Upload Nano TTS checkpoint 500000

__init__.py

@@ -0,0 +1,31 @@
+from .configuration_nanotts import NanoTTSConfig
+from .modeling_nanotts_global_local import (
+    NanoTTSGenerationOutput,
+    NanoTTSGlobalLocalForCausalLM,
+    NanoTTSOutput,
+)
+from .tokenization_nanotts_sentencepiece import NanoTTSSentencePieceTokenizer
+
+try:
+    NanoTTSConfig.register_for_auto_class()
+except Exception:
+    pass
+
+for auto_class_name in ("AutoModel", "AutoModelForCausalLM"):
+    try:
+        NanoTTSGlobalLocalForCausalLM.register_for_auto_class(auto_class_name)
+    except Exception:
+        pass
+
+try:
+    NanoTTSSentencePieceTokenizer.register_for_auto_class("AutoTokenizer")
+except Exception:
+    pass
+
+__all__ = [
+    "NanoTTSConfig",
+    "NanoTTSGlobalLocalForCausalLM",
+    "NanoTTSSentencePieceTokenizer",
+    "NanoTTSGenerationOutput",
+    "NanoTTSOutput",
+]

config.json

@@ -0,0 +1,197 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "NanoTTSGlobalLocalForCausalLM"
+  ],
+  "attn_implementation": "flash_attention_2",
+  "audio_assistant_slot_token_id": 9,
+  "audio_codebook_sizes": [
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024,
+    1024
+  ],
+  "audio_end_token_id": 7,
+  "audio_pad_token_id": 1024,
+  "audio_start_token_id": 6,
+  "audio_tokenizer_pretrained_name_or_path": "OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano",
+  "audio_tokenizer_sample_rate": 48000,
+  "audio_tokenizer_type": "moss-audio-tokenizer-nano",
+  "audio_user_slot_token_id": 8,
+  "audio_vocab_size": 1024,
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": null,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "dtype": "float32",
+  "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,
+  "gpt2_config": {
+    "_name_or_path": "",
+    "activation_function": "gelu_new",
+    "add_cross_attention": false,
+    "architectures": null,
+    "attn_pdrop": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 1,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dtype": null,
+    "early_stopping": false,
+    "embd_pdrop": 0.0,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_epsilon": 1e-05,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "gpt2",
+    "n_ctx": 32768,
+    "n_embd": 768,
+    "n_head": 12,
+    "n_inner": 3072,
+    "n_layer": 12,
+    "n_positions": 32768,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 3,
+    "position_embedding_type": "rope",
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "reorder_and_upcast_attn": false,
+    "repetition_penalty": 1.0,
+    "resid_pdrop": 0.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rope_base": 10000.0,
+    "scale_attn_by_inverse_layer_idx": false,
+    "scale_attn_weights": true,
+    "sep_token_id": null,
+    "summary_activation": null,
+    "summary_first_dropout": 0.1,
+    "summary_proj_to_labels": true,
+    "summary_type": "cls_index",
+    "summary_use_proj": true,
+    "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,
+    "torchscript": false,
+    "transformers_version": "4.57.1",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 16384
+  },
+  "hidden_size": 768,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "im_end_token_id": 5,
+  "im_start_token_id": 4,
+  "initializer_range": 0.02,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "length_penalty": 1.0,
+  "local_transformer_attn_implementation": "flash_attention_2",
+  "local_transformer_layers": 1,
+  "max_length": 20,
+  "max_position_embeddings": 32768,
+  "min_length": 0,
+  "model_architecture": "global_local_transformer",
+  "model_type": "nano_tts",
+  "n_vq": 16,
+  "no_repeat_ngram_size": 0,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": 3,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "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": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": true,
+  "tokenizer_class": "NanoTTSSentencePieceTokenizer",
+  "tokenizer_use_fast": false,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": "4.57.1",
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "vocab_size": 16384,
+  "auto_map": {
+    "AutoConfig": "configuration_nanotts.NanoTTSConfig",
+    "AutoModel": "modeling_nanotts_global_local.NanoTTSGlobalLocalForCausalLM",
+    "AutoModelForCausalLM": "modeling_nanotts_global_local.NanoTTSGlobalLocalForCausalLM"
+  }
+}

configuration_nanotts.py

@@ -0,0 +1,105 @@
+# coding=utf-8
+from typing import Any, Dict, Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+
+class NanoTTSConfig(PretrainedConfig):
+    model_type = "nano_tts"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        gpt2_config: Optional[Union[GPT2Config, Dict[str, Any]]] = None,
+        n_vq: int = 8,
+        audio_vocab_size: Optional[int] = 1024,
+        audio_codebook_sizes: Optional[list[int]] = None,
+        audio_pad_token_id: int = 1024,
+        pad_token_id: int = 151643,
+        im_start_token_id: int = 151644,
+        im_end_token_id: int = 151645,
+        audio_start_token_id: int = 151652,
+        audio_end_token_id: int = 151653,
+        audio_user_slot_token_id: int = 151654,
+        audio_assistant_slot_token_id: int = 151656,
+        tokenizer_use_fast: bool = False,
+        audio_tokenizer_type: str = "moss-audio-tokenizer-nano",
+        audio_tokenizer_pretrained_name_or_path: Optional[str] = "OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano",
+        audio_tokenizer_sample_rate: int = 48000,
+        attn_implementation: str = "flash_attention_2",
+        initializer_range: float = 0.02,
+        model_architecture: str = "global_local_transformer",
+        local_transformer_layers: int = 4,
+        local_transformer_attn_implementation: Optional[str] = None,
+        **kwargs: Any,
+    ) -> None:
+        if isinstance(gpt2_config, dict):
+            self.gpt2_config = GPT2Config(**gpt2_config)
+        elif gpt2_config is None:
+            self.gpt2_config = GPT2Config()
+        else:
+            self.gpt2_config = gpt2_config
+
+        self.n_vq = int(n_vq)
+        if audio_codebook_sizes is None:
+            if audio_vocab_size is None:
+                raise ValueError("audio_vocab_size must be set when audio_codebook_sizes is not provided.")
+            resolved_audio_codebook_sizes = [int(audio_vocab_size)] * self.n_vq
+        else:
+            resolved_audio_codebook_sizes = [int(codebook_size) for codebook_size in audio_codebook_sizes]
+        if len(resolved_audio_codebook_sizes) != self.n_vq:
+            raise ValueError(
+                "audio_codebook_sizes must have length n_vq "
+                f"(expected {self.n_vq}, got {len(resolved_audio_codebook_sizes)})."
+            )
+        if any(codebook_size <= 0 for codebook_size in resolved_audio_codebook_sizes):
+            raise ValueError("audio_codebook_sizes must contain positive integers.")
+
+        max_audio_codebook_size = max(resolved_audio_codebook_sizes)
+        if audio_vocab_size is not None and int(audio_vocab_size) < max_audio_codebook_size:
+            raise ValueError(
+                "audio_vocab_size must be >= max(audio_codebook_sizes) "
+                f"(got {audio_vocab_size}, expected at least {max_audio_codebook_size})."
+            )
+
+        self.audio_codebook_sizes = resolved_audio_codebook_sizes
+        self.audio_vocab_size = (
+            max_audio_codebook_size if audio_vocab_size is None else int(audio_vocab_size)
+        )
+        self.audio_pad_token_id = int(audio_pad_token_id)
+        if self.audio_pad_token_id < max_audio_codebook_size:
+            raise ValueError(
+                "audio_pad_token_id must be >= max(audio_codebook_sizes) so pad stays outside every codebook "
+                f"(got {self.audio_pad_token_id}, max codebook size {max_audio_codebook_size})."
+            )
+        self.pad_token_id = pad_token_id
+        self.im_start_token_id = im_start_token_id
+        self.im_end_token_id = im_end_token_id
+        self.audio_start_token_id = audio_start_token_id
+        self.audio_end_token_id = audio_end_token_id
+        self.audio_user_slot_token_id = audio_user_slot_token_id
+        self.audio_assistant_slot_token_id = audio_assistant_slot_token_id
+        self.tokenizer_use_fast = tokenizer_use_fast
+        self.audio_tokenizer_type = audio_tokenizer_type
+        self.audio_tokenizer_pretrained_name_or_path = audio_tokenizer_pretrained_name_or_path
+        self.audio_tokenizer_sample_rate = audio_tokenizer_sample_rate
+        self.attn_implementation = attn_implementation
+        self.initializer_range = initializer_range
+        self.model_architecture = model_architecture
+        self.local_transformer_layers = local_transformer_layers
+        self.local_transformer_attn_implementation = (
+            attn_implementation
+            if local_transformer_attn_implementation is None
+            else local_transformer_attn_implementation
+        )
+        self.vocab_size = self.gpt2_config.vocab_size
+        self.hidden_size = self.gpt2_config.hidden_size
+        self.max_position_embeddings = self.gpt2_config.n_positions
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+    def to_dict(self) -> Dict[str, Any]:
+        output = super().to_dict()
+        output["gpt2_config"] = self.gpt2_config.to_dict()
+        return output

gpt2_decoder.py

@@ -0,0 +1,605 @@
+# coding=utf-8
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import pad_input, unpad_input
+
+    _FLASH_ATTN_AVAILABLE = True
+except Exception:
+    flash_attn_func = None
+    flash_attn_varlen_func = None
+    pad_input = None
+    unpad_input = None
+    _FLASH_ATTN_AVAILABLE = False
+
+
+@dataclass
+class PackedSequenceMetadata:
+    cu_seqlens: torch.Tensor
+    max_seqlen: int
+    indices: Optional[torch.Tensor] = None
+    batch_size: Optional[int] = None
