how to initialize a model from scratch. For example, you can initialize a dummy model with:
+
+```sh
+python helpers/model_init_scripts/init_dummy_model.py ./parler-tts-untrained-dummy --text_model "google-t5/t5-small" --audio_model "parler-tts/dac_44khZ_8kbps"
+```
+
+In the rest of this guide, and to reproduce the Parler-TTS Mini v0.1 training recipe, we'll use a 600M parameters model that we'll initialize with:
+
+```sh
+python helpers/model_init_scripts/init_model_600M.py ./parler-tts-untrained-600M --text_model "google/flan-t5-base" --audio_model "parler-tts/dac_44khZ_8kbps"
+```
+
+
+### Create or find datasets
+
+To train your own Parler-TTS, you need datasets with 3 main features:
+- speech data
+- text transcription of the speech data
+- conditionning text description - that you can create using [Data-Speech](https://github.com/huggingface/dataspeech), a library that allows you to annotate the speaker and utterance characteristics with natural language description.
+
+Note that we made the choice to use description of the main speech characteristics (speaker pitch, speaking rate, level of noise, etc.) but that you are free to use any handmade or generated text description that makes sense.
+
+To train Parler-TTS Mini v0.1, we used:
+* The full [LibriTTS-R dataset](https://huggingface.co/datasets/blabble-io/libritts_r), a 1K hours high-quality speech dataset.
+* A [10K hours subset](https://huggingface.co/datasets/parler-tts/mls_eng_10k) of [Multilingual LibriSpeech](https://huggingface.co/datasets/facebook/multilingual_librispeech).
+
+Both datasets have been annotated using the [Data-Speech](https://github.com/huggingface/dataspeech) recipe, respectively [here](https://huggingface.co/datasets/parler-tts/libritts_r_tags_tagged_10k_generated) and [here](https://huggingface.co/datasets/parler-tts/mls-eng-10k-tags_tagged_10k_generated).
+
+
+## 3. Training
+
+The script [`run_parler_tts_training.py`](/training/run_parler_tts_training.py) is an end-to-end script that:
+1. load dataset(s) and merge them to the annotation dataset(s) if necessary
+2. pre-compute audio tokens
+3. train Parler-TTS
+
+To train Parler-TTS Mini v0.1, we roughly used:
+
+```sh
+accelerate launch ./training/run_parler_tts_training.py \
+    --model_name_or_path "./parler-tts-untrained-600M/parler-tts-untrained-600M/" \
+    --feature_extractor_name "parler-tts/dac_44khZ_8kbps" \
+    --description_tokenizer_name "google/flan-t5-base" \
+    --prompt_tokenizer_name "google/flan-t5-base" \
+    --report_to "wandb" \
+    --overwrite_output_dir true \
+    --train_dataset_name "blabble-io/libritts_r+blabble-io/libritts_r+blabble-io/libritts_r+parler-tts/mls_eng_10k" \
+    --train_metadata_dataset_name "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated" \
+    --train_dataset_config_name "clean+clean+other+default" \
+    --train_split_name "train.clean.360+train.clean.100+train.other.500+train" \
+    --eval_dataset_name "blabble-io/libritts_r+parler-tts/mls_eng_10k" \
+    --eval_metadata_dataset_name "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated" \
+    --eval_dataset_config_name "other+default" \
+    --eval_split_name "test.other+test" \
+    --target_audio_column_name "audio" \
+    --description_column_name "text_description" \
+    --prompt_column_name "text" \
+    --max_duration_in_seconds 30 \
+    --min_duration_in_seconds 2.0 \
+    --max_text_length 400 \
+    --add_audio_samples_to_wandb true \
+    --id_column_name "id" \
+    --preprocessing_num_workers 8 \
+    --do_train true \
+    --num_train_epochs 40 \
+    --gradient_accumulation_steps 8 \
+    --gradient_checkpointing false \
+    --per_device_train_batch_size 3 \
+    --learning_rate 0.00095 \
+    --adam_beta1 0.9 \
+    --adam_beta2 0.99 \
+    --weight_decay 0.01 \
+    --lr_scheduler_type "constant_with_warmup" \
+    --warmup_steps 20000 \
+    --logging_steps 1000 \
+    --freeze_text_encoder true \
+    --do_eval true \
+    --predict_with_generate true \
+    --include_inputs_for_metrics true \
+    --evaluation_strategy steps \
+    --eval_steps 10000 \
+    --save_steps 10000 \
+    --per_device_eval_batch_size 12 \
+    --audio_encoder_per_device_batch_size 20 \
+    --dtype "bfloat16" \
+    --seed 456 \
+    --output_dir "./output_dir_training/" \
+    --temporary_save_to_disk "./audio_code_tmp/" \
+    --save_to_disk "./tmp_dataset_audio/" \
+    --max_eval_samples 96 \
+    --dataloader_num_workers 8 \
+    --group_by_length true
+```
+
+In particular, note how multiple training datasets, metadataset, configurations and splits can be loaded by separating the dataset arguments by + symbols:
+```sh
+    "train_dataset_name": "blabble-io/libritts_r+blabble-io/libritts_r+blabble-io/libritts_r+parler-tts/mls_eng_10k",
+    "train_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated",
+    "train_dataset_config_name": "clean+clean+other+default",
+    "train_split_name": "train.clean.360+train.clean.100+train.other.500+train",
+```
+
+
+Additionally, you can also write a JSON config file. Here, [starting_point_0.01.json](helpers/training_configs/starting_point_0.01.json) contains the exact same hyper-parameters than above and can be launched like that:
+```sh
+accelerate launch ./training/run_parler_tts_training.py ./helpers/training_configs/starting_point_0.01.json
+```
+
+Training logs will be reported to wandb, provided that you passed `--report_to "wandb"` to the arguments. An example of what a training log from the above training looks like can be found [here](https://wandb.ai/ylacombe/parler-tts-300M-punctuated/runs/q6h7hspc?nw=nwuserylacombe).
+
+> [!TIP]
+> Starting training a new model from scratch can easily be overwhelming, so here's what training looked like for v0.1: [logs](https://api.wandb.ai/links/ylacombe/ea449l81)
+
+Scaling to multiple GPUs using [distributed data parallelism (DDP)](https://pytorch.org/tutorials/beginner/ddp_series_theory.html) is trivial: simply run `accelerate config` and select the multi-GPU option, specifying the IDs of the GPUs you wish to use. The above script can then be run using DDP with no code changes. In our case, we used a node of 8 H100 80GB to train Parler-TTS v0.1 for around 4 days.
+
+
+There are a few other noteworthy arguments:
+1. `train_metadata_dataset_name` and `eval_metadata_dataset_name` specify, if necessary, the names of the dataset(s) that contain(s) the conditionning text descriptions. For example, this [dataset resulting from the Data-Speech annotation process](https://huggingface.co/datasets/parler-tts/libritts_r_tags_tagged_10k_generated) is saved without the audio column, as it's costly to write and push audio data, so it needs to be concatenated back to the original LibriTTS-R dataset.
+2. As noted above, the script pre-computes audio tokens as computing audio codes is costly and only needs to be done once, since we're freezing the audio encoder. `audio_encoder_per_device_batch_size` is used to precise the per devie batch size for this pre-processing step.
+3. Additionnally, when scaling up the training data and iterating on the hyper-parameters or the model architecture, we might want to avoid recomputing the audio tokens at each training run. That's why we introduced two additional parameters, `save_to_disk` and `temporary_save_to_disk` that serves as temporary buffers to save intermediary datasets. Note that processed data is made of text and audio tokens which are much more memory efficient, so the additional required space is negligible.
+4. `predict_with_generate` and `add_audio_samples_to_wandb` are required to store generated audios and to compute WER and CLAP similarity.
+5. `freeze_text_encoder`: which allows to freeze the text encoder, to save compute resources.
+
+And finally, two additional comments:
+1. `lr_scheduler_stype`: defines the learning rate schedule, one of `constant_with_warmup` or `cosine`. When experimenting with a training set-up or training for very few epochs, using `constant_with_warmup` is typically beneficial, since the learning rate remains high over the short training run. When performing longer training runs, using a `cosine` schedule shoud give better results.
+2. `dtype`: data type (dtype) in which the model computation should be performed. Note that this only controls the dtype of the computations (forward and backward pass), and not the dtype of the parameters or optimiser states.
+
+> [!TIP]
+> Fine-tuning is as easy as modifying `model_name_or_path` to a pre-trained model.
+> For example: `--model_name_or_path parler-tts/parler_tts_mini_v0.1`.