+    seq_len: Optional[int] = None
+
+
+class NanoGPT2RotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, base: float = 10000.0) -> None:
+        super().__init__()
+        if dim % 2 != 0:
+            raise ValueError(f"RoPE head_dim must be even, got {dim}")
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(
+        self,
+        position_ids: torch.LongTensor,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if position_ids.ndim == 1:
+            position_ids = position_ids.unsqueeze(0)
+        freqs = torch.einsum("bs,d->bsd", position_ids.to(device=device, dtype=self.inv_freq.dtype), self.inv_freq)
+        cos = freqs.cos().repeat_interleave(2, dim=-1).unsqueeze(2).to(dtype=dtype)
+        sin = freqs.sin().repeat_interleave(2, dim=-1).unsqueeze(2).to(dtype=dtype)
+        return cos, sin
+
+
+def rotate_half(hidden_states: torch.Tensor) -> torch.Tensor:
+    even = hidden_states[..., ::2]
+    odd = hidden_states[..., 1::2]
+    return torch.stack((-odd, even), dim=-1).reshape_as(hidden_states)
+
+
+def apply_rotary_pos_emb(
+    hidden_states: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+) -> torch.Tensor:
+    return (hidden_states * cos) + (rotate_half(hidden_states) * sin)
+
+
+class NanoGPT2MLP(nn.Module):
+    def __init__(self, config: GPT2Config) -> None:
+        super().__init__()
+        hidden_size = int(config.hidden_size)
+        inner_size = int(config.n_inner or 4 * hidden_size)
+        self.fc_in = nn.Linear(hidden_size, inner_size)
+        self.fc_out = nn.Linear(inner_size, hidden_size)
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc_out(hidden_states)
+        return self.dropout(hidden_states)
+
+
+class NanoGPT2Attention(nn.Module):
+    def __init__(self, config: GPT2Config, layer_idx: int, attn_implementation: str) -> None:
+        super().__init__()
+        hidden_size = int(config.hidden_size)
+        num_heads = int(config.num_attention_heads)
+        if hidden_size % num_heads != 0:
+            raise ValueError(f"hidden_size={hidden_size} must be divisible by num_attention_heads={num_heads}")
+
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.embed_dim = hidden_size
+        self.layer_idx = layer_idx
+        self.attn_implementation = attn_implementation
+        self.attn_dropout = float(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.scale_attn_weights = bool(getattr(config, "scale_attn_weights", True))
+        self.scale_attn_by_inverse_layer_idx = bool(getattr(config, "scale_attn_by_inverse_layer_idx", False))
+        self.position_embedding_type = str(getattr(config, "position_embedding_type", "absolute")).lower()
+        if self.position_embedding_type not in {"absolute", "rope"}:
+            raise ValueError(f"Unsupported position_embedding_type={self.position_embedding_type!r}")
+
+        self.c_attn = nn.Linear(hidden_size, 3 * hidden_size)
+        self.c_proj = nn.Linear(hidden_size, hidden_size)
+        self.rotary_emb = None
+        if self.position_embedding_type == "rope":
+            self.rotary_emb = NanoGPT2RotaryEmbedding(
+                self.head_dim,
+                base=float(getattr(config, "rope_base", 10000.0)),
+            )
+
+    def _split_heads(self, tensor: torch.Tensor) -> torch.Tensor:
+        if tensor.ndim == 3:
+            batch_size, seq_len, _ = tensor.shape
+            return tensor.view(batch_size, seq_len, self.num_heads, self.head_dim)
+        if tensor.ndim == 2:
+            total_tokens, _ = tensor.shape
+            return tensor.view(total_tokens, self.num_heads, self.head_dim)
+        raise ValueError(f"Unsupported tensor rank for attention split: {tensor.ndim}")
+
+    def _merge_heads(self, tensor: torch.Tensor) -> torch.Tensor:
+        if tensor.ndim == 4:
+            batch_size, seq_len, _, _ = tensor.shape
+            return tensor.reshape(batch_size, seq_len, self.embed_dim)
+        if tensor.ndim == 3:
+            total_tokens, _, _ = tensor.shape
+            return tensor.reshape(total_tokens, self.embed_dim)
+        raise ValueError(f"Unsupported tensor rank for attention merge: {tensor.ndim}")
+
+    def _causal_attention_mask(
+        self,
+        attention_mask: Optional[torch.Tensor],
+        query_length: int,
+        key_length: int,
+        device: torch.device,
+    ) -> torch.Tensor:
+        query_positions = torch.arange(query_length, device=device, dtype=torch.long)
+        query_positions = query_positions + max(key_length - query_length, 0)
+        key_positions = torch.arange(key_length, device=device, dtype=torch.long)
+        causal = key_positions.unsqueeze(0) <= query_positions.unsqueeze(1)
+        causal = causal.unsqueeze(0).unsqueeze(0)
+        if attention_mask is None:
+            return causal
+        key_mask = attention_mask[:, None, None, :].to(dtype=torch.bool)
+        return causal & key_mask
+
+    def _eager_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        scale = 1.0
+        if self.scale_attn_weights:
+            scale /= self.head_dim ** 0.5
+        if self.scale_attn_by_inverse_layer_idx:
+            scale /= float(self.layer_idx + 1)
+
+        scores = torch.matmul(query, key.transpose(-1, -2)) * scale
+        causal_mask = self._causal_attention_mask(
+            attention_mask=attention_mask,
+            query_length=query.shape[-2],
+            key_length=key.shape[-2],
+            device=query.device,
+        )
+        scores = scores.masked_fill(~causal_mask, torch.finfo(scores.dtype).min)
+        probs = torch.softmax(scores, dim=-1)
+        if self.training and self.attn_dropout > 0:
+            probs = torch.dropout(probs, self.attn_dropout, train=True)
+        output = torch.matmul(probs, value)
+        return output.transpose(1, 2).contiguous()
+
+    def _sdpa_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        mask = None
+        if attention_mask is not None:
+            mask = self._causal_attention_mask(
+                attention_mask=attention_mask,
+                query_length=query.shape[-2],
+                key_length=key.shape[-2],
+                device=query.device,
+            )
+        output = torch.nn.functional.scaled_dot_product_attention(
+            query,
+            key,
+            value,
+            attn_mask=mask,
+            dropout_p=self.attn_dropout if self.training else 0.0,
+            is_causal=mask is None,
+        )
+        return output.transpose(1, 2).contiguous()
+
+    def _flash_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        packed_metadata: Optional[PackedSequenceMetadata],
+    ) -> torch.Tensor:
+        if not _FLASH_ATTN_AVAILABLE:
+            raise ImportError("flash_attn is not installed, but attn_implementation='flash_attention_2' was requested.")
+        if query.device.type != "cuda":
+            raise ValueError("flash_attention_2 requires CUDA tensors.")
+        if query.dtype not in (torch.float16, torch.bfloat16):
+            raise ValueError(
+                f"flash_attention_2 requires fp16/bf16 tensors, but received dtype={query.dtype}."
+            )
+
+        dropout_p = self.attn_dropout if self.training else 0.0
+        if packed_metadata is not None:
+            if packed_metadata.indices is not None:
+                query = query.reshape(-1, self.num_heads, self.head_dim).index_select(0, packed_metadata.indices)
+                key = key.reshape(-1, self.num_heads, self.head_dim).index_select(0, packed_metadata.indices)
+                value = value.reshape(-1, self.num_heads, self.head_dim).index_select(0, packed_metadata.indices)
+            output = flash_attn_varlen_func(
+                query,
+                key,
+                value,
+                packed_metadata.cu_seqlens,
+                packed_metadata.cu_seqlens,
+                packed_metadata.max_seqlen,
+                packed_metadata.max_seqlen,
+                dropout_p=dropout_p,
+                causal=True,
+            )
+            if packed_metadata.indices is None:
+                return output
+            return pad_input(
+                output,
+                packed_metadata.indices,
+                packed_metadata.batch_size,
+                packed_metadata.seq_len,
+            )
+
+        if attention_mask is None or bool(attention_mask.all()):
+            return flash_attn_func(
+                query,
+                key,
+                value,
+                dropout_p=dropout_p,
+                causal=True,
+            )
+
+        unpadded_query, indices, cu_seqlens, max_seqlen, _ = unpad_input(query, attention_mask)
+        unpadded_key, _, _, _, _ = unpad_input(key, attention_mask)
+        unpadded_value, _, _, _, _ = unpad_input(value, attention_mask)
+        output = flash_attn_varlen_func(
+            unpadded_query,
+            unpadded_key,
+            unpadded_value,
+            cu_seqlens,
+            cu_seqlens,
+            max_seqlen,
+            max_seqlen,
+            dropout_p=dropout_p,
+            causal=True,
+        )
+        return pad_input(output, indices, query.shape[0], query.shape[1])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        packed_metadata: Optional[PackedSequenceMetadata] = None,
+        layer_past: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        use_cache: bool = False,
+    ) -> tuple[torch.Tensor, Optional[tuple[torch.Tensor, torch.Tensor]]]:
+        qkv = self.c_attn(hidden_states)
+        query, key, value = qkv.split(self.embed_dim, dim=-1)
+        query = self._split_heads(query)
+        key = self._split_heads(key)
+        value = self._split_heads(value)
+
+        if self.rotary_emb is not None:
+            if position_ids is None:
+                raise ValueError("position_ids must be provided when position_embedding_type='rope'.")