training/arguments.py

@@ -0,0 +1,313 @@
+from dataclasses import dataclass, field
+from typing import Optional
+
+from transformers import Seq2SeqTrainingArguments
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    feature_extractor_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained feature extractor name or path if not the same as model_name"}
+    )
+    description_tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained description tokenizer name or path if not the same as model_name"}
+    )
+    prompt_tokenizer_name: Optional[str] = field(
+        default=None,
+        metadata={"help": "Pretrained prompt tokenizer name or path if not the same as description_tokenizer_name"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    model_revision: str = field(
+        default="main",
+        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
+    )
+    pad_token_id: int = field(
+        default=None,
+        metadata={"help": "If specified, change the model pad token id."},
+    )
+    decoder_start_token_id: int = field(
+        default=None,
+        metadata={"help": "If specified, change the model decoder start token id."},
+    )
+    freeze_text_encoder: bool = field(
+        default=False,
+        metadata={"help": "Whether to freeze the text encoder."},
+    )
+    do_sample: bool = field(
+        default=True,
+        metadata={"help": "Whether to do sampling or greedy decoding."},
+    )
+    temperature: float = field(
+        default=1.0,
+        metadata={"help": "Temperature if sampling."},
+    )
+    max_length: int = field(
+        default=2580,
+        metadata={"help": "Generation max length."},
+    )
+    bandwidth: float = field(
+        default=6,
+        metadata={"help": "Audio encoder bandwidth."},
+    )
+    asr_model_name_or_path: str = field(
+        default="distil-whisper/distil-large-v2",
+        metadata={
+            "help": "Used to compute WER during evaluation. Path to pretrained model or model identifier from huggingface.co/models"
+        },
+    )
+    clap_model_name_or_path: str = field(
+        default="laion/larger_clap_music_and_speech",
+        metadata={
+            "help": "Used to compute audio similarity during evaluation. Path to pretrained model or model identifier from huggingface.co/models"
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    train_dataset_name: str = field(
+        default=None,
+        metadata={
+            "help": "The name of the training dataset to use (via the datasets library). Load and combine "
+            "multiple datasets by separating dataset ids by a '+' symbol. For example, to load and combine "
+            " librispeech and common voice, set `train_dataset_name='librispeech_asr+common_voice'`."
+        },
+    )
+    train_dataset_config_name: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The configuration name of the training dataset to use (via the datasets library). Load and combine "
+            "multiple datasets by separating dataset configs by a '+' symbol."
+        },
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": ("The name of the training data set split to use (via the datasets library). Defaults to 'train'")
+        },
+    )
+    train_dataset_samples: str = field(
+        default=None,
+        metadata={
+            "help": "Number of samples in the training data. Load and combine "
+            "multiple datasets by separating dataset samples by a '+' symbol."
+        },
+    )
+    train_metadata_dataset_name: str = field(
+        default=None,
+        metadata={
+            "help": "The name of the metadata training dataset to use (via the datasets library). Load and combine "
+            "multiple datasets by separating dataset ids by a '+' symbol. For example, to load and combine "
+            " librispeech and common voice, set `train_dataset_name='librispeech_asr+common_voice'`."
+        },
+    )
+    eval_dataset_name: str = field(
+        default=None,
+        metadata={
+            "help": "The name of the evaluation dataset to use (via the datasets library). Defaults to the training dataset name if unspecified."
+        },
+    )
+    eval_dataset_config_name: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The configuration name of the evaluation dataset to use (via the datasets library). Defaults to the training dataset config name if unspecified"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'test'"
+        },
+    )
+    eval_metadata_dataset_name: str = field(
+        default=None,
+        metadata={
+            "help": "The name of the metadata training dataset to use (via the datasets library). Load and combine "
+            "multiple datasets by separating dataset ids by a '+' symbol. For example, to load and combine "
+            " librispeech and common voice, set `train_dataset_name='librispeech_asr+common_voice'`."
+        },
+    )
+    target_audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the target audio data. Defaults to 'audio'"},
+    )
+    description_column_name: str = field(
+        default=None,
+        metadata={"help": "The name of the dataset column containing the description text data. Defaults to 'None'."},
+    )
+    prompt_column_name: str = field(
+        default=None,
+        metadata={"help": "The name of the dataset column containing the prompt text data. Defaults to 'None'."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "For debugging purposes or quicker training, truncate the number of training examples to this "
+                "value if set."
+            )
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "For debugging purposes or quicker training, truncate the number of validation examples to this "
+                "value if set."
+            )
+        },
+    )
+    max_duration_in_seconds: float = field(
+        default=35.0,
+        metadata={
+            "help": (
+                "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`."
+                "Also, used to set maximum audio length if `pad_to_max_length=True`."
+            )
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    max_text_length: int = field(
+        default=500, metadata={"help": "If set, max description lengths in number of characters."}
+    )
+    max_prompt_token_length: int = field(
+        default=None,
+        metadata={
+            "help": (
+                "If set, filter samples with prompts that are longer than `max_prompt_token_length` tokens."
+                "Also, used to set maximum prompt token length if `pad_to_max_length=True`."
+            )
+        },
+    )
+    max_description_token_length: int = field(
+        default=None,
+        metadata={
+            "help": (
+                "If set, filter samples with descriptions that are longer than `max_description_token_length` tokens."
+                "Also, used to set maximum description token length if `pad_to_max_length=True`."
+            )
+        },
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "If `True`, pad audio, prompt and description to a maximum length set with respectively "
+                "`max_duration_in_seconds`, `max_prompt_token_length`, `max_description_token_length`."
+            )
+        },
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether to only do data preprocessing and skip training. This is especially useful when data"
+                " preprocessing errors out in distributed training due to timeout. In this case, one should run the"
+                " preprocessing in a non-distributed setup with `preprocessing_only=True` so that the cached datasets"
+                " can consequently be loaded in distributed training."
+                " In this training script, `save_to_disk` must be set to the path in which the dataset should be saved. "
+            )
+        },
+    )
+    token: str = field(
+        default=None,
+        metadata={
+            "help": (
+                "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
+                "generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
+            )
+        },
+    )
+    use_auth_token: bool = field(
+        default=None,
+        metadata={
+            "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead."
+        },
+    )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option "
+                "should only be set to `True` for repositories you trust and in which you have read the code, as it will "
+                "execute code present on the Hub on your local machine."
+            )
+        },
+    )
+    add_audio_samples_to_wandb: bool = field(
+        default=False,
+        metadata={"help": "If set and if `wandb` in args.report_to, will add generated audio samples to wandb logs."},
+    )
+    id_column_name: str = field(default=None, metadata={"help": "id column name."})
+    wandb_project: str = field(
+        default="parler-speech",
+        metadata={"help": "The name of the wandb project."},
+    )
+    wandb_run_name: str = field(
+        default=None,
+        metadata={
+            "help": "If specified, the name of the run. If not specified, wandb will give a random name to this run."
+        },
+    )
+    save_to_disk: str = field(
+        default=None,
+        metadata={
+            "help": "If set, will save the dataset to this path if this is an empyt folder. If not empty, will load the datasets from it."
+        },
+    )
+    temporary_save_to_disk: str = field(default=None, metadata={"help": "Temporarily save audio labels here."})
+    pad_to_multiple_of: Optional[int] = field(
+        default=2,
+        metadata={"help": ("Pad to multiple of for tokenizers.")},
+    )
+
+
+@dataclass
+class ParlerTTSTrainingArguments(Seq2SeqTrainingArguments):
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": (
+                "The data type (dtype) in which to run training. One of `float32` (full-precision), "
+                "`float16` or `bfloat16` (both half-precision)."
+            )
+        },
+    )
+    audio_encoder_per_device_batch_size: int = field(
+        default=8,
+        metadata={"help": ("Specify the batch size of the audio encoding pre-processing steps.")},
+    )

training/data.py

@@ -0,0 +1,310 @@
+import logging
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Union, Set
+
+import torch
+import numpy as np
+import datasets
+from datasets import load_dataset, Dataset, IterableDataset, interleave_datasets, concatenate_datasets
+from transformers import AutoFeatureExtractor, AutoTokenizer
+from tqdm import tqdm
+
+from accelerate import Accelerator
+
+
+@dataclass
+class DataCollatorEncodecWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received to the longest sequence in the batch or
+    to `max_length` if `max_length` is set and `padding=max_length`.
+    """
+
+    feature_extractor: AutoFeatureExtractor
+    audio_column_name: str
+    feature_extractor_input_name: Optional[str] = "input_values"
+    max_length: Optional[int] = None
+    padding: Optional[str] = "longest"
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lengths and need
+        # different padding methods
+        audios = [feature[self.audio_column_name]["array"] for feature in features]
+        len_audio = [len(audio) for audio in audios]
+
+        # since resampling has already been performed in the 'load_multiple_datasets' function,
+        # a fixed sampling_rate(44100hz) is passed to the feature_extractor.
+        sampling_rate = self.feature_extractor.sampling_rate
+        batch = self.feature_extractor(
+            audios, sampling_rate=sampling_rate, return_tensors="pt", padding=self.padding, max_length=self.max_length
+        )
+        batch["len_audio"] = torch.tensor(len_audio).unsqueeze(1)
+        return batch
+
+
+@dataclass
+class DataCollatorParlerTTSWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        prompt_tokenizer (:class:`~transformers.AutoTokenizer`)
+            The prompt_tokenizer used for proccessing the data.