+            cos, sin = self.rotary_emb(
+                position_ids.to(device=query.device),
+                device=query.device,
+                dtype=query.dtype,
+            )
+            query = apply_rotary_pos_emb(query, cos, sin)
+            key = apply_rotary_pos_emb(key, cos, sin)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            key = torch.cat([past_key.to(device=key.device, dtype=key.dtype), key], dim=1)
+            value = torch.cat([past_value.to(device=value.device, dtype=value.dtype), value], dim=1)
+
+        present = (key, value) if use_cache else None
+
+        if self.attn_implementation == "flash_attention_2" and layer_past is None:
+            attn_output = self._flash_attention(
+                query=query,
+                key=key,
+                value=value,
+                attention_mask=attention_mask,
+                packed_metadata=packed_metadata,
+            )
+        elif self.attn_implementation == "sdpa":
+            attn_output = self._sdpa_attention(
+                query=query,
+                key=key,
+                value=value,
+                attention_mask=attention_mask,
+            )
+        else:
+            attn_output = self._eager_attention(
+                query=query,
+                key=key,
+                value=value,
+                attention_mask=attention_mask,
+            )
+
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.c_proj(attn_output)
+        return self.resid_dropout(attn_output), present
+
+
+class NanoGPT2Block(nn.Module):
+    def __init__(self, config: GPT2Config, layer_idx: int, attn_implementation: str) -> None:
+        super().__init__()
+        hidden_size = int(config.hidden_size)
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = NanoGPT2Attention(config, layer_idx=layer_idx, attn_implementation=attn_implementation)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.mlp = NanoGPT2MLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        packed_metadata: Optional[PackedSequenceMetadata] = None,
+        layer_past: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        use_cache: bool = False,
+    ) -> tuple[torch.Tensor, Optional[tuple[torch.Tensor, torch.Tensor]]]:
+        attn_output, present = self.attn(
+            self.ln_1(hidden_states),
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            packed_metadata=packed_metadata,
+            layer_past=layer_past,
+            use_cache=use_cache,
+        )
+        hidden_states = hidden_states + attn_output
+        hidden_states = hidden_states + self.mlp(self.ln_2(hidden_states))
+        return hidden_states, present
+
+
+class NanoGPT2Model(nn.Module):
+    def __init__(self, config: GPT2Config, attn_implementation: str = "eager") -> None:
+        super().__init__()
+        self.config = config
+        self.attn_implementation = attn_implementation
+        self.position_embedding_type = str(getattr(config, "position_embedding_type", "absolute")).lower()
+        if self.position_embedding_type not in {"absolute", "rope"}:
+            raise ValueError(f"Unsupported position_embedding_type={self.position_embedding_type!r}")
+        hidden_size = int(config.hidden_size)
+        self.wte = nn.Embedding(config.vocab_size, hidden_size)
+        self.wpe = nn.Embedding(config.n_positions, hidden_size) if self.position_embedding_type == "absolute" else nn.Identity()
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList(
+            [NanoGPT2Block(config, layer_idx=index, attn_implementation=attn_implementation) for index in range(config.n_layer)]
+        )
+        self.ln_f = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.gradient_checkpointing = False
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        init_std = float(self.config.initializer_range)
+        for module in self.modules():
+            if isinstance(module, nn.Linear):
+                nn.init.normal_(module.weight, mean=0.0, std=init_std)
+                if module.bias is not None:
+                    nn.init.zeros_(module.bias)
+            elif isinstance(module, nn.Embedding):
+                nn.init.normal_(module.weight, mean=0.0, std=init_std)
+            elif isinstance(module, nn.LayerNorm):
+                nn.init.ones_(module.weight)
+                nn.init.zeros_(module.bias)
+
+    @staticmethod
+    def _normalize_num_sequences(
+        cu_seqlens: torch.Tensor,
+        num_sequences: Optional[torch.Tensor],
+        device: torch.device,
+    ) -> torch.Tensor:
+        if cu_seqlens.ndim == 1:
+            cu_seqlens = cu_seqlens.unsqueeze(0)
+        if num_sequences is None:
+            counts = []
+            for boundary in cu_seqlens:
+                diffs = boundary[1:] - boundary[:-1]
+                counts.append(int((diffs > 0).sum().item()))
+            return torch.tensor(counts, dtype=torch.int32, device=device)
+        if num_sequences.ndim == 0:
+            return num_sequences.unsqueeze(0)
+        return num_sequences
+
+    @staticmethod
+    def build_packed_position_ids(
+        attention_mask: Optional[torch.Tensor],
+        cu_seqlens: torch.Tensor,
+        num_sequences: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        if cu_seqlens.ndim == 1:
+            cu_seqlens = cu_seqlens.unsqueeze(0)
+        batch_size, seq_len = cu_seqlens.shape[0], cu_seqlens.shape[1] - 1
+        device = cu_seqlens.device
+        position_ids = torch.zeros((batch_size, seq_len), dtype=torch.long, device=device)
+        counts = NanoGPT2Model._normalize_num_sequences(cu_seqlens, num_sequences, device=device)
+        for batch_index in range(batch_size):
+            sequence_count = int(counts[batch_index].item())
+            boundaries = cu_seqlens[batch_index, : sequence_count + 1].tolist()
+            for start, end in zip(boundaries[:-1], boundaries[1:]):
+                start = int(start)
+                end = int(end)
+                if end > start:
+                    position_ids[batch_index, start:end] = torch.arange(end - start, device=device)
+        if attention_mask is not None:
+            position_ids = position_ids * attention_mask.to(dtype=position_ids.dtype)
+        return position_ids
+
+    @staticmethod
+    def build_packed_metadata(
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        num_sequences: Optional[torch.Tensor],
+    ) -> PackedSequenceMetadata:
+        if cu_seqlens.ndim == 1:
+            cu_seqlens = cu_seqlens.unsqueeze(0)
+        device = hidden_states.device
+        counts = NanoGPT2Model._normalize_num_sequences(cu_seqlens, num_sequences, device=device)
+        flat_indices = []
+        cumulative = [0]
+        max_seqlen = 0
+        seq_len = hidden_states.shape[1]
+
+        for batch_index in range(hidden_states.shape[0]):
+            sequence_count = int(counts[batch_index].item())
+            boundaries = cu_seqlens[batch_index, : sequence_count + 1].tolist()
+            for start, end in zip(boundaries[:-1], boundaries[1:]):
+                start = int(start)
+                end = int(end)
+                if end <= start:
+                    continue
+                segment_indices = batch_index * seq_len + torch.arange(start, end, device=device)
+                flat_indices.append(segment_indices)
+                cumulative.append(cumulative[-1] + (end - start))
+                max_seqlen = max(max_seqlen, end - start)
+
+        if not flat_indices:
+            raise ValueError("cu_seqlens did not describe any non-empty packed sequences.")
+
+        indices = torch.cat(flat_indices, dim=0)
+        return PackedSequenceMetadata(
+            cu_seqlens=torch.tensor(cumulative, dtype=torch.int32, device=device),
+            max_seqlen=max_seqlen,
+            indices=indices,
+            batch_size=hidden_states.shape[0],
+            seq_len=hidden_states.shape[1],
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[tuple[tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: bool = True,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        num_sequences: Optional[torch.Tensor] = None,
+    ) -> BaseModelOutputWithPast:
+        del input_ids, output_attentions
+
+        if inputs_embeds is None:
+            raise ValueError("inputs_embeds must be provided.")
+
+        use_cache = bool(use_cache)
+        if use_cache and cu_seqlens is not None:
+            raise ValueError("use_cache=True is not supported together with cu_seqlens packing.")
+
+        hidden_states = inputs_embeds
+        if attention_mask is None:
+            attention_mask = torch.ones(hidden_states.shape[:2], dtype=torch.bool, device=hidden_states.device)
+        else:
+            attention_mask = attention_mask.to(dtype=torch.bool, device=hidden_states.device)
+        query_attention_mask = attention_mask[:, -hidden_states.shape[1] :]
+
+        packed_metadata = None
+        if position_ids is None:
+            if cu_seqlens is not None:
+                position_ids = self.build_packed_position_ids(
+                    attention_mask=attention_mask,
+                    cu_seqlens=cu_seqlens.to(device=hidden_states.device),
+                    num_sequences=num_sequences.to(device=hidden_states.device) if num_sequences is not None else None,
+                )
+            elif attention_mask is not None:
+                position_ids = attention_mask.long().cumsum(dim=-1) - 1
+                position_ids = position_ids.masked_fill(~attention_mask, 0)
+                position_ids = position_ids[:, -hidden_states.shape[1] :]
+            else:
+                past_length = 0
+                if past_key_values is not None and len(past_key_values) > 0:
+                    past_length = past_key_values[0][0].shape[1]
+                position_ids = torch.arange(hidden_states.shape[1], device=hidden_states.device, dtype=torch.long)
+                position_ids = position_ids + past_length
+                position_ids = position_ids.unsqueeze(0).expand(hidden_states.shape[0], -1)
+
+        if cu_seqlens is not None and self.attn_implementation == "flash_attention_2":
+            packed_metadata = self.build_packed_metadata(
+                hidden_states=hidden_states,
+                cu_seqlens=cu_seqlens.to(device=hidden_states.device),
+                num_sequences=num_sequences.to(device=hidden_states.device) if num_sequences is not None else None,
+            )
+
+        if self.position_embedding_type == "absolute":
+            hidden_states = hidden_states + self.wpe(position_ids)
+        hidden_states = self.drop(hidden_states)
+        hidden_states = hidden_states * query_attention_mask.unsqueeze(-1).to(dtype=hidden_states.dtype)
+
+        all_hidden_states = () if output_hidden_states else None
+        presents = [] if use_cache else None
+        for layer_index, block in enumerate(self.h):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                if use_cache:
+                    raise ValueError("use_cache=True is not supported when gradient checkpointing is enabled during training.")
+
+                def custom_forward(*inputs):
+                    output, _ = block(
+                        inputs[0],
+                        attention_mask=inputs[1],
+                        position_ids=inputs[2],
+                        packed_metadata=packed_metadata,
+                        layer_past=None,
+                        use_cache=False,
+                    )
+                    return output
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    custom_forward,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    use_reentrant=False,
+                )
+                present = None
+            else:
+                hidden_states, present = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    packed_metadata=packed_metadata,
+                    layer_past=None if past_key_values is None else past_key_values[layer_index],
+                    use_cache=use_cache,
+                )
+            hidden_states = hidden_states * query_attention_mask.unsqueeze(-1).to(dtype=hidden_states.dtype)
+            if presents is not None:
+                presents.append(present)
+
+        hidden_states = self.ln_f(hidden_states)
+        hidden_states = hidden_states * query_attention_mask.unsqueeze(-1).to(dtype=hidden_states.dtype)
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return (hidden_states, tuple(presents) if presents is not None else None, all_hidden_states, None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=tuple(presents) if presents is not None else None,
+            hidden_states=all_hidden_states,
+            attentions=None,
+        )

modeling_nanotts_global_local.py

@@ -0,0 +1,1757 @@
+# coding=utf-8
+from __future__ import annotations
+
+import json
+import logging
+import os
+from contextlib import nullcontext
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torchaudio
+from transformers import AutoModel, AutoTokenizer
+from transformers.modeling_outputs import ModelOutput
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+from .configuration_nanotts import NanoTTSConfig
+from .gpt2_decoder import NanoGPT2Block, NanoGPT2Model
+from .prompting import (
+    build_assistant_prompt_prefix,
+    build_prompt_token_ids,
+    build_user_prompt_after_reference,
+    build_user_prompt_prefix,
+)
+from .tokenization_nanotts_sentencepiece import NanoTTSSentencePieceTokenizer
+
+
+@dataclass
+class NanoTTSOutput(ModelOutput):
+    global_hidden_states: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[tuple[tuple[torch.Tensor, torch.Tensor], ...]] = None
+    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+
+
+@dataclass
+class NanoTTSGenerationOutput(ModelOutput):
+    audio_token_ids: torch.LongTensor
+    prompt_input_ids: Optional[torch.LongTensor] = None
+
+
+MOSS_AUDIO_TOKENIZER_NANO_TYPE = "moss-audio-tokenizer-nano"
+DEFAULT_MOSS_AUDIO_TOKENIZER_PRETRAINED_NAME_OR_PATH = "OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano"
+DEFAULT_VOICE_CLONE_MAX_TEXT_TOKENS = 50
+DEFAULT_VOICE_CLONE_MAX_MEMORY_PER_SAMPLE_GB = 1.0
+DEFAULT_VOICE_CLONE_INTER_CHUNK_PAUSE_SHORT_SECONDS = 0.40
+DEFAULT_VOICE_CLONE_INTER_CHUNK_PAUSE_LONG_SECONDS = 0.24
+_SENTENCE_END_PUNCTUATION = frozenset(".!?。！？；;")
+_CLAUSE_SPLIT_PUNCTUATION = frozenset(",，、；;：:")
+_CLOSING_PUNCTUATION = frozenset("\"'”’)]}）】》」』")
+
+
+class NanoTTSPreTrainedModel(PreTrainedModel):
+    config_class = NanoTTSConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["NanoGPT2Block"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+
+
+class NanoTTSGlobalLocalForCausalLM(NanoTTSPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"local_transformer\.wte\.weight"]
+
+    def __init__(self, config: NanoTTSConfig) -> None:
+        super().__init__(config)
+        config.gpt2_config.pad_token_id = config.pad_token_id