+        description_tokenizer (:class:`~transformers.AutoTokenizer`)
+            The description_tokenizer used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    prompt_tokenizer: AutoTokenizer
+    description_tokenizer: AutoTokenizer
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    prompt_max_length: Optional[int] = None
+    description_max_length: Optional[int] = None
+    audio_max_length: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lengths and need
+        # different padding methods
+
+        labels = [torch.tensor(feature["labels"]).transpose(0, 1) for feature in features]
+        # (bsz, seq_len, num_codebooks)
+        labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=-100)
+        if self.audio_max_length is not None and self.padding == "max_length":
+            labels = torch.nn.functional.pad(labels, pad=(0, 0, 0, max(self.audio_max_length - labels.shape[1], 0)))
+
+        input_ids = [{"input_ids": feature["input_ids"]} for feature in features]
+
+        input_ids = self.description_tokenizer.pad(
+            input_ids,
+            return_tensors="pt",
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            max_length=self.description_max_length,
+        )
+
+        batch = {"labels": labels, **input_ids}
+
+        if self.audio_max_length is not None and self.padding == "max_length":
+            # if we do torch.compile, we need to also specify the attention_mask
+            decoder_attention_mask = torch.ones(labels.shape[:2], dtype=input_ids["attention_mask"].dtype)
+            batch["decoder_attention_mask"] = decoder_attention_mask
+
+        prompt_input_ids = [{"input_ids": feature["prompt_input_ids"]} for feature in features]
+        prompt_input_ids = self.prompt_tokenizer.pad(
+            prompt_input_ids,
+            return_tensors="pt",
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            max_length=self.prompt_max_length,
+        )
+
+        batch["prompt_input_ids"] = prompt_input_ids["input_ids"]
+        if "attention_mask" in prompt_input_ids:
+            batch["prompt_attention_mask"] = prompt_input_ids["attention_mask"]
+
+        return batch
+
+
+def convert_dataset_str_to_list(
+    dataset_names,
+    dataset_config_names,
+    metadata_dataset_names=None,
+    splits=None,
+    dataset_samples=None,
+    default_split="train",
+):
+    if isinstance(dataset_names, str):
+        dataset_names = dataset_names.split("+")
+        dataset_config_names = dataset_config_names.split("+")
+        splits = splits.split("+") if splits is not None else None
+        dataset_samples = dataset_samples.split("+") if dataset_samples is not None else None
+        metadata_dataset_names = metadata_dataset_names.split("+") if metadata_dataset_names is not None else None
+
+    # basic checks to ensure we've got the right number of datasets/configs/splits/columns/probs
+    if len(dataset_names) != len(dataset_config_names):
+        raise ValueError(
+            f"Ensure one config is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(dataset_config_names)} configs."
+        )
+
+    if splits is not None and len(splits) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one split is passed for each dataset, got {len(dataset_names)} datasets and {len(splits)} splits."
+        )
+
+    if metadata_dataset_names is not None and len(metadata_dataset_names) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one metadata dataset is passed for each dataset, got {len(dataset_names)} datasets and {len(metadata_dataset_names)} metadata datasets."
+        )
+
+    if dataset_samples is not None:
+        if len(dataset_samples) != len(dataset_names):
+            raise ValueError(
+                f"Ensure one sample is passed for each dataset, got {len(dataset_names)} datasets and "
+                f"{len(dataset_samples)} samples."
+            )
+        dataset_samples = [float(ds_sample) for ds_sample in dataset_samples]
+    else:
+        dataset_samples = [None] * len(dataset_names)
+
+    splits = splits if splits is not None else [default_split for _ in range(len(dataset_names))]
+
+    dataset_names_dict = []
+    for i, ds_name in enumerate(dataset_names):
+        dataset_names_dict.append(
+            {
+                "name": ds_name,
+                "config": dataset_config_names[i],
+                "split": splits[i],
+                "metadata_dataset_name": metadata_dataset_names[i],
+                "samples": dataset_samples[i],
+            }
+        )
+    return dataset_names_dict
+
+
+def load_multiple_datasets(
+    accelerator: Accelerator,
+    dataset_names: Union[List, str],
+    dataset_config_names: Union[List, str],
+    metadata_dataset_names: Optional[str] = None,
+    splits: Optional[Union[List, str]] = None,
+    label_column_names: Optional[List] = None,
+    stopping_strategy: Optional[str] = "first_exhausted",
+    dataset_samples: Optional[Union[List, np.array]] = None,
+    streaming: Optional[bool] = False,
+    seed: Optional[int] = None,
+    id_column_name: Optional[str] = None,
+    columns_to_keep: Optional[Set[str]] = None,
+    prompt_column_name: Optional[str] = None,
+    sampling_rate: Optional[int] = None,
+    audio_column_name: Optional[str] = None,
+    logger: Optional[logging.Logger] = None,
+    **kwargs,
+) -> Union[Dataset, IterableDataset]:
+    dataset_names_dict = convert_dataset_str_to_list(
+        dataset_names, dataset_config_names, metadata_dataset_names, splits, label_column_names, dataset_samples
+    )
+
+    if dataset_samples is not None:
+        dataset_samples = [ds_dict["samples"] for ds_dict in dataset_names_dict]
+        probabilities = np.array(dataset_samples) / np.sum(dataset_samples)
+    else:
+        probabilities = None
+
+    all_datasets = []
+    # iterate over the datasets we want to interleave
+    for dataset_dict in tqdm(dataset_names_dict, desc="Combining datasets..."):
+        with accelerator.main_process_first():
+            dataset = load_dataset(
+                dataset_dict["name"],
+                dataset_dict["config"],
+                split=dataset_dict["split"],
+                streaming=streaming,
+                **kwargs,
+            )
+            dataset_features = dataset.features.keys()
+
+            if sampling_rate is not None and audio_column_name is not None:
+                # resample target audio
+                dataset = dataset.cast_column(audio_column_name, datasets.features.Audio(sampling_rate=sampling_rate))
+
+            metadata_dataset_name = dataset_dict["metadata_dataset_name"]
+            if metadata_dataset_name is not None:
+                logger.info(
+                    f'Merging {dataset_dict["name"]} - {dataset_dict["split"]} with {metadata_dataset_name} - {dataset_dict["split"]}'
+                )
+                metadata_dataset = load_dataset(
+                    metadata_dataset_name,
+                    dataset_dict["config"],
+                    split=dataset_dict["split"],
+                    streaming=streaming,
+                    **kwargs,
+                )
+
+                # TODO(YL): I forgot to create unique ids for MLS english.
+                # To iterate faster, I bypass the original id check and do another one. - Done once because assuming it won't change next time
+                # if dataset_dict["name"] == "parler-tts/mls_eng_10k":
+                #     def concat_ids(book_id, speaker_id, begin_time):
+                #         return {"id": f"{book_id}_{speaker_id}_{str(begin_time).replace('.', '_')}"}
+                #     dataset = dataset.map(concat_ids, input_columns=["book_id", "speaker_id", "begin_time"], num_proc=24)
+                #     metadata_dataset = metadata_dataset.map(concat_ids, input_columns=["book_id", "speaker_id", "begin_time"], num_proc=24)
+                #     metadata_dataset = metadata_dataset.rename_column(id_column_name, f"metadata_{id_column_name}")
+
+                if dataset_dict["name"] != "parler-tts/mls_eng_10k":
+                    if id_column_name is not None and id_column_name not in dataset.column_names:
+                        raise ValueError(
+                            f"id_column_name={id_column_name} but has not been found in the dataset columns"
+                            f"- one of {', '.join(list(dataset.column_names))}."
+                        )
+                    if id_column_name is not None and id_column_name not in metadata_dataset.column_names:
+                        raise ValueError(
+                            f"id_column_name={id_column_name} but has not been found in the metadata dataset columns"
+                            f"- one of {', '.join(list(metadata_dataset.column_names))}."
+                        )
+                    elif id_column_name is not None:
+                        metadata_dataset = metadata_dataset.rename_column(id_column_name, f"metadata_{id_column_name}")
+
+                metadata_columns_to_remove = set(metadata_dataset.column_names).intersection(set(dataset.column_names))
+
+                if prompt_column_name is not None:
+                    # We might have applied some transformations to the prompts (e.g  punctuation restoration)
+                    # so we make sure to remove it from the original dataset
+                    if prompt_column_name in dataset.column_names:
+                        logger.info(
+                            f"REMOVE {prompt_column_name} from dataset {dataset_dict['name']} - dataset_dict['split']"
+                        )
+                        dataset.remove_columns(prompt_column_name)
+
+                metadata_columns_to_remove = set(metadata_dataset.column_names).intersection(set(dataset.column_names))
+                metadata_dataset = metadata_dataset.remove_columns(metadata_columns_to_remove)
+
+                dataset = concatenate_datasets([dataset, metadata_dataset], axis=1)
+
+                if id_column_name is not None and dataset_dict["name"] != "parler-tts/mls_eng_10k":
+                    if (
+                        len(
+                            dataset.filter(
+                                lambda id1, id2: id1 != id2,
+                                input_columns=[id_column_name, f"metadata_{id_column_name}"],
+                            )
+                        )
+                        != 0
+                    ):
+                        raise ValueError(
+                            f"Concatenate didn't work. Some ids don't correspond on dataset {dataset_dict['name']}"
+                        )
+
+                dataset_features = dataset.features.keys()
+
+            if columns_to_keep is not None:
+                dataset = dataset.remove_columns(set(dataset_features - columns_to_keep))
+        all_datasets.append(dataset)
+
+    if len(all_datasets) == 1:
+        # we have a single dataset so just return it as is
+        return all_datasets[0]
+
+    if streaming:
+        interleaved_dataset = interleave_datasets(
+            all_datasets,
+            stopping_strategy=stopping_strategy,
+            probabilities=probabilities,
+            seed=seed,
+        )
+    else:
+        with accelerator.main_process_first():
+            interleaved_dataset = concatenate_datasets(all_datasets)
+
+    return interleaved_dataset

training/eval.py

@@ -0,0 +1,59 @@
+import torch
+import evaluate
+from transformers import AutoModel, AutoProcessor, pipeline, WhisperForConditionalGeneration, WhisperTokenizer, WhisperTokenizerFast
+
+
+def clap_similarity(clap_model_name_or_path, texts, audios, device):
+    clap = AutoModel.from_pretrained(clap_model_name_or_path)
+    clap_processor = AutoProcessor.from_pretrained(clap_model_name_or_path)
+    clap_inputs = clap_processor(text=texts, audios=audios, padding=True, return_tensors="pt").to(device)
+    clap.to(device)
+    with torch.no_grad():
+        text_features = clap.get_text_features(
+            clap_inputs["input_ids"], attention_mask=clap_inputs.get("attention_mask", None)
+        )
+        audio_features = clap.get_audio_features(clap_inputs["input_features"])
+
+        cosine_sim = torch.nn.functional.cosine_similarity(audio_features, text_features, dim=1, eps=1e-8)
+
+    clap.to("cpu")
+    clap_inputs.to("cpu")
+    return cosine_sim.mean().to("cpu")
+
+
+def wer(asr_model_name_or_path, prompts, audios, device, per_device_eval_batch_size, sampling_rate):
+    metric = evaluate.load("wer")
+    asr_pipeline = pipeline(model=asr_model_name_or_path, device=device)
+
+    return_language = None
+    if isinstance(asr_pipeline.model, WhisperForConditionalGeneration):
+        return_language = True
+
+    transcriptions = asr_pipeline(
+        [{"raw": audio, "sampling_rate": sampling_rate} for audio in audios],
+        batch_size=int(per_device_eval_batch_size),
+        return_language=return_language,
+    )
+
+    if isinstance(asr_pipeline.tokenizer, (WhisperTokenizer, WhisperTokenizerFast)):
+        tokenizer = asr_pipeline.tokenizer
+    else:
+        tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-large-v3")
+
+    english_normalizer = tokenizer.normalize
+    basic_normalizer = tokenizer.basic_normalize
+
+    normalized_predictions = []
+    normalized_references = []
+
+    for pred, ref in zip(transcriptions, prompts):
+        normalizer = english_normalizer if return_language and pred["chunks"][0]["language"] == "english" else basic_normalizer
+        norm_ref = normalizer(ref)
+        if len(norm_ref) > 0:
+            norm_pred = normalizer(pred["text"])
+            normalized_predictions.append(norm_pred)
+            normalized_references.append(norm_ref)
+
+    word_error = 100 * metric.compute(predictions=normalized_predictions, references=normalized_references)
+
+    return word_error, [t["text"] for t in transcriptions]

training/run_parler_tts_training.py

@@ -0,0 +1,1038 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+""" Train Parler-TTS using 🤗 Accelerate"""
+
+import logging
+import os
+import re
+import sys
+import time
+from multiprocess import set_start_method
+from datetime import timedelta
+
+from tqdm import tqdm
+from pathlib import Path
+
+import torch
+from torch.utils.data import DataLoader
+
+import datasets
+from datasets import DatasetDict, Dataset, IterableDataset, concatenate_datasets
+
+from huggingface_hub import HfApi
+
+import transformers
+from transformers import AutoFeatureExtractor, AutoTokenizer, HfArgumentParser
+from transformers.trainer_pt_utils import LengthGroupedSampler
+from transformers.optimization import get_scheduler
+from transformers.utils import send_example_telemetry
+
+
+from accelerate import Accelerator
+from accelerate.utils import set_seed, AutocastKwargs, InitProcessGroupKwargs, TorchDynamoPlugin
+from accelerate.utils.memory import release_memory
+
+from parler_tts import (
+    ParlerTTSConfig,
+    ParlerTTSForConditionalGeneration,
+    build_delay_pattern_mask,
+)
+
+from training.utils import get_last_checkpoint, rotate_checkpoints, log_pred, log_metric
+from training.arguments import ModelArguments, DataTrainingArguments, ParlerTTSTrainingArguments
+from training.data import load_multiple_datasets, DataCollatorParlerTTSWithPadding, DataCollatorEncodecWithPadding
+from training.eval import clap_similarity, wer
+
+
+logger = logging.getLogger(__name__)
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, ParlerTTSTrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your Python/PyTorch versions.
+    send_example_telemetry("run_parler_tts", model_args, data_args)
+
+    if training_args.dtype == "float16":
+        mixed_precision = "fp16"
+    elif training_args.dtype == "bfloat16":
+        mixed_precision = "bf16"
+    else:
+        mixed_precision = "no"
+
+    if data_args.pad_to_max_length and (
+        data_args.max_duration_in_seconds is None
+        or data_args.max_prompt_token_length is None
+        or data_args.max_description_token_length is None
+    ):
+        raise ValueError(
+            "`pad_to_max_length` is `True` but one of the following parameters has not been set: `max_duration_in_seconds`, `max_prompt_token_length`, `max_description_token_length`"
+        )
+
+    padding = "max_length" if data_args.pad_to_max_length else "longest"
+
+    ####### A. Preparation
+    kwargs_handlers = [InitProcessGroupKwargs(timeout=timedelta(minutes=60))]
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=training_args.gradient_accumulation_steps,
+        mixed_precision=mixed_precision,
+        log_with=training_args.report_to,
+        project_dir=training_args.output_dir,
+        kwargs_handlers=kwargs_handlers,
+    )
+
+    accelerator.init_trackers(
+        project_name=data_args.wandb_project,
+        config={
+            "learning_rate": training_args.learning_rate,
+            "model_name_or_path": model_args.model_name_or_path,
+            "num_train_epochs": training_args.num_train_epochs,
+            "gradient_accumulation_steps": training_args.gradient_accumulation_steps,
+            "per_device_train_batch_size": training_args.per_device_train_batch_size,
+            "global_batch_size": training_args.per_device_train_batch_size * accelerator.num_processes,
+            "mixed_precision": mixed_precision,
+            "lr_scheduler_type": training_args.lr_scheduler_type,
+            "warmup_steps": training_args.warmup_steps,
+            "freeze_text_encoder": model_args.freeze_text_encoder,
+            "max_duration_in_seconds": data_args.max_duration_in_seconds,
+            "weight_decay": training_args.weight_decay,
+            "adam_beta1": training_args.adam_beta1,
+            "adam_beta2": training_args.adam_beta2,
+            "temperature": model_args.temperature,
+        },
+        init_kwargs={"wandb": {"name": data_args.wandb_run_name}} if data_args.wandb_run_name else {},
+    )
+
+    # Detecting last checkpoint and eventually continue from last checkpoint
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            logger.info(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if accelerator.is_main_process else logging.WARN)
+
+    # Log a small summary on each proces
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
+        f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
+    )
+
+    # Set the verbosity to info of the Transformers logger (on main process only)
+    if accelerator.is_local_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+    num_workers = data_args.preprocessing_num_workers
+
+    # 1. First, lett's instantiate the feature extractor, tokenizers and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature extractor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name or model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
+    )
+    sampling_rate = feature_extractor.sampling_rate
+
+    # load prompt tokenizer
+    prompt_tokenizer = AutoTokenizer.from_pretrained(
+        model_args.prompt_tokenizer_name or model_args.description_tokenizer_name or model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
+        use_fast=model_args.use_fast_tokenizer,
+    )
+
+    # load description tokenizer
+    description_tokenizer = AutoTokenizer.from_pretrained(
+        model_args.description_tokenizer_name or model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
+        use_fast=model_args.use_fast_tokenizer,
+        padding_side="left",
+    )
+
+    if model_args.use_fast_tokenizer:
+        logger.warning(
+            "Disabling fast tokenizer warning: https://github.com/huggingface/transformers/blob/main/src/transformers/tokenization_utils_base.py#L3231-L3235"
+        )
+        prompt_tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True
+        description_tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True
+
+    # 2. Now, let's load the dataset
+
+    if data_args.save_to_disk is not None:
+        os.makedirs(data_args.save_to_disk, exist_ok=True)
+
+    # assume that the dataset has been saved to `save_to_disk` if the latter is not empty
+    dataset_was_precomputed = len(os.listdir(data_args.save_to_disk)) > 0
+    if dataset_was_precomputed:
+        vectorized_datasets = datasets.load_from_disk(data_args.save_to_disk)
+    else:
+        raw_datasets = DatasetDict()
+
+        columns_to_keep = {
+            "target_audio_column_name": data_args.target_audio_column_name,
+            "prompt_column_name": data_args.prompt_column_name,
+        }
+        if data_args.description_column_name is not None:
+            columns_to_keep["description_column_name"] = data_args.description_column_name
+
+        if training_args.do_train:
+            raw_datasets["train"] = load_multiple_datasets(
+                accelerator,
+                data_args.train_dataset_name,
+                data_args.train_dataset_config_name,
+                metadata_dataset_names=data_args.train_metadata_dataset_name,
+                splits=data_args.train_split_name,
+                dataset_samples=data_args.train_dataset_samples,
+                seed=training_args.seed,
+                cache_dir=model_args.cache_dir,
+                num_proc=data_args.preprocessing_num_workers,
+                id_column_name=data_args.id_column_name,
+                columns_to_keep=columns_to_keep.values(),
+                prompt_column_name=data_args.prompt_column_name,
+                audio_column_name=data_args.target_audio_column_name,
+                sampling_rate=sampling_rate,
+                logger=logger,
+                # streaming=data_args.streaming, TODO(SG): optionally enable streaming mode
+            )
+
+            for key in columns_to_keep:
+                if columns_to_keep[key] not in raw_datasets["train"].column_names:
+                    raise ValueError(
+                        f"--{key} '{columns_to_keep[key]}' not found in dataset '{data_args.train_dataset_name}'."