+        config.gpt2_config._attn_implementation = config.attn_implementation
+
+        self.transformer = NanoGPT2Model(
+            config.gpt2_config,
+            attn_implementation=config.attn_implementation,
+        )
+        hidden_size = config.gpt2_config.hidden_size
+        init_std = config.gpt2_config.initializer_range
+
+        self.audio_embeddings = nn.ModuleList(
+            [
+                nn.Embedding(int(config.audio_codebook_sizes[index]), hidden_size)
+                for index in range(config.n_vq)
+            ]
+        )
+        self.text_lm_head = nn.Linear(hidden_size, config.gpt2_config.vocab_size, bias=False)
+        self.audio_lm_heads = nn.ModuleList(
+            [
+                nn.Linear(hidden_size, int(config.audio_codebook_sizes[index]), bias=False)
+                for index in range(config.n_vq)
+            ]
+        )
+
+        local_gpt2_config = config.gpt2_config.to_dict()
+        local_gpt2_config["n_layer"] = int(config.local_transformer_layers)
+        local_gpt2_config["n_positions"] = config.n_vq + 1
+        local_gpt2_config["n_ctx"] = config.n_vq + 1
+        self.local_transformer = NanoGPT2Model(
+            GPT2Config(**local_gpt2_config),
+            attn_implementation=str(config.local_transformer_attn_implementation),
+        )
+        self.local_transformer.wte = nn.Identity()
+
+        for module in list(self.audio_embeddings) + [self.text_lm_head] + list(self.audio_lm_heads):
+            if hasattr(module, "weight") and module.weight is not None:
+                nn.init.normal_(module.weight, mean=0.0, std=init_std)
+
+        self._tied_weights_keys = tuple(self.all_tied_weights_keys.keys())
+        self.tie_weights()
+
+    @property
+    def all_tied_weights_keys(self) -> dict[str, str]:
+        tied_weights = {"text_lm_head.weight": "transformer.wte.weight"}
+        tied_weights.update(
+            {
+                f"audio_lm_heads.{index}.weight": f"audio_embeddings.{index}.weight"
+                for index in range(self.config.n_vq)
+            }
+        )
+        return tied_weights
+
+    def tie_weights(self, *args, **kwargs) -> None:
+        del args, kwargs
+        self.text_lm_head.weight = self.transformer.wte.weight
+        for embedding, lm_head in zip(self.audio_embeddings, self.audio_lm_heads):
+            lm_head.weight = embedding.weight
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.transformer.wte
+
+    def set_input_embeddings(self, value: nn.Embedding) -> None:
+        self.transformer.wte = value
+        self.tie_weights()
+
+    def _build_inputs_embeds(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
+        if input_ids.ndim != 3 or input_ids.shape[-1] != self.config.n_vq + 1:
+            raise ValueError(
+                f"Expected input_ids shape [batch, seq, {self.config.n_vq + 1}], got {tuple(input_ids.shape)}"
+            )
+
+        text_ids = input_ids[..., 0]
+        inputs_embeds = self.transformer.wte(text_ids)
+
+        for channel_index, embedding in enumerate(self.audio_embeddings):
+            channel_ids = input_ids[..., channel_index + 1]
+            valid_mask = channel_ids.ne(self.config.audio_pad_token_id)
+            invalid_mask = valid_mask & ((channel_ids < 0) | (channel_ids >= embedding.num_embeddings))
+            if invalid_mask.any():
+                invalid_token_ids = channel_ids[invalid_mask]
+                raise ValueError(
+                    "Found out-of-range audio token ids for channel "
+                    f"{channel_index}: min={int(invalid_token_ids.min().item())} "
+                    f"max={int(invalid_token_ids.max().item())} "
+                    f"codebook_size={embedding.num_embeddings} "
+                    f"audio_pad_token_id={self.config.audio_pad_token_id}"
+                )
+            safe_ids = channel_ids.masked_fill(~valid_mask, 0)
+            audio_embeds = embedding(safe_ids)
+            audio_embeds = audio_embeds * valid_mask.unsqueeze(-1)
+            inputs_embeds = inputs_embeds + audio_embeds
+
+        return inputs_embeds
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[tuple[tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ):
+        labels = kwargs.pop("labels", None)
+        if labels is not None:
+            raise NotImplementedError("This open-source package is inference-only and does not support training forward.")
+        if kwargs:
+            ignored = ", ".join(sorted(kwargs.keys()))
+            logging.debug("ignoring unsupported forward kwargs: %s", ignored)
+
+        return_dict = self.config.use_return_dict if return_dict is None else return_dict
+        if inputs_embeds is None:
+            if input_ids is None:
+                raise ValueError("Either input_ids or inputs_embeds must be provided.")
+            inputs_embeds = self._build_inputs_embeds(input_ids)
+
+        outputs = self.transformer(
+            input_ids=None,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            position_ids=None,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cu_seqlens=None,
+            num_sequences=None,
+        )
+
+        if not return_dict:
+            return (
+                outputs.last_hidden_state,
+                outputs.past_key_values,
+                outputs.hidden_states,
+                outputs.attentions,
+            )
+
+        return NanoTTSOutput(
+            global_hidden_states=outputs.last_hidden_state,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _build_text_rows(
+        self,
+        token_ids: list[int],
+        device: torch.device,
+    ) -> torch.LongTensor:
+        rows = torch.full(
+            (len(token_ids), self.config.n_vq + 1),
+            self.config.audio_pad_token_id,
+            dtype=torch.long,
+            device=device,
+        )
+        if token_ids:
+            rows[:, 0] = torch.tensor(token_ids, dtype=torch.long, device=device)
+        return rows
+
+    def _encode_text(self, tokenizer, text: str) -> list[int]:
+        try:
+            return list(tokenizer.encode(text, add_special_tokens=False))
+        except TypeError:
+            return list(tokenizer.encode(text))
+
+    @staticmethod
+    def _contains_cjk(text: str) -> bool:
+        return any(
+            "\u4e00" <= ch <= "\u9fff"
+            or "\u3400" <= ch <= "\u4dbf"
+            or "\u3040" <= ch <= "\u30ff"
+            or "\uac00" <= ch <= "\ud7af"
+            for ch in str(text)
+        )
+
+    @staticmethod
+    def _prepare_text_for_sentence_chunking(text: str) -> str:
+        normalized_text = str(text).strip()
+        if normalized_text == "":
+            raise ValueError("Text prompt cannot be empty.")
+
+        normalized_text = normalized_text.replace("\n", " ").replace("\r", " ")
+        while "  " in normalized_text:
+            normalized_text = normalized_text.replace("  ", " ")
+
+        contains_cjk = NanoTTSGlobalLocalForCausalLM._contains_cjk(normalized_text)
+        if contains_cjk:
+            if normalized_text[-1] not in _SENTENCE_END_PUNCTUATION:
+                normalized_text = normalized_text + "。"
+            return normalized_text
+
+        if not normalized_text[0].isupper():
+            normalized_text = normalized_text[0].upper() + normalized_text[1:]
+        if normalized_text[-1].isalnum():
+            normalized_text = normalized_text + "."
+        if len(normalized_text.split()) < 5:
+            normalized_text = " " * 8 + normalized_text
+        return normalized_text
+
+    @staticmethod
+    def _split_text_by_punctuation(text: str, punctuation: set[str] | frozenset[str]) -> list[str]:
+        sentences: list[str] = []
+        current_chars: list[str] = []
+        text = str(text)
+        index = 0
+        while index < len(text):
+            char = text[index]
+            current_chars.append(char)
+            if char in punctuation:
+                lookahead = index + 1
+                while lookahead < len(text) and text[lookahead] in _CLOSING_PUNCTUATION:
+                    current_chars.append(text[lookahead])
+                    lookahead += 1
+                sentence = "".join(current_chars).strip()
+                if sentence:
+                    sentences.append(sentence)
+                current_chars = []
+                while lookahead < len(text) and text[lookahead].isspace():
+                    lookahead += 1
+                index = lookahead
+                continue
+            index += 1
+
+        tail = "".join(current_chars).strip()
+        if tail:
+            sentences.append(tail)
+        return sentences
+
+    def _count_text_tokens(self, text_tokenizer, text: str) -> int:
+        return len(self._encode_text(text_tokenizer, text))
+
+    def _split_text_by_token_budget(
+        self,
+        text_tokenizer,
+        text: str,
+        max_tokens: int,
+    ) -> list[str]:
+        remaining_text = str(text).strip()
+        if remaining_text == "":
+            return []
+
+        pieces: list[str] = []
+        preferred_boundary_chars = _CLAUSE_SPLIT_PUNCTUATION | _SENTENCE_END_PUNCTUATION | frozenset({" "})
+        while remaining_text:
+            if self._count_text_tokens(text_tokenizer, remaining_text) <= int(max_tokens):
+                pieces.append(remaining_text)
+                break
+
+            low = 1
+            high = len(remaining_text)
+            best_prefix_length = 1
+            while low <= high:
+                middle = (low + high) // 2
+                candidate = remaining_text[:middle].strip()
+                if not candidate:
+                    low = middle + 1
+                    continue
+                if self._count_text_tokens(text_tokenizer, candidate) <= int(max_tokens):
+                    best_prefix_length = middle
+                    low = middle + 1
+                else:
+                    high = middle - 1
+
+            cut_index = best_prefix_length
+            prefix = remaining_text[:best_prefix_length]
+            preferred_index = -1
+            for scan_index in range(len(prefix) - 1, max(-1, len(prefix) - 25), -1):
+                if prefix[scan_index] in preferred_boundary_chars:
+                    preferred_index = scan_index + 1
+                    break
+            if preferred_index > 0:
+                cut_index = preferred_index
+
+            piece = remaining_text[:cut_index].strip()
+            if not piece:
+                piece = remaining_text[:best_prefix_length].strip()
+                cut_index = best_prefix_length
+            pieces.append(piece)
+            remaining_text = remaining_text[cut_index:].strip()
+        return pieces
+
+    @staticmethod
+    def _join_sentence_parts(left: str, right: str) -> str:
+        if not left:
+            return right
+        if not right:
+            return left
+        if NanoTTSGlobalLocalForCausalLM._contains_cjk(left) or NanoTTSGlobalLocalForCausalLM._contains_cjk(right):
+            return left + right
+        return left + " " + right
+
+    def _split_text_into_best_sentences(
+        self,
+        text_tokenizer,
+        text: str,
+        max_tokens: int,
+    ) -> list[str]:
+        if int(max_tokens) <= 0:
+            return [str(text)]
+
+        prepared_text = self._prepare_text_for_sentence_chunking(text)
+        sentence_candidates = self._split_text_by_punctuation(prepared_text, punctuation=_SENTENCE_END_PUNCTUATION)
+        if not sentence_candidates:
+            sentence_candidates = [prepared_text.strip()]
+
+        sentence_slices: list[tuple[int, str]] = []
+        for sentence_text in sentence_candidates:
+            normalized_sentence = sentence_text.strip()
+            if not normalized_sentence:
+                continue
+            sentence_token_count = self._count_text_tokens(text_tokenizer, normalized_sentence)
+            if sentence_token_count <= int(max_tokens):
+                sentence_slices.append((sentence_token_count, normalized_sentence))
+                continue
+
+            clause_candidates = self._split_text_by_punctuation(
+                normalized_sentence,
+                punctuation=_CLAUSE_SPLIT_PUNCTUATION,
+            )
+            if len(clause_candidates) <= 1:
+                clause_candidates = [normalized_sentence]
+
+            for clause_text in clause_candidates:
+                normalized_clause = clause_text.strip()
+                if not normalized_clause:
+                    continue
+                clause_token_count = self._count_text_tokens(text_tokenizer, normalized_clause)
+                if clause_token_count <= int(max_tokens):
+                    sentence_slices.append((clause_token_count, normalized_clause))
+                    continue
+                for piece in self._split_text_by_token_budget(
+                    text_tokenizer=text_tokenizer,
+                    text=normalized_clause,
+                    max_tokens=max_tokens,
+                ):
+                    normalized_piece = piece.strip()
+                    if normalized_piece:
+                        sentence_slices.append(
+                            (self._count_text_tokens(text_tokenizer, normalized_piece), normalized_piece)
+                        )
+
+        chunks: list[str] = []
+        current_chunk = ""
+        current_chunk_token_count = 0
+        for sentence_token_count, sentence_text in sentence_slices:
+            if current_chunk == "":
+                current_chunk = sentence_text
+                current_chunk_token_count = sentence_token_count
+                continue
+            if current_chunk_token_count + sentence_token_count > int(max_tokens):
+                chunks.append(current_chunk.strip())
+                current_chunk = sentence_text
+                current_chunk_token_count = sentence_token_count
+            else:
+                current_chunk = self._join_sentence_parts(current_chunk, sentence_text)
+                current_chunk_token_count = self._count_text_tokens(text_tokenizer, current_chunk)
+
+        if current_chunk:
+            chunks.append(current_chunk.strip())
+        return chunks or [prepared_text.strip()]
+
+    @staticmethod
+    def _estimate_voice_clone_inter_chunk_pause_seconds(text_chunk: str) -> float:
+        return (
+            DEFAULT_VOICE_CLONE_INTER_CHUNK_PAUSE_SHORT_SECONDS
+            if len(str(text_chunk).strip().split()) <= 4
+            else DEFAULT_VOICE_CLONE_INTER_CHUNK_PAUSE_LONG_SECONDS
+        )
+
+    def _concat_voice_clone_waveform_chunks(
+        self,
+        waveform_chunks: list[torch.FloatTensor],
+        text_chunks: list[str],
+        sample_rate: int,
+    ) -> torch.FloatTensor:
+        if not waveform_chunks:
+            return torch.zeros((1, 0), dtype=torch.float32)
+        if len(waveform_chunks) != len(text_chunks):
+            raise ValueError("waveform_chunks and text_chunks must have the same length.")