+                        f" Make sure to set `--{key}` to the correct audio column - one of"
+                        f" {', '.join(raw_datasets['train'].column_names)}."
+                    )
+
+            if data_args.max_train_samples is not None:
+                raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+        if training_args.do_eval:
+            raw_datasets["eval"] = load_multiple_datasets(
+                accelerator,
+                data_args.eval_dataset_name if data_args.eval_dataset_name else data_args.train_dataset_name,
+                data_args.eval_dataset_config_name
+                if data_args.eval_dataset_config_name
+                else data_args.train_dataset_config_name,
+                metadata_dataset_names=data_args.eval_metadata_dataset_name,
+                splits=data_args.eval_split_name,
+                cache_dir=model_args.cache_dir,
+                num_proc=data_args.preprocessing_num_workers,
+                id_column_name=data_args.id_column_name,
+                columns_to_keep=columns_to_keep.values(),
+                prompt_column_name=data_args.prompt_column_name,
+                audio_column_name=data_args.target_audio_column_name,
+                sampling_rate=sampling_rate,
+                logger=logger,
+                # streaming=data_args.streaming, TODO(SG): optionally enable streaming mode
+            )
+
+            if data_args.max_eval_samples is not None:
+                raw_datasets["eval"] = (
+                    raw_datasets["eval"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
+                )
+
+    # 3. Next, let's load the config.
+    config = ParlerTTSConfig.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
+    )
+
+    # update pad token id and decoder_start_token_id
+    config.update(
+        {
+            "pad_token_id": model_args.pad_token_id if model_args.pad_token_id is not None else config.pad_token_id,
+            "decoder_start_token_id": model_args.decoder_start_token_id
+            if model_args.decoder_start_token_id is not None
+            else config.decoder_start_token_id,
+        }
+    )
+
+    # create model
+    model = ParlerTTSForConditionalGeneration.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        token=data_args.token,
+        trust_remote_code=data_args.trust_remote_code,
+    )
+    generation_config = model.generation_config
+
+    # enable gradient checkpointing if necessary
+    if training_args.gradient_checkpointing:
+        model.gradient_checkpointing_enable()
+
+    # 4. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # derive max & min input length for sample rate & max duration
+    sampling_rate = feature_extractor.sampling_rate
+    max_target_length = data_args.max_duration_in_seconds * sampling_rate
+    min_target_length = data_args.min_duration_in_seconds * sampling_rate
+    target_audio_column_name = data_args.target_audio_column_name
+    description_column_name = data_args.description_column_name
+    prompt_column_name = data_args.prompt_column_name
+    feature_extractor_input_name = feature_extractor.model_input_names[0]
+    audio_encoder_pad_token_id = config.decoder.pad_token_id
+    audio_encoder_eos_token_id = config.decoder.eos_token_id
+    audio_encoder_bos_token_id = generation_config.decoder_start_token_id
+    max_length = generation_config.max_length
+    num_codebooks = model.decoder.config.num_codebooks
+    bandwidth = model_args.bandwidth
+
+    # Freeze Encoders
+    model.freeze_encoders(model_args.freeze_text_encoder)
+
+    # Test all gather - used for warmout and avoiding timeout
+    test_tensor = torch.tensor([accelerator.process_index], device=accelerator.device)
+    gathered_tensor = accelerator.gather(test_tensor)
+    print("gathered_tensor", gathered_tensor)
+    accelerator.wait_for_everyone()
+
+    if not dataset_was_precomputed:
+        # Filter on text length
+        if description_column_name is not None and data_args.max_text_length is not None:
+            with accelerator.main_process_first():
+                # filter description that is shorter than max_text_length
+                raw_datasets = raw_datasets.filter(
+                    lambda x: len(x) < data_args.max_text_length,
+                    num_proc=num_workers,
+                    input_columns=[description_column_name],
+                )
+
+        # Preprocessing the dataset.
+        # We need to tokenize the texts.
+        def pass_through_processors(description, prompt):
+            batch = {}
+
+            batch["input_ids"] = description_tokenizer(description.strip())["input_ids"]
+            batch["prompt_input_ids"] = prompt_tokenizer(prompt.strip())["input_ids"]
+
+            return batch
+
+        with accelerator.main_process_first():
+            # this is a trick to avoid to rewrite the entire audio column which takes ages
+            vectorized_datasets = raw_datasets.map(
+                pass_through_processors,
+                remove_columns=next(iter(raw_datasets.values())).column_names,
+                input_columns=[description_column_name, prompt_column_name],
+                num_proc=num_workers,
+                desc="preprocess datasets",
+            )
+
+        # We use Accelerate to perform distributed inference
+        # T5 doesn't support fp16
+        autocast_kwargs = AutocastKwargs(enabled=(mixed_precision != "fp16"))
+
+        # Now we encode the audio labels with encodec.
+        ####### B. Encode audio
+
+        logger.info("*** Encode target audio with encodec ***")
+
+        # no need to prepare audio_decoder because used for inference without mixed precision
+        # see: https://huggingface.co/docs/accelerate/main/en/package_reference/accelerator#accelerate.Accelerator.prepare
+        if training_args.torch_compile:
+            audio_decoder = accelerator.prepare_model(model.audio_encoder, evaluation_mode=True)
+        else:
+            audio_decoder = model.audio_encoder
+
+        encoder_data_collator = DataCollatorEncodecWithPadding(
+            feature_extractor,
+            audio_column_name=target_audio_column_name,
+            feature_extractor_input_name=feature_extractor_input_name,
+            max_length=max_target_length,
+            padding=padding,
+        )
+
+        def apply_audio_decoder(batch):
+            len_audio = batch.pop("len_audio")
+            audio_decoder.to(batch["input_values"].device).eval()
+            with torch.no_grad():
+                labels = audio_decoder.encode(**batch, bandwidth=bandwidth)["audio_codes"]
+            output = {}
+            output["len_audio"] = len_audio
+            # (1, bsz, codebooks, seq_len) -> (bsz, seq_len, codebooks)
+            output["labels"] = labels.squeeze(0).transpose(1, 2)
+            output["ratio"] = torch.ones_like(len_audio) * labels.shape[-1] / len_audio.max()
+            return output
+
+        for split in vectorized_datasets:
+            data_loader = DataLoader(
+                raw_datasets[split],
+                batch_size=training_args.audio_encoder_per_device_batch_size,
+                collate_fn=encoder_data_collator,
+                num_workers=training_args.dataloader_num_workers,
+                pin_memory=True,
+            )
+            data_loader = accelerator.prepare(data_loader)
+
+            all_generated_labels = []
+            all_lens = []
+            for batch in tqdm(data_loader, disable=not accelerator.is_local_main_process):
+                generate_labels = apply_audio_decoder(batch)
+                generate_labels = accelerator.pad_across_processes(generate_labels, dim=1, pad_index=0)
+                generate_labels = accelerator.gather_for_metrics(generate_labels)
+
+                if accelerator.is_main_process:
+                    lab = generate_labels["labels"].cpu().transpose(1, 2).to(torch.int16)
+                    rat = generate_labels["ratio"].cpu().squeeze()
+                    lens = generate_labels["len_audio"].cpu().squeeze()
+                    lab = [l[:, : int(ratio * length)] for (l, ratio, length) in zip(lab, rat, lens)]
+
+                    all_generated_labels.extend(lab)
+                    all_lens.extend(lens)
+
+            # (1, codebooks, seq_len) where seq_len=1
+            bos_labels = torch.ones((1, num_codebooks, 1)) * audio_encoder_bos_token_id
+
+            if accelerator.is_main_process:
+                tmp_labels = Dataset.from_dict({"labels": all_generated_labels, "target_length": all_lens})
+                tmp_labels.save_to_disk(
+                    os.path.join(data_args.temporary_save_to_disk, split),
+                    num_proc=1 if split == "eval" else data_args.preprocessing_num_workers,
+                )
+            accelerator.wait_for_everyone()
+            del all_generated_labels
+
+            tmp_labels = datasets.load_from_disk(os.path.join(data_args.temporary_save_to_disk, split))
+            with accelerator.main_process_first():
+                vectorized_datasets[split] = concatenate_datasets([vectorized_datasets[split], tmp_labels], axis=1)
+
+            def postprocess_dataset(labels):
+                # (1, codebooks, seq_len)
+                labels = torch.tensor(labels).unsqueeze(0)
+                # add bos
+                labels = torch.cat([bos_labels, labels], dim=-1)
+
+                labels, delay_pattern_mask = build_delay_pattern_mask(
+                    labels,
+                    bos_token_id=audio_encoder_bos_token_id,
+                    pad_token_id=audio_encoder_eos_token_id,
+                    max_length=labels.shape[-1] + num_codebooks,
+                    num_codebooks=num_codebooks,
+                )
+
+                # the first ids of the delay pattern mask are precisely labels, we use the rest of the labels mask
+                # to take care of EOS
+                # we want labels to look like this:
+                #  - [B, a, b, E, E, E, E]
+                #  - [B, B, c, d, E, E, E]
+                #  - [B, B, B, e, f, E, E]
+                #  - [B, B, B, B, g, h, E]
+                labels = torch.where(delay_pattern_mask == -1, audio_encoder_eos_token_id, delay_pattern_mask)
+
+                # the first timestamp is associated to a row full of BOS, let's get rid of it
+                # we also remove the last timestampts (full of PAD)
+                output = {"labels": labels[:, 1:]}
+                return output
+
+            with accelerator.main_process_first():
+                vectorized_datasets[split] = vectorized_datasets[split].map(
+                    postprocess_dataset,
+                    num_proc=data_args.preprocessing_num_workers,  # this one is resource consuming if many processor.