+        if len(waveform_chunks) == 1:
+            return waveform_chunks[0]
+
+        segments: list[torch.FloatTensor] = []
+        for chunk_index, waveform_chunk in enumerate(waveform_chunks):
+            segments.append(waveform_chunk)
+            if chunk_index >= len(waveform_chunks) - 1:
+                continue
+            pause_seconds = self._estimate_voice_clone_inter_chunk_pause_seconds(text_chunks[chunk_index])
+            pause_samples = max(0, int(round(float(sample_rate) * pause_seconds)))
+            if pause_samples > 0:
+                silence = torch.zeros((waveform_chunk.shape[0], pause_samples), dtype=waveform_chunk.dtype)
+                segments.append(silence)
+        return torch.cat(segments, dim=-1)
+
+    @staticmethod
+    def _resolve_inference_mode(
+        mode: str,
+        has_prompt_text: bool,
+        has_prompt_audio: bool,
+    ) -> str:
+        normalized_mode = str(mode or "continuation").strip().lower() or "continuation"
+        if normalized_mode not in {"continuation", "voice_clone"}:
+            raise ValueError(f"Unsupported inference mode {mode!r}.")
+        if normalized_mode == "voice_clone":
+            if not has_prompt_audio:
+                raise ValueError("voice_clone mode requires prompt_audio_path.")
+            if has_prompt_text:
+                raise ValueError("voice_clone mode does not accept prompt_text.")
+        elif has_prompt_text != has_prompt_audio:
+            raise ValueError(
+                "continuation mode accepts either target text only, or prompt_text and prompt_audio_path together."
+            )
+        return normalized_mode
+
+    def _resolve_inference_nq(self, nq: Optional[int] = None) -> int:
+        if nq is None:
+            return int(self.config.n_vq)
+        resolved_nq = int(nq)
+        if resolved_nq < 1 or resolved_nq > int(self.config.n_vq):
+            raise ValueError(f"nq must be in [1, {self.config.n_vq}], got {resolved_nq}.")
+        return resolved_nq
+
+    def _mask_unused_audio_channels(
+        self,
+        audio_token_ids: torch.LongTensor,
+        nq: int,
+    ) -> torch.LongTensor:
+        tensor = torch.as_tensor(audio_token_ids, dtype=torch.long)
+        if tensor.shape[-1] != self.config.n_vq:
+            raise ValueError(
+                f"Expected audio token ids with trailing dim {self.config.n_vq}, got {tuple(tensor.shape)}"
+            )
+        if nq < self.config.n_vq:
+            tensor = tensor.clone()
+            tensor[..., nq:] = self.config.audio_pad_token_id
+        return tensor
+
+    def _build_audio_prefix_rows(
+        self,
+        prompt_audio_codes: torch.LongTensor,
+        slot_token_id: int,
+        device: torch.device,
+    ) -> torch.LongTensor:
+        rows = torch.full(
+            (int(prompt_audio_codes.shape[0]), self.config.n_vq + 1),
+            self.config.audio_pad_token_id,
+            dtype=torch.long,
+            device=device,
+        )
+        if rows.shape[0] > 0:
+            rows[:, 0] = int(slot_token_id)
+            rows[:, 1:] = prompt_audio_codes
+        return rows
+
+    def build_inference_input_ids(
+        self,
+        text: str,
+        text_tokenizer,
+        mode: str = "continuation",
+        prompt_text: Optional[str] = None,
+        prompt_audio_codes: Optional[torch.LongTensor] = None,
+        device: Optional[Union[str, torch.device]] = None,
+    ) -> tuple[torch.LongTensor, torch.BoolTensor]:
+        resolved_device = self._resolve_device(device)
+        resolved_mode = self._resolve_inference_mode(
+            mode=mode,
+            has_prompt_text=prompt_text is not None,
+            has_prompt_audio=prompt_audio_codes is not None,
+        )
+
+        if resolved_mode == "voice_clone":
+            assert prompt_audio_codes is not None
+            text_token_ids = self._encode_text(text_tokenizer, text)
+            prompt_token_ids = build_user_prompt_prefix(text_tokenizer, self.config) + [self.config.audio_start_token_id]
+            suffix_token_ids = (
+                [self.config.audio_end_token_id]
+                + build_user_prompt_after_reference(text_tokenizer)
+                + text_token_ids
+                + build_assistant_prompt_prefix(text_tokenizer, self.config)
+                + [self.config.audio_start_token_id]
+            )
+            sections = [
+                self._build_text_rows(prompt_token_ids, device=resolved_device),
+                self._build_audio_prefix_rows(
+                    prompt_audio_codes=prompt_audio_codes.to(resolved_device),
+                    slot_token_id=self.config.audio_user_slot_token_id,
+                    device=resolved_device,
+                ),
+                self._build_text_rows(suffix_token_ids, device=resolved_device),
+            ]
+            input_ids = torch.cat(sections, dim=0).unsqueeze(0)
+            attention_mask = torch.ones(input_ids.shape[:2], dtype=torch.bool, device=resolved_device)
+            return input_ids, attention_mask
+
+        effective_text = text if prompt_text is None else prompt_text + text
+        prompt_token_ids = build_prompt_token_ids(
+            tokenizer=text_tokenizer,
+            config=self.config,
+            text_token_ids=self._encode_text(text_tokenizer, effective_text),
+        )
+        sections = [
+            self._build_text_rows(prompt_token_ids, device=resolved_device),
+            self._build_text_rows([self.config.audio_start_token_id], device=resolved_device),
+        ]
+        if prompt_audio_codes is not None:
+            sections.append(
+                self._build_audio_prefix_rows(
+                    prompt_audio_codes=prompt_audio_codes.to(resolved_device),
+                    slot_token_id=self.config.audio_assistant_slot_token_id,
+                    device=resolved_device,
+                )
+            )
+        input_ids = torch.cat(sections, dim=0).unsqueeze(0)
+        attention_mask = torch.ones(input_ids.shape[:2], dtype=torch.bool, device=resolved_device)
+        return input_ids, attention_mask
+
+    def _left_pad_inference_batch(
+        self,
+        input_id_batches: list[torch.LongTensor],
+        attention_mask_batches: list[torch.BoolTensor],
+        device: torch.device,
+    ) -> tuple[torch.LongTensor, torch.BoolTensor]:
+        if not input_id_batches:
+            raise ValueError("input_id_batches must not be empty.")
+        if len(input_id_batches) != len(attention_mask_batches):
+            raise ValueError("input_id_batches and attention_mask_batches must have the same length.")
+
+        batch_size = len(input_id_batches)
+        max_seq_len = max(int(batch.shape[1]) for batch in input_id_batches)
+        row_width = self.config.n_vq + 1
+
+        padded_input_ids = torch.full(
+            (batch_size, max_seq_len, row_width),
+            self.config.audio_pad_token_id,
+            dtype=torch.long,
+            device=device,
+        )
+        padded_input_ids[:, :, 0] = self.config.pad_token_id
+        padded_attention_mask = torch.zeros((batch_size, max_seq_len), dtype=torch.bool, device=device)
+
+        for batch_index, (input_ids, attention_mask) in enumerate(zip(input_id_batches, attention_mask_batches)):
+            normalized_input_ids = input_ids.squeeze(0).to(device=device, dtype=torch.long)
+            normalized_attention_mask = attention_mask.squeeze(0).to(device=device, dtype=torch.bool)
+            seq_len = int(normalized_input_ids.shape[0])
+            padded_input_ids[batch_index, -seq_len:, :] = normalized_input_ids
+            padded_attention_mask[batch_index, -seq_len:] = normalized_attention_mask
+
+        return padded_input_ids, padded_attention_mask
+
+    def _trim_generated_audio_token_ids(
+        self,
+        audio_token_ids: torch.LongTensor,
+        effective_nq: int,
+    ) -> torch.LongTensor:
+        tensor = self._mask_unused_audio_channels(audio_token_ids, nq=effective_nq)
+        if tensor.ndim != 2:
+            raise ValueError(f"Expected a 2D audio token tensor, got {tuple(tensor.shape)}")
+        valid_rows = tensor[:, :effective_nq].ne(self.config.audio_pad_token_id).any(dim=-1)
+        if not bool(valid_rows.any()):
+            return tensor[:0]
+        last_valid_index = int(torch.nonzero(valid_rows, as_tuple=False)[-1].item()) + 1
+        return tensor[:last_valid_index]
+
+    def _resolve_voice_clone_chunk_batch_size(
+        self,
+        *,
+        resolved_device: torch.device,
+        chunk_count: int,
+        max_memory_per_sample_gb: float,
+    ) -> int:
+        if chunk_count <= 1 or max_memory_per_sample_gb <= 0 or resolved_device.type != "cuda":
+            return 1
+        if not hasattr(torch.cuda, "mem_get_info"):
+            return 1
+        try:
+            free_bytes, _ = torch.cuda.mem_get_info(resolved_device)
+        except Exception:
+            return 1
+        bytes_per_sample = int(float(max_memory_per_sample_gb) * (1024**3))
+        if bytes_per_sample <= 0:
+            return 1
+        usable_free_bytes = max(0, int(free_bytes * 0.9))
+        batch_size = max(1, usable_free_bytes // bytes_per_sample)
+        resolved_batch_size = max(1, min(int(chunk_count), int(batch_size)))
+        logging.info(
+            "voice_clone chunk batching device=%s free_gb=%.2f max_memory_per_sample_gb=%.2f resolved_batch_size=%d chunk_count=%d",
+            resolved_device,
+            float(free_bytes) / float(1024**3),
+            float(max_memory_per_sample_gb),
+            resolved_batch_size,
+            int(chunk_count),
+        )
+        return resolved_batch_size
+
+    def _generate_audio_token_ids_with_fallback(
+        self,
+        *,
+        prompt_input_ids: torch.LongTensor,
+        attention_mask: torch.BoolTensor,
+        effective_nq: int,
+        max_new_frames: int,
+        do_sample: bool,
+        text_temperature: float,
+        text_top_p: float,
+        text_top_k: int,
+        audio_temperature: float,
+        audio_top_p: float,
+        audio_top_k: int,
+        audio_repetition_penalty: float,
+        use_kv_cache: bool,
+        resolved_device: torch.device,
+    ) -> torch.LongTensor:
+        try:
+            generation = self.generate(
+                input_ids=prompt_input_ids,
+                attention_mask=attention_mask,
+                nq=effective_nq,
+                max_new_frames=max_new_frames,
+                do_sample=do_sample,
+                text_temperature=text_temperature,
+                text_top_p=text_top_p,
+                text_top_k=text_top_k,
+                audio_temperature=audio_temperature,
+                audio_top_p=audio_top_p,
+                audio_top_k=audio_top_k,
+                audio_repetition_penalty=audio_repetition_penalty,
+                use_kv_cache=use_kv_cache,
+                return_dict_in_generate=True,
+            )
+        except (RuntimeError, ValueError) as exc:
+            if not self._is_generation_stability_error(exc):
+                raise
+            self._apply_inference_stability_fallback(resolved_device)
+            generation = self.generate(
+                input_ids=prompt_input_ids,
+                attention_mask=attention_mask,
+                nq=effective_nq,
+                max_new_frames=max_new_frames,
+                do_sample=do_sample,
+                text_temperature=text_temperature,
+                text_top_p=text_top_p,
+                text_top_k=text_top_k,
+                audio_temperature=audio_temperature,
+                audio_top_p=audio_top_p,
+                audio_top_k=audio_top_k,
+                audio_repetition_penalty=audio_repetition_penalty,
+                use_kv_cache=use_kv_cache,
+                return_dict_in_generate=True,
+            )
+        return self._mask_unused_audio_channels(generation.audio_token_ids, nq=effective_nq)
+
+    def _decode_audio_token_ids_to_waveform(
+        self,
+        *,
+        audio_tokenizer,
+        audio_token_ids: torch.LongTensor,
+        target_sample_rate: int,
+        effective_nq: int,
+        resolved_device: torch.device,
+    ) -> tuple[torch.FloatTensor, int]:
+        decoded = self._call_audio_decode(
+            audio_tokenizer=audio_tokenizer,
+            audio_token_ids=audio_token_ids.to(resolved_device),
+            sample_rate=target_sample_rate,
+            nq=effective_nq,
+        )
+        return self._extract_waveform_and_sample_rate(decoded, fallback_sample_rate=target_sample_rate)
+
+    def _build_generation_row(
+        self,
+        batch_size: int,
+        device: torch.device,
+        audio_token_ids: torch.LongTensor,
+    ) -> torch.LongTensor:
+        row = torch.full(
+            (batch_size, 1, self.config.n_vq + 1),
+            self.config.audio_pad_token_id,
+            dtype=torch.long,
+            device=device,
+        )
+        row[:, :, 0] = self.config.audio_assistant_slot_token_id
+        row[:, :, 1:] = audio_token_ids.unsqueeze(1)
+        return row
+
+    def _sample_next_token(
+        self,
+        logits: torch.FloatTensor,
+        do_sample: bool,
+        temperature: float,
+        top_k: Optional[int],
+        top_p: Optional[float],
+        previous_token_ids: Optional[torch.LongTensor] = None,
+        repetition_penalty: float = 1.0,
+    ) -> torch.LongTensor:
+        scores = self._apply_repetition_penalty(
+            logits=logits,
+            previous_token_ids=previous_token_ids,
+            repetition_penalty=repetition_penalty,
+        )
+        if not do_sample:
+            return scores.argmax(dim=-1)
+        if temperature <= 0:
+            raise ValueError("temperature must be positive when do_sample=True")
+
+        scores = scores / temperature
+        if top_k is not None and top_k > 0:
+            top_k = min(top_k, scores.shape[-1])
+            threshold = torch.topk(scores, top_k, dim=-1).values[..., -1, None]
+            scores = scores.masked_fill(scores < threshold, float("-inf"))
+
+        if top_p is not None and 0.0 < top_p < 1.0:
+            sorted_scores, sorted_indices = torch.sort(scores, descending=True, dim=-1)
+            sorted_probs = torch.softmax(sorted_scores, dim=-1)
+            sorted_cumsum = torch.cumsum(sorted_probs, dim=-1)
+            sorted_remove = sorted_cumsum > top_p
+            sorted_remove[..., 1:] = sorted_remove[..., :-1].clone()
+            sorted_remove[..., 0] = False
+            sorted_scores = sorted_scores.masked_fill(sorted_remove, float("-inf"))
+            scores = torch.full_like(scores, float("-inf"))
+            scores.scatter_(dim=-1, index=sorted_indices, src=sorted_scores)
+
+        probs = torch.softmax(scores, dim=-1)
+        return torch.multinomial(probs, num_samples=1).squeeze(-1)
+
+    @staticmethod
+    def _ensure_finite_generation_logits(logits: torch.FloatTensor, name: str) -> None:
+        if torch.isfinite(logits).all():
+            return
+        finite_mask = torch.isfinite(logits)
+        finite_logits = logits[finite_mask]
+        min_value = float(finite_logits.min().item()) if finite_logits.numel() > 0 else float("nan")
+        max_value = float(finite_logits.max().item()) if finite_logits.numel() > 0 else float("nan")
+        raise RuntimeError(
+            f"Non-finite {name} during generation: dtype={logits.dtype} "
+            f"shape={tuple(logits.shape)} finite={int(finite_mask.sum().item())}/{int(logits.numel())} "
+            f"min={min_value} max={max_value}"
+        )
+
+    def _apply_repetition_penalty(
+        self,
+        logits: torch.FloatTensor,
+        previous_token_ids: Optional[torch.LongTensor],
+        repetition_penalty: float,
+    ) -> torch.FloatTensor:
+        if repetition_penalty <= 0:
+            raise ValueError("repetition_penalty must be positive")
+        if repetition_penalty == 1.0 or previous_token_ids is None:
+            return logits
+
+        token_ids = torch.as_tensor(previous_token_ids, device=logits.device, dtype=torch.long)
+        if token_ids.ndim == 1:
+            token_ids = token_ids.unsqueeze(0)
+        elif token_ids.ndim > 2:
+            token_ids = token_ids.reshape(token_ids.shape[0], -1)
+
+        scores = logits.clone()
+        vocab_size = scores.shape[-1]
+        for batch_index in range(scores.shape[0]):
+            valid_token_ids = token_ids[batch_index]
+            valid_token_ids = valid_token_ids[(valid_token_ids >= 0) & (valid_token_ids < vocab_size)]
+            if valid_token_ids.numel() == 0:
+                continue
+            unique_token_ids = torch.unique(valid_token_ids)
+            token_scores = scores[batch_index].index_select(0, unique_token_ids)
+            token_scores = torch.where(
+                token_scores < 0,
+                token_scores * repetition_penalty,
+                token_scores / repetition_penalty,
+            )
+            scores[batch_index].scatter_(0, unique_token_ids, token_scores)
+        return scores
+
+    def _sample_next_assistant_text_token(
+        self,
+        logits: torch.FloatTensor,
+        do_sample: bool,
+        temperature: float,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+    ) -> torch.LongTensor:
+        candidate_ids = torch.tensor(
+            [
+                self.config.audio_assistant_slot_token_id,
+                self.config.audio_end_token_id,
+            ],
+            dtype=torch.long,
+            device=logits.device,
+        )
+        candidate_logits = logits.index_select(dim=-1, index=candidate_ids)
+        sampled_indices = self._sample_next_token(
+            logits=candidate_logits,
+            do_sample=do_sample,
+            temperature=temperature,
+            top_k=top_k,
+            top_p=top_p,
+        )
+        return candidate_ids[sampled_indices]
+
+    def _resolve_device(self, device: Optional[Union[str, torch.device]] = None) -> torch.device:
+        return torch.device(device) if device is not None else next(self.parameters()).device
+
+    @staticmethod
+    def _looks_like_hf_tokenizer_dir(candidate_path: Path) -> bool:
+        if not candidate_path.is_dir():
+            return False
+        if (candidate_path / "tokenizer.model").is_file():
+            return True
+        if (candidate_path / "tokenizer.json").is_file():
+            return True
+        if (candidate_path / "tokenizer_config.json").is_file() and (
+            (candidate_path / "vocab.json").is_file()
+            or (candidate_path / "merges.txt").is_file()
+            or (candidate_path / "special_tokens_map.json").is_file()
+        ):
+            return True
+        return False
+
+    def _resolve_text_tokenizer_path(self, raw_path: Union[str, Path]) -> Path:
+        candidate_path = Path(raw_path)
+        if candidate_path.is_file() and candidate_path.suffix == ".model":
+            return candidate_path
+        if not candidate_path.exists():
+            raise FileNotFoundError(f"Tokenizer path does not exist: {candidate_path}")
+        if candidate_path.is_dir():
+            if (candidate_path / "tokenizer.model").is_file():
+                return candidate_path
+            if self._looks_like_hf_tokenizer_dir(candidate_path):
+                return candidate_path
+            hf_dir = candidate_path / "hf_tokenizer"
+            if self._looks_like_hf_tokenizer_dir(hf_dir):
+                return hf_dir
+            sentencepiece_model = candidate_path / "sentencepiece" / "nanotts_spm_bpe.model"
+            if sentencepiece_model.is_file():
+                return sentencepiece_model
+            final_summary_path = candidate_path / "final_summary.json"
+            if final_summary_path.is_file():
+                final_summary = json.loads(final_summary_path.read_text(encoding="utf-8"))
+                latest_hf_dir = final_summary.get("latest_hf_tokenizer_dir")
+                if latest_hf_dir:
+                    latest_hf_path = Path(str(latest_hf_dir))
+                    if self._looks_like_hf_tokenizer_dir(latest_hf_path):
+                        return latest_hf_path
+        raise ValueError(
+            "Could not resolve a tokenizer from the provided path. Expected a tokenizer dir, experiment dir, or SentencePiece .model file."