+                    input_columns=["labels"],
+                    desc="Postprocessing labeling",
+                )
+
+        accelerator.free_memory()
+        del generate_labels, all_lens
+
+        with accelerator.main_process_first():
+            # NOTE: filtering is done at the end because in the `datasets` library, caching audio files is done after most operations
+            # caching audio files is time and disk-space consuming, so we want to avoid it at all costs, especially for large (>1Kh) audio datasets.
+            # That's also why we avoid to concat the processed datasets (vectorized_datasets) with the audio column present in raw_datasets.
+
+            def is_audio_in_length_range(length):
+                return length > min_target_length and length < max_target_length
+
+            # filter data that is shorter than min_target_length
+            vectorized_datasets = vectorized_datasets.filter(
+                is_audio_in_length_range,
+                num_proc=num_workers,
+                input_columns=["target_length"],
+            )
+
+            if description_column_name is not None and data_args.max_description_token_length is not None:
+                with accelerator.main_process_first():
+                    # filter description that is shorter than max_text_length
+                    vectorized_datasets = vectorized_datasets.filter(
+                        lambda x: len(x) < data_args.max_description_token_length,
+                        num_proc=num_workers,
+                        input_columns=["input_ids"],
+                    )
+
+            if data_args.max_prompt_token_length is not None:
+                with accelerator.main_process_first():
+                    # filter description that is shorter than max_text_length
+                    vectorized_datasets = vectorized_datasets.filter(
+                        lambda x: len(x) < data_args.max_prompt_token_length,
+                        num_proc=num_workers,
+                        input_columns=["prompt_input_ids"],
+                    )
+
+    if data_args.save_to_disk is not None and not dataset_was_precomputed:
+        if accelerator.is_main_process:
+            vectorized_datasets.save_to_disk(
+                data_args.save_to_disk,
+                num_proc=min(data_args.preprocessing_num_workers, len(vectorized_datasets["eval"]) - 1),
+            )
+        logger.info(f"Dataset saved at {data_args.save_to_disk}")
+
+    audio_max_length = None
+    if padding == "max_length":
+        audio_max_length = max(vectorized_datasets["train"]["target_length"])
+        with accelerator.main_process_first():
+            max_sample = vectorized_datasets["train"].filter(
+                lambda x: x == audio_max_length,
+                num_proc=num_workers,
+                input_columns=["target_length"],
+            )
+        audio_max_length = torch.tensor(max_sample[0]["labels"]).shape[1]
+
+    if training_args.group_by_length:
+        # apply a simple heuristic to take into account audio and text lengths
+        def add_target_lengths(target_length, prompt, description):
+            return {"target_length": target_length + len(prompt) + len(description)}
+
+        with accelerator.main_process_first():
+            vectorized_datasets = vectorized_datasets.map(
+                add_target_lengths,
+                num_proc=num_workers,
+                input_columns=["target_length", "prompt_input_ids", "input_ids"],
+            )
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only and data_args.save_to_disk is None:
+        raise ValueError(
+            "`preprocessing_only=True` but `save_to_disk` is not set. The latter should indicates where to save the dataset locally."
+        )
+    elif data_args.preprocessing_only:
+        logger.info(f"Data preprocessing finished. Files save at {data_args.save_to_disk}")
+        return
+
+    # 6. Next, we can prepare the training.
+
+    # Let's use word CLAP similary and WER metrics as our evaluation metrics,
+    def compute_metrics(audios, descriptions, prompts, device="cpu"):
+        results = {}
+        input_ids = descriptions
+        texts = description_tokenizer.batch_decode(input_ids, skip_special_tokens=True)
+        prompts = prompt_tokenizer.batch_decode(prompts, skip_special_tokens=True)
+        audios = [a.cpu().numpy() for a in audios]
+
+        clap_score = clap_similarity(model_args.clap_model_name_or_path, texts, audios, device)
+        results["clap"] = clap_score
+
+        word_error, transcriptions = wer(
+            model_args.asr_model_name_or_path,
+            prompts,
+            audios,
+            device,
+            training_args.per_device_eval_batch_size,
+            sampling_rate,
+        )
+        results["wer"] = word_error
+
+        return results, texts, prompts, audios, transcriptions
+
+    # Define Training Schedule
+    # Store some constants
+    per_device_train_batch_size = int(training_args.per_device_train_batch_size)
+    train_batch_size = per_device_train_batch_size * accelerator.num_processes
+    gradient_accumulation_steps = int(training_args.gradient_accumulation_steps)
+    per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
+
+    if training_args.max_steps < 0:
+        num_epochs = int(training_args.num_train_epochs)
+        steps_per_epoch = len(vectorized_datasets["train"]) // (train_batch_size * gradient_accumulation_steps)
+        total_train_steps = steps_per_epoch * num_epochs
+    elif training_args.max_steps > 0:
+        logger.info("max_steps is given, it will override any value given in num_train_epochs")
+        total_train_steps = int(training_args.max_steps)
+        # Setting a very large number of epochs so we go as many times as necessary over the iterator.
+        num_epochs = sys.maxsize
+        steps_per_epoch = total_train_steps
+
+    if training_args.eval_steps is None:
+        logger.info(f"eval_steps is not set, evaluating at the end of each epoch")
+        eval_steps = steps_per_epoch
+    else:
+        eval_steps = training_args.eval_steps
+
+    # T5 doesn't support fp16
+    autocast_kwargs = AutocastKwargs(enabled=(mixed_precision != "fp16"))
+
+    # Define optimizer, LR scheduler, collator
+    optimizer = torch.optim.AdamW(
+        params=model.parameters(),
+        lr=training_args.learning_rate,
+        betas=(training_args.adam_beta1, training_args.adam_beta2),
+        eps=training_args.adam_epsilon,
+        weight_decay=training_args.weight_decay,
+    )
+
+    # LR scheduler gets stepped by `num_processes` each time -> account for this in warmup / total steps
+    lr_scheduler = get_scheduler(
+        name=training_args.lr_scheduler_type,
+        optimizer=optimizer,
+        num_warmup_steps=training_args.get_warmup_steps(total_train_steps) * accelerator.num_processes,
+        num_training_steps=total_train_steps * accelerator.num_processes,
+    )
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorParlerTTSWithPadding(
+        prompt_tokenizer=prompt_tokenizer,
+        description_tokenizer=description_tokenizer,
+        pad_to_multiple_of=data_args.pad_to_multiple_of,
+        padding=padding,
+        prompt_max_length=data_args.max_prompt_token_length,
+        description_max_length=data_args.max_description_token_length,
+        audio_max_length=audio_max_length,
+    )
+
+    # Prepare everything with accelerate
+    model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {total_train_steps * train_batch_size * gradient_accumulation_steps}")
+    logger.info("  Instantaneous batch size per device =" f" {per_device_train_batch_size}")
+    logger.info("  Gradient accumulation steps =" f" {gradient_accumulation_steps}")
+    logger.info(
+        f"  Total train batch size (w. parallel & distributed) = {train_batch_size * gradient_accumulation_steps}"
+    )
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    # ======================== Training ================================
+    train_time = 0
+    train_start = time.time()
+    steps_trained_progress_bar = tqdm(
+        range(total_train_steps), desc="Train steps ... ", position=0, disable=not accelerator.is_local_main_process
+    )
+    continue_training = True
+    epochs_trained = 0
+    cur_step = 0
+
+    checkpoint = None
+    if training_args.resume_from_checkpoint is not None:
+        checkpoint = training_args.resume_from_checkpoint
+    elif last_checkpoint is not None:
+        checkpoint = last_checkpoint
+
+    if accelerator.is_main_process:
+        if training_args.output_dir is not None:
+            os.makedirs(training_args.output_dir, exist_ok=True)
+        if training_args.push_to_hub:
+            api = HfApi(token=training_args.hub_token)
+
+            # Create repo (repo_name from args or inferred)
+            repo_name = training_args.hub_model_id
+            if repo_name is None:
+                repo_name = Path(training_args.output_dir).absolute().name
+            repo_id = api.create_repo(repo_name, exist_ok=True).repo_id
+
+            with open(os.path.join(training_args.output_dir, ".gitignore"), "w+") as gitignore:
+                if "wandb" not in gitignore:
+                    gitignore.write("wandb\n")
+    accelerator.wait_for_everyone()
+
+    # Now save everything to be able to create a single processor later
+    # make sure all processes wait until data is saved
+    with accelerator.main_process_first():
+        # only the main process saves them
+        if accelerator.is_main_process:
+            # save feature extractor, tokenizer and config
+            if (
+                model_args.prompt_tokenizer_name is None
+                and model_args.description_tokenizer_name
+                or (model_args.prompt_tokenizer_name == model_args.description_tokenizer_name)
+            ):
+                prompt_tokenizer.save_pretrained(training_args.output_dir)
+            else:
+                logger.warning(
+                    f"Prompt tokenizer ('{model_args.prompt_tokenizer_name}') and description tokenizer ('{model_args.description_tokenizer_name}') are not the same. Saving only the prompt tokenizer."
+                )
+                prompt_tokenizer.save_pretrained(training_args.output_dir)
+
+            feature_extractor.save_pretrained(training_args.output_dir)
+            config.save_pretrained(training_args.output_dir)
+
+    if checkpoint is not None:
+        accelerator.load_state(checkpoint)
+        # Find num steps and epoch from saved state string pattern
+        pattern = r"checkpoint-(\d+)-epoch-(\d+)"
+        match = re.search(pattern, checkpoint)
+        cur_step = int(match.group(1))
+        epochs_trained = int(match.group(2))
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info(f"  Continuing training from epoch {epochs_trained}")
+        logger.info(f"  Continuing training from global step {cur_step}")
+
+        steps_trained_progress_bar.update(cur_step)
+
+        for epoch in range(0, epochs_trained):
+            vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
+
+        if training_args.max_steps < 0:
+            # we know exactly the number of steps per epoch, so can skip through the required number of batches
+            resume_step = (cur_step - epochs_trained * steps_per_epoch) * gradient_accumulation_steps
+        else:
+            # Currently we don't know how many steps we've taken in the current epoch
+            # So we just shuffle the dataset one extra time and start from a fresh epoch
+            # This is "good enough" for our purposes but not fully correct
+            resume_step = None
+            vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
+    else:
+        resume_step = None
+
+    gen_kwargs = {
+        "do_sample": model_args.do_sample,
+        "temperature": model_args.temperature,
+        "max_length": model_args.max_length,
+        # Because of the delayed pattern mask, generation might stop earlier because of unexpected behaviour
+        # on the first tokens of the codebooks that are delayed.
+        # This fix the issue.
+        "min_new_tokens": num_codebooks + 1,
+    }
+    for key in gen_kwargs:
+        generation_config.key = gen_kwargs[key]
+
+    # Define gradient update step fn
+    def train_step(
+        batch,
+        accelerator,
+        autocast_kwargs,
+    ):
+        model.train()
+
+        if mixed_precision == "fp16":
+            # fp16 doesn't work with T5-like models
+            with accelerator.autocast(autocast_handler=autocast_kwargs):
+                if training_args.parallel_mode.value != "distributed":
+                    encoder_outputs = model.text_encoder(
+                        input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
+                    )
+                else:
+                    encoder_outputs = model.module.text_encoder(
+                        input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
+                    )
+                batch["encoder_outputs"] = encoder_outputs
+
+        outputs = model(**batch)
+        # CE (data) loss
+        ce_loss = outputs.loss
+
+        metrics = {"loss": ce_loss}
+        return ce_loss, metrics
+
+    # Define eval fn
+    def eval_step(
+        batch,
+        accelerator,
+        autocast_kwargs,
+    ):
+        eval_model = model if not training_args.torch_compile else model._orig_mod
+        eval_model.eval()
+
+        if mixed_precision == "fp16":
+            # fp16 doesn't work with T5-like models
+            with accelerator.autocast(autocast_handler=autocast_kwargs):
+                with torch.no_grad():
+                    if training_args.parallel_mode.value != "distributed" or training_args.torch_compile:
+                        encoder_outputs = eval_model.text_encoder(
+                            input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
+                        )
+                    else:
+                        encoder_outputs = eval_model.module.text_encoder(
+                            input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
+                        )
+                batch["encoder_outputs"] = encoder_outputs
+
+        with torch.no_grad():
+            outputs = eval_model(**batch)
+        # CE (data) loss
+        ce_loss = outputs.loss
+        metrics = {"loss": ce_loss}
+        return metrics
+
+    def generate_step(batch):
+        batch.pop("decoder_attention_mask", None)
+        eval_model = accelerator.unwrap_model(model, keep_fp32_wrapper=mixed_precision != "fp16").eval()
+        if training_args.torch_compile:
+            eval_model = model._orig_mod
+
+        output_audios = eval_model.generate(**batch, **gen_kwargs)
+        output_audios = accelerator.pad_across_processes(output_audios, dim=1, pad_index=0)
+        return output_audios
+
+    for epoch in range(epochs_trained, num_epochs):
+        vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
+        sampler = None
+        if training_args.group_by_length:
+            sampler = LengthGroupedSampler(train_batch_size, lengths=vectorized_datasets["train"]["target_length"])
+        train_dataloader = DataLoader(
+            vectorized_datasets["train"],
+            collate_fn=data_collator,
+            batch_size=per_device_train_batch_size,
+            sampler=sampler,
+            num_workers=training_args.dataloader_num_workers,
+            pin_memory=training_args.dataloader_pin_memory,
+        )
+        train_dataloader = accelerator.prepare(train_dataloader)
+        if hasattr(train_dataloader, "dataset") and isinstance(train_dataloader.dataset, IterableDataset):
+            train_dataloader.dataset.set_epoch(epoch)
+
+        if resume_step is not None:
+            # Skip the first N batches in the dataloader when resuming from a checkpoint
+            train_dataloader = accelerator.skip_first_batches(train_dataloader, resume_step)
+            resume_step = None
+
+        for batch in train_dataloader:
+            with accelerator.accumulate(model):
+                loss, train_metric = train_step(batch, accelerator, autocast_kwargs)
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    accelerator.clip_grad_norm_(model.parameters(), training_args.max_grad_norm)
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            # Check if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                steps_trained_progress_bar.update(1)
+                cur_step += 1
+
+                if cur_step % training_args.logging_steps == 0:
+                    steps_trained_progress_bar.write(
+                        f"Step... ({cur_step} / {total_train_steps} | Loss:"
+                        f" {train_metric['loss']}, Learning Rate:"
+                        f" {lr_scheduler.get_last_lr()[0]})"
+                    )
+                    log_metric(
+                        accelerator,
+                        metrics=train_metric,
+                        learning_rate=lr_scheduler.get_last_lr()[0],
+                        train_time=train_time + time.time() - train_start,
+                        step=cur_step,
+                        epoch=epoch,
+                        prefix="train",
+                    )
+
+                # save checkpoint and weights after each save_steps and at the end of training
+                if (cur_step % training_args.save_steps == 0) or cur_step == total_train_steps:
+                    intermediate_dir = os.path.join(training_args.output_dir, f"checkpoint-{cur_step}-epoch-{epoch}")
+                    # safe_serialization=False to avoid shared tensors saving issue (TODO(YL): it's a temporary fix)
+                    # https://github.com/huggingface/transformers/issues/27293#issuecomment-1872560074
+                    accelerator.save_state(output_dir=intermediate_dir, safe_serialization=False)
+                    config.save_pretrained(intermediate_dir)
+                    generation_config.save_pretrained(intermediate_dir)
+                    accelerator.wait_for_everyone()
+                    if accelerator.is_main_process:
+                        checkpoints_to_be_deleted = rotate_checkpoints(
+                            training_args.save_total_limit, output_dir=training_args.output_dir, logger=logger
+                        )
+
+                        if cur_step == total_train_steps:
+                            # un-wrap student model for save
+                            unwrapped_model = accelerator.unwrap_model(model)
+                            unwrapped_model.save_pretrained(training_args.output_dir)
+
+                        if training_args.push_to_hub:
+                            api.upload_folder(
+                                repo_id=repo_id,
+                                folder_path=training_args.output_dir,
+                                commit_message=f"Saving train state of step {cur_step}",
+                                run_as_future=True,
+                                delete_patterns=checkpoints_to_be_deleted,
+                            )
+
+                if training_args.do_eval and (cur_step % eval_steps == 0 or cur_step == total_train_steps):
+                    train_time += time.time() - train_start
+                    # ======================== Evaluating ==============================
+                    eval_metrics = []
+                    eval_preds = []
+                    eval_descriptions = []
+                    eval_prompts = []
+                    eval_start = time.time()
+
+                    # release training input batch
+                    batch = release_memory(batch)
+
+                    validation_dataloader = DataLoader(
+                        vectorized_datasets["eval"],
+                        collate_fn=data_collator,
+                        batch_size=per_device_eval_batch_size,
+                        drop_last=False,
+                        num_workers=training_args.dataloader_pin_memory,
+                        pin_memory=training_args.dataloader_pin_memory,
+                    )
+                    validation_dataloader = accelerator.prepare(validation_dataloader)
+
+                    for batch in tqdm(
+                        validation_dataloader,
+                        desc=f"Evaluating - Inference ...",
+                        position=2,
+                        disable=not accelerator.is_local_main_process,
+                    ):
+                        # Model forward
+                        eval_metric = eval_step(batch, accelerator, autocast_kwargs)
+                        eval_metric = accelerator.gather_for_metrics(eval_metric)
+                        eval_metrics.append(eval_metric)
+
+                    if training_args.predict_with_generate:
+                        validation_dataloader = DataLoader(
+                            vectorized_datasets["eval"],
+                            collate_fn=data_collator,
+                            batch_size=per_device_eval_batch_size,
+                            drop_last=False,
+                            num_workers=training_args.dataloader_pin_memory,
+                            pin_memory=training_args.dataloader_pin_memory,
+                        )
+                        validation_dataloader = accelerator.prepare(validation_dataloader)
+                        # generation
+                        for batch in tqdm(
+                            validation_dataloader,
+                            desc=f"Evaluating - Generation ...",
+                            position=2,
+                            disable=not accelerator.is_local_main_process,
+                        ):
+                            generated_audios = generate_step(batch)
+                            # Gather all predictions and targets
+                            generated_audios, input_ids, prompts = accelerator.pad_across_processes(
+                                (generated_audios, batch["input_ids"], batch["prompt_input_ids"]), dim=1, pad_index=0
+                            )
+                            generated_audios, input_ids, prompts = accelerator.gather_for_metrics(
+                                (generated_audios, input_ids, prompts)
+                            )
+                            eval_preds.extend(generated_audios.to("cpu"))
+                            eval_descriptions.extend(input_ids.to("cpu"))
+                            eval_prompts.extend(prompts.to("cpu"))
+
+                    eval_time = time.time() - eval_start
+                    # normalize eval metrics
+                    eval_metrics = {
+                        key: torch.mean(torch.cat([d[key].unsqueeze(0) for d in eval_metrics]))
+                        for key in eval_metrics[0]
+                    }
+
+                    # compute metrics
+                    metrics_desc = ""
+                    if training_args.predict_with_generate:
+                        metric_values, pred_descriptions, pred_prompts, audios, transcriptions = compute_metrics(
+                            eval_preds, eval_descriptions, eval_prompts, accelerator.device
+                        )
+                        eval_metrics.update(metric_values)
+                        metrics_desc = " ".join([f"Eval {key}: {value} |" for key, value in metric_values.items()])
+                        if "wandb" in training_args.report_to:
+                            log_pred(
+                                accelerator,
+                                pred_descriptions,
+                                pred_prompts,
+                                transcriptions,
+                                audios,
+                                sampling_rate=sampling_rate,
+                                step=cur_step,
+                                prefix="eval",
+                            )
+
+                    # Print metrics and update progress bar
+                    steps_trained_progress_bar.write(
+                        f"Eval results for step ({cur_step} / {total_train_steps} | Eval Loss: {eval_metrics['loss']} |"
+                        f" {metrics_desc})"
+                    )
+
+                    log_metric(
+                        accelerator,
+                        metrics=eval_metrics,
+                        train_time=eval_time,
+                        step=cur_step,
+                        epoch=epoch,
+                        prefix="eval",
+                    )
+
+                    # release eval batch and relax metrics
+                    eval_metrics = []
+                    eval_preds = []
+                    eval_descriptions = []
+                    eval_prompts = []
+                    batch = release_memory(batch)
+
+                    # flush the train metrics
+                    train_start = time.time()
+
+                # break condition
+                if cur_step == total_train_steps:
+                    continue_training = False
+                    break
+
+        if not continue_training:
+            break
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    set_start_method("spawn")
+    main()

training/utils.py

@@ -0,0 +1,127 @@
+import os
+import re
+import shutil
+from pathlib import Path
+from dataclasses import field
+from typing import Dict, List, Union
+
+import torch
+from wandb import Audio
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+_RE_CHECKPOINT = re.compile(r"^checkpoint-(\d+)-epoch-(\d+)$")
+
+
+def get_last_checkpoint(folder):
+    content = os.listdir(folder)
+    checkpoints = [
+        path
+        for path in content
+        if _RE_CHECKPOINT.search(path) is not None and os.path.isdir(os.path.join(folder, path))
+    ]
+    if len(checkpoints) == 0:
+        return
+    return os.path.join(folder, max(checkpoints, key=lambda x: int(_RE_CHECKPOINT.search(x).groups()[0])))
+
+
+def sorted_checkpoints(output_dir=None, checkpoint_prefix="checkpoint") -> List[str]:
+    """Helper function to sort saved checkpoints from oldest to newest."""
+    ordering_and_checkpoint_path = []
+
+    glob_checkpoints = [str(x) for x in Path(output_dir).glob(f"{checkpoint_prefix}-*") if os.path.isdir(x)]
+
+    for path in glob_checkpoints:
+        regex_match = re.match(f".*{checkpoint_prefix}-([0-9]+)", path)
+        if regex_match is not None and regex_match.groups() is not None:
+            ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))
+
+    checkpoints_sorted = sorted(ordering_and_checkpoint_path)
+    checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
+    return checkpoints_sorted
+
+
+def rotate_checkpoints(save_total_limit=None, output_dir=None, checkpoint_prefix="checkpoint", logger=None) -> Union[List, None]:
+    """Helper function to delete old checkpoints."""
+    if save_total_limit is None or save_total_limit <= 0:
+        return
+    # Check if we should delete older checkpoint(s)
+    checkpoints_sorted = sorted_checkpoints(output_dir=output_dir, checkpoint_prefix=checkpoint_prefix)
+    if len(checkpoints_sorted) <= save_total_limit:
+        return
+
+    number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - save_total_limit)
+    checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
+    for checkpoint in checkpoints_to_be_deleted:
+        logger.info(f"Deleting older checkpoint [{checkpoint}] due to args.save_total_limit")
+        shutil.rmtree(checkpoint, ignore_errors=True)
+    checkpoints_to_be_deleted = [f"*{Path(checkpoint).absolute().name}*"  for checkpoint in checkpoints_to_be_deleted]
+    return checkpoints_to_be_deleted
+
+
+def log_metric(
+    accelerator,
+    metrics: Dict,
+    train_time: float,
+    step: int,
+    epoch: int,
+    learning_rate: float = None,
+    prefix: str = "train",
+):
+    """Helper function to log all training/evaluation metrics with the correct prefixes and styling."""
+    log_metrics = {}
+    for k, v in metrics.items():
+        log_metrics[f"{prefix}/{k}"] = v
+    log_metrics[f"{prefix}/time"] = train_time
+    log_metrics[f"{prefix}/epoch"] = epoch
+    if learning_rate is not None:
+        log_metrics[f"{prefix}/learning_rate"] = learning_rate
+    accelerator.log(log_metrics, step=step)
+
+
+def log_pred(
+    accelerator,
+    pred_descriptions: List[str],
+    pred_prompts: List[str],
+    transcriptions: List[str],
+    audios: List[torch.Tensor],
+    sampling_rate: int,
+    step: int,
+    prefix: str = "eval",
+    num_lines: int = 200000,
+):
+    """Helper function to log target/predicted transcriptions to weights and biases (wandb)."""
+    if accelerator.is_main_process:
+        wandb_tracker = accelerator.get_tracker("wandb")
+        # pretty name for current step: step 50000 -> step 50k
+        cur_step_pretty = f"{int(step // 1000)}k" if step > 1000 else step
+        prefix_pretty = prefix.replace("/", "-")
+
+        # convert str data to a wandb compatible format
+        str_data = [[pred_descriptions[i], pred_prompts[i], transcriptions[i]] for i in range(len(pred_descriptions))]
+        # log as a table with the appropriate headers
+        wandb_tracker.log_table(
+            table_name=f"predictions/{prefix_pretty}-step-{cur_step_pretty}",
+            columns=["Target descriptions", "Target prompts", "Predicted transcriptions"],
+            data=str_data[:num_lines],
+            step=step,
+            commit=False,
+        )
+
+        # wandb can only loads 100 audios per step
+        wandb_tracker.log(
+            {
+                "Speech samples": [
+                    Audio(
+                        audio,
+                        caption=f"{pred_prompts[i]} --- DESCRIPTION: {pred_descriptions[i]}",
+                        sample_rate=sampling_rate,
+                    )
+                    for (i, audio) in enumerate(audios[: min(len(audios), 100)])
+                ]
+            },
+            step=step,
+        )