+        )
+
+    def _load_resolved_text_tokenizer(self, resolved_path: Path, cache_dir: str):
+        if resolved_path.is_file() and resolved_path.suffix == ".model":
+            return NanoTTSSentencePieceTokenizer(vocab_file=str(resolved_path))
+        try:
+            return AutoTokenizer.from_pretrained(
+                str(resolved_path),
+                trust_remote_code=True,
+                use_fast=bool(self.config.tokenizer_use_fast),
+                local_files_only=True,
+                cache_dir=cache_dir,
+            )
+        except Exception:
+            model_path = resolved_path / "tokenizer.model"
+            if model_path.is_file():
+                return NanoTTSSentencePieceTokenizer(vocab_file=str(model_path))
+            raise
+
+    @staticmethod
+    def _resolve_hf_cache_dir() -> str:
+        cache_dir = Path(__file__).resolve().parent / ".cache" / "huggingface"
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        return str(cache_dir)
+
+    @staticmethod
+    def _patch_hf_dynamic_module_cache_dir(cache_dir: str) -> None:
+        import transformers.dynamic_module_utils as dynamic_module_utils
+
+        modules_cache_dir = str(Path(cache_dir) / "modules")
+        Path(modules_cache_dir).mkdir(parents=True, exist_ok=True)
+        os.environ["HF_MODULES_CACHE"] = modules_cache_dir
+        dynamic_module_utils.HF_MODULES_CACHE = modules_cache_dir
+
+    def _resolve_default_text_tokenizer_path(self) -> Path:
+        candidates: list[Path] = []
+
+        raw_name_or_path = getattr(self.config, "_name_or_path", None)
+        if raw_name_or_path:
+            candidates.append(Path(str(raw_name_or_path)).expanduser())
+
+        raw_model_name_or_path = getattr(self, "name_or_path", None)
+        if raw_model_name_or_path:
+            candidates.append(Path(str(raw_model_name_or_path)).expanduser())
+
+        candidates.append(Path(__file__).resolve().parent)
+
+        checked: set[str] = set()
+        for candidate in candidates:
+            resolved_candidate = candidate.resolve()
+            key = str(resolved_candidate)
+            if key in checked:
+                continue
+            checked.add(key)
+
+            if (resolved_candidate / "tokenizer.model").is_file():
+                return resolved_candidate
+            if self._looks_like_hf_tokenizer_dir(resolved_candidate):
+                return resolved_candidate
+
+        return candidates[0].resolve()
+
+    def _load_text_tokenizer(self, text_tokenizer=None, text_tokenizer_path: Optional[str] = None):
+        if text_tokenizer is not None:
+            return text_tokenizer
+
+        resolved_path = (
+            self._resolve_text_tokenizer_path(text_tokenizer_path)
+            if text_tokenizer_path is not None
+            else self._resolve_default_text_tokenizer_path()
+        )
+        normalized_path = str(resolved_path.resolve())
+        cached = getattr(self, "_cached_text_tokenizer", None)
+        cached_path = getattr(self, "_cached_text_tokenizer_path", None)
+        if cached is not None and cached_path == normalized_path:
+            return cached
+
+        cache_dir = self._resolve_hf_cache_dir()
+        self._patch_hf_dynamic_module_cache_dir(cache_dir)
+        tokenizer = self._load_resolved_text_tokenizer(resolved_path=resolved_path, cache_dir=cache_dir)
+        if tokenizer.pad_token_id is None and tokenizer.eos_token is not None:
+            tokenizer.pad_token = tokenizer.eos_token
+        self._cached_text_tokenizer = tokenizer
+        self._cached_text_tokenizer_path = normalized_path
+        return tokenizer
+
+    @staticmethod
+    def _normalize_audio_tokenizer_type(audio_tokenizer_type: Optional[str]) -> Optional[str]:
+        if audio_tokenizer_type is None:
+            return None
+        normalized = str(audio_tokenizer_type).strip().lower()
+        if not normalized:
+            return None
+        if normalized == MOSS_AUDIO_TOKENIZER_NANO_TYPE:
+            return MOSS_AUDIO_TOKENIZER_NANO_TYPE
+        raise ValueError(
+            "Unsupported audio tokenizer type. "
+            f"The open-source package only supports '{MOSS_AUDIO_TOKENIZER_NANO_TYPE}'."
+        )
+
+    def _resolve_audio_tokenizer_type(self, audio_tokenizer_type: Optional[str]) -> str:
+        explicit_type = self._normalize_audio_tokenizer_type(audio_tokenizer_type)
+        if explicit_type is not None:
+            return explicit_type
+        config_type = self._normalize_audio_tokenizer_type(getattr(self.config, "audio_tokenizer_type", None))
+        return MOSS_AUDIO_TOKENIZER_NANO_TYPE if config_type is None else config_type
+
+    @staticmethod
+    def _set_decoder_attention_implementation(decoder, attn_implementation: str) -> None:
+        decoder.attn_implementation = str(attn_implementation)
+        if getattr(decoder, "config", None) is not None:
+            decoder.config._attn_implementation = str(attn_implementation)
+        for block in getattr(decoder, "h", []):
+            block.attn.attn_implementation = str(attn_implementation)
+
+    def _set_attention_implementation(
+        self,
+        attn_implementation: str,
+        local_attn_implementation: Optional[str] = None,
+    ) -> None:
+        resolved_global = str(attn_implementation)
+        resolved_local = resolved_global if local_attn_implementation is None else str(local_attn_implementation)
+        self.config.attn_implementation = resolved_global
+        self.config.gpt2_config._attn_implementation = resolved_global
+        self._set_decoder_attention_implementation(self.transformer, resolved_global)
+        self.config.local_transformer_attn_implementation = resolved_local
+        self._set_decoder_attention_implementation(self.local_transformer, resolved_local)
+
+    @staticmethod
+    def _select_fallback_attention_implementation(device: torch.device) -> str:
+        return "sdpa" if device.type == "cuda" else "eager"
+
+    @staticmethod
+    def _is_generation_stability_error(exc: Exception) -> bool:
+        message = str(exc)
+        return any(
+            marker in message
+            for marker in (
+                "Non-finite",
+                "device-side assert triggered",
+                "probability tensor contains either",
+                "flash_attention_2 requires fp16/bf16 tensors",
+            )
+        )
+
+    def _apply_inference_stability_fallback(self, device: torch.device) -> None:
+        fallback_attn = self._select_fallback_attention_implementation(device)
+        if next(self.parameters()).dtype != torch.float32:
+            self.to(device=device, dtype=torch.float32)
+        self._set_attention_implementation(fallback_attn)
+        logging.warning(
+            "retrying inference with dtype=float32 attn_implementation=%s due to numerical instability",
+            fallback_attn,
+        )
+
+    def _load_audio_tokenizer(
+        self,
+        audio_tokenizer=None,
+        audio_tokenizer_type: Optional[str] = None,
+        audio_tokenizer_pretrained_name_or_path: Optional[str] = None,
+        device: Optional[Union[str, torch.device]] = None,
+    ):
+        if audio_tokenizer is not None:
+            return audio_tokenizer
+
+        resolved_type = self._resolve_audio_tokenizer_type(audio_tokenizer_type=audio_tokenizer_type)
+        if resolved_type != MOSS_AUDIO_TOKENIZER_NANO_TYPE:
+            raise ValueError(
+                f"Unsupported audio tokenizer type {resolved_type!r}; expected '{MOSS_AUDIO_TOKENIZER_NANO_TYPE}'."
+            )
+
+        resolved_pretrained_name_or_path = (
+            audio_tokenizer_pretrained_name_or_path
+            or getattr(self.config, "audio_tokenizer_pretrained_name_or_path", None)
+            or DEFAULT_MOSS_AUDIO_TOKENIZER_PRETRAINED_NAME_OR_PATH
+        )
+        candidate_path = Path(str(resolved_pretrained_name_or_path)).expanduser()
+        if candidate_path.exists():
+            load_source = str(candidate_path.resolve())
+            load_kwargs: dict[str, object] = {
+                "trust_remote_code": True,
+                "local_files_only": True,
+                "force_download": True,
+            }
+            cache_key = f"{resolved_type}|{load_source}"
+        else:
+            load_source = str(resolved_pretrained_name_or_path)
+            load_kwargs = {
+                "trust_remote_code": True,
+            }
+            cache_key = f"{resolved_type}|{load_source}"
+
+        cached = getattr(self, "_cached_audio_tokenizer", None)
+        cached_path = getattr(self, "_cached_audio_tokenizer_path", None)
+        if cached is not None and cached_path == cache_key:
+            tokenizer = cached
+        else:
+            tokenizer = AutoModel.from_pretrained(load_source, **load_kwargs)
+            if hasattr(tokenizer, "eval"):
+                tokenizer.eval()
+            self._cached_audio_tokenizer = tokenizer
+            self._cached_audio_tokenizer_path = cache_key
+
+        resolved_device = self._resolve_device(device)
+        return tokenizer.to(resolved_device) if hasattr(tokenizer, "to") else tokenizer
+
+    @staticmethod
+    def _extract_tensor_candidate(output: Any) -> Any:
+        if torch.is_tensor(output) or isinstance(output, np.ndarray):
+            return output
+        for attr_name in ("audio_codes", "audio_token_ids", "codes", "tokens", "input_ids"):
+            value = getattr(output, attr_name, None)
+            if value is not None:
+                return value
+        if isinstance(output, dict):
+            for key in ("audio_codes", "audio_token_ids", "codes", "tokens", "input_ids"):
+                if key in output:
+                    return output[key]
+            if len(output) == 1:
+                return next(iter(output.values()))
+        if isinstance(output, (list, tuple)) and output:
+            if len(output) == 2 and isinstance(output[1], (int, float)):
+                return output[0]
+            return NanoTTSGlobalLocalForCausalLM._extract_tensor_candidate(output[0])
+        raise TypeError(f"Unsupported audio tokenizer output type: {type(output)!r}")
+
+    @staticmethod
+    def _extract_audio_code_length(output: Any) -> Optional[int]:
+        for attr_name in ("audio_codes_lengths", "audio_token_ids_lengths", "codes_lengths", "lengths"):
+            candidate = getattr(output, attr_name, None)
+            if candidate is not None:
+                lengths = torch.as_tensor(candidate).reshape(-1)
+                if lengths.numel() > 0:
+                    return int(lengths[0].item())
+        if isinstance(output, dict):
+            for key in ("audio_codes_lengths", "audio_token_ids_lengths", "codes_lengths", "lengths"):
+                if key in output:
+                    lengths = torch.as_tensor(output[key]).reshape(-1)
+                    if lengths.numel() > 0:
+                        return int(lengths[0].item())
+        if isinstance(output, (list, tuple)) and len(output) >= 2:
+            candidate = output[1]
+            if torch.is_tensor(candidate) or isinstance(candidate, np.ndarray):
+                lengths = torch.as_tensor(candidate).reshape(-1)
+                if lengths.numel() > 0:
+                    return int(lengths[0].item())
+            if isinstance(candidate, (int, float)):
+                return int(candidate)
+        return None
+
+    def _normalize_audio_codes(self, audio_codes: Any) -> torch.LongTensor:
+        code_length = self._extract_audio_code_length(audio_codes)
+        tensor = torch.as_tensor(self._extract_tensor_candidate(audio_codes))
+        if tensor.ndim == 1:
+            tensor = tensor.unsqueeze(-1)
+        if tensor.ndim == 3:
+            if tensor.shape[1] == 1 and tensor.shape[0] >= self.config.n_vq:
+                tensor = tensor[: self.config.n_vq, 0, :].transpose(0, 1)
+            elif tensor.shape[0] == 1:
+                tensor = tensor[0]
+            elif tensor.shape[1] == self.config.n_vq:
+                tensor = tensor.transpose(1, 2)[0]
+            elif tensor.shape[-1] == self.config.n_vq:
+                tensor = tensor[0]
+            else:
+                raise ValueError(f"Unable to normalize audio codes with shape {tuple(tensor.shape)}")
+
+        if tensor.ndim != 2:
+            raise ValueError(f"Expected audio codes with 2 dims after normalization, got {tuple(tensor.shape)}")
+        if tensor.shape[-1] != self.config.n_vq and tensor.shape[0] == self.config.n_vq:
+            tensor = tensor.transpose(0, 1)
+        elif tensor.shape[-1] != self.config.n_vq and tensor.shape[0] > self.config.n_vq:
+            tensor = tensor[: self.config.n_vq].transpose(0, 1)
+        elif tensor.shape[-1] > self.config.n_vq:
+            tensor = tensor[:, : self.config.n_vq]
+        if tensor.shape[-1] != self.config.n_vq:
+            raise ValueError(
+                f"Expected normalized audio codes with trailing dim {self.config.n_vq}, got {tuple(tensor.shape)}"
+            )
+        if code_length is not None:
+            tensor = tensor[:code_length]
+        return tensor.to(dtype=torch.long)
+
+    def _extract_waveform_and_sample_rate(
+        self,
+        decode_output: Any,
+        fallback_sample_rate: int,
+    ) -> tuple[torch.FloatTensor, int]:
+        sample_rate = fallback_sample_rate
+        waveform = decode_output
+        waveform_length = None
+
+        for key in ("sample_rate", "sampling_rate"):
+            value = getattr(decode_output, key, None)
+            if value is not None:
+                sample_rate = int(value)
+                break
+        for key in ("waveform", "audio", "wav", "samples"):
+            value = getattr(decode_output, key, None)
+            if value is not None:
+                waveform = value
+                break
+        for key in ("audio_lengths", "waveform_lengths", "lengths"):
+            value = getattr(decode_output, key, None)
+            if value is not None:
+                lengths = torch.as_tensor(value).reshape(-1)
+                if lengths.numel() > 0:
+                    waveform_length = int(lengths[0].item())
+                    break
+
+        if isinstance(decode_output, dict):
+            for key in ("sample_rate", "sampling_rate"):
+                if key in decode_output:
+                    sample_rate = int(decode_output[key])
+                    break
+            for key in ("waveform", "audio", "wav", "samples"):
+                if key in decode_output:
+                    waveform = decode_output[key]
+                    break
+            for key in ("audio_lengths", "waveform_lengths", "lengths"):
+                if key in decode_output:
+                    lengths = torch.as_tensor(decode_output[key]).reshape(-1)
+                    if lengths.numel() > 0:
+                        waveform_length = int(lengths[0].item())
+                        break
+        elif isinstance(decode_output, (list, tuple)) and decode_output:
+            if len(decode_output) == 2 and isinstance(decode_output[1], (int, float)):
+                waveform = decode_output[0]
+                sample_rate = int(decode_output[1])
+            else:
+                waveform = decode_output[0]
+
+        waveform_tensor = torch.as_tensor(waveform, dtype=torch.float32)
+        if waveform_tensor.ndim == 3 and waveform_tensor.shape[0] == 1:
+            waveform_tensor = waveform_tensor[0]
+        if waveform_tensor.ndim == 2 and waveform_tensor.shape[0] > waveform_tensor.shape[1]:
+            waveform_tensor = waveform_tensor.transpose(0, 1)
+        if waveform_tensor.ndim == 1:
+            waveform_tensor = waveform_tensor.unsqueeze(0)
+        if waveform_tensor.ndim != 2:
+            raise ValueError(f"Expected decoded waveform with 2 dims, got {tuple(waveform_tensor.shape)}")
+        if waveform_length is not None:
+            waveform_tensor = waveform_tensor[..., : max(0, waveform_length)]
+        return waveform_tensor.cpu(), sample_rate
+
+    def _call_audio_encode(
+        self,
+        audio_tokenizer,
+        waveform: torch.FloatTensor,
+        sample_rate: int,
+    ) -> Any:
+        del sample_rate
+        batch_encode_fn = getattr(audio_tokenizer, "batch_encode", None)
+        if batch_encode_fn is None:
+            raise AttributeError("audio_tokenizer must provide a batch_encode method.")
+
+        waveform_tensor = torch.as_tensor(waveform, dtype=torch.float32, device=self._resolve_device(waveform.device))
+        if waveform_tensor.ndim == 1:
+            waveform_tensor = waveform_tensor.unsqueeze(0)
+        if waveform_tensor.ndim != 2:
+            raise ValueError(
+                f"MOSS audio tokenizer encode expects waveform shaped like (C, T), got {tuple(waveform_tensor.shape)}"
+            )
+
+        with self._audio_tokenizer_inference_context(audio_tokenizer, waveform_tensor.device):
+            return batch_encode_fn([waveform_tensor], chunk_duration=None)
+
+    def _call_audio_decode(
+        self,
+        audio_tokenizer,
+        audio_token_ids: torch.LongTensor,
+        sample_rate: int,
+        nq: Optional[int] = None,
+    ) -> Any:
+        del sample_rate
+        batch_decode_fn = getattr(audio_tokenizer, "batch_decode", None)
+        if batch_decode_fn is None:
+            raise AttributeError("audio_tokenizer must provide a batch_decode method.")
+
+        effective_nq = self._resolve_inference_nq(nq)
+        decode_codes = self._prepare_audio_codes_for_decode(audio_token_ids, nq=effective_nq)
+        with self._audio_tokenizer_inference_context(audio_tokenizer, decode_codes.device):
+            return batch_decode_fn([decode_codes], num_quantizers=effective_nq, chunk_duration=None)
+
+    @staticmethod
+    def _resolve_audio_tokenizer_downsample_rate(audio_tokenizer) -> int:
+        for holder in (audio_tokenizer, getattr(audio_tokenizer, "config", None)):
+            if holder is None:
+                continue
+            for attr_name in ("downsample_rate", "hop_length", "frame_size"):
+                value = getattr(holder, attr_name, None)
+                if value is not None:
+                    return int(value)
+            sampling_rate = getattr(holder, "sampling_rate", None)
+            frame_rate = getattr(holder, "frame_rate", None)
+            if sampling_rate is not None and frame_rate not in (None, 0):
+                return int(round(float(sampling_rate) / float(frame_rate)))
+        raise ValueError("audio_tokenizer.downsample_rate is required for prompt-audio decoding.")
+
+    def _resolve_audio_tokenizer_sample_rate(self, audio_tokenizer) -> int:
+        for holder in (audio_tokenizer, getattr(audio_tokenizer, "config", None)):
+            if holder is None:
+                continue
+            for attr_name in ("sampling_rate", "sample_rate"):
+                value = getattr(holder, attr_name, None)
+                if value is not None:
+                    return int(value)
+        return int(self.config.audio_tokenizer_sample_rate)
+
+    @staticmethod
+    def _resolve_audio_tokenizer_channels(audio_tokenizer) -> int:
+        for holder in (audio_tokenizer, getattr(audio_tokenizer, "config", None)):
+            if holder is None:
+                continue
+            for attr_name in ("number_channels", "channels_numbers", "audio_channels", "channels", "num_channels"):
+                value = getattr(holder, attr_name, None)
+                if value is not None:
+                    return int(value)
+        return 1
+
+    @staticmethod
+    def _audio_tokenizer_inference_context(audio_tokenizer, device: Union[str, torch.device]):
+        del audio_tokenizer, device
+        return nullcontext()
+
+    def _prepare_audio_codes_for_decode(
+        self,
+        audio_token_ids: torch.LongTensor,
+        nq: Optional[int] = None,
+    ) -> torch.LongTensor:
+        effective_nq = self._resolve_inference_nq(nq)
+        tensor = torch.as_tensor(audio_token_ids, dtype=torch.long)
+        if tensor.ndim == 2:
+            if tensor.shape[-1] == self.config.n_vq and tensor.shape[0] != self.config.n_vq:
+                return tensor[:, :effective_nq].transpose(0, 1).contiguous()
+            if tensor.shape[0] == self.config.n_vq:
+                return tensor[:effective_nq].contiguous()
+        elif tensor.ndim == 3:
+            if tensor.shape[-1] == self.config.n_vq:
+                return tensor[..., :effective_nq].permute(2, 0, 1).contiguous()
+            if tensor.shape[0] == self.config.n_vq:
+                return tensor[:effective_nq].contiguous()
+        raise ValueError(
+            f"Expected generated audio token ids shaped like (T, {self.config.n_vq}) or ({self.config.n_vq}, T); got {tuple(tensor.shape)}"
+        )
+
+    def _load_reference_audio(
+        self,
+        reference_audio_path: Union[str, Path],
+        target_sample_rate: int,
+        target_channels: int,
+    ) -> tuple[torch.FloatTensor, int]:
+        waveform, sample_rate = torchaudio.load(str(reference_audio_path))
+        waveform = waveform.to(torch.float32)
+        if sample_rate != target_sample_rate:
+            waveform = torchaudio.functional.resample(waveform, sample_rate, target_sample_rate)
+            sample_rate = target_sample_rate
+        current_channels = int(waveform.shape[0])
+        if current_channels == target_channels:
+            return waveform, sample_rate
+        if current_channels == 1 and target_channels > 1:
+            return waveform.repeat(target_channels, 1), sample_rate
+        if current_channels > 1 and target_channels == 1:
+            return waveform.mean(dim=0, keepdim=True), sample_rate
+        raise ValueError(f"Unsupported reference audio channel conversion: {current_channels} -> {target_channels}")
+
+    def _decode_local_last_hidden_state(
+        self,
+        local_inputs_embeds: torch.FloatTensor,
+    ) -> torch.FloatTensor:
+        local_attention_mask = torch.ones(
+            local_inputs_embeds.shape[:2],
+            dtype=torch.bool,
+            device=local_inputs_embeds.device,
+        )
+        local_outputs = self.local_transformer(
+            input_ids=None,
+            attention_mask=local_attention_mask,
+            position_ids=None,
+            inputs_embeds=local_inputs_embeds,
+            use_cache=False,
+            output_attentions=False,
+            output_hidden_states=False,
+            return_dict=True,
+            cu_seqlens=None,
+            num_sequences=None,
+        )
+        return local_outputs.last_hidden_state[:, -1, :]
+
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        nq: Optional[int] = None,
+        max_new_frames: int = 300,
+        do_sample: bool = False,
+        text_temperature: float = 1.5,
+        text_top_p: float = 1.0,
+        text_top_k: int = 50,
+        audio_temperature: float = 1.7,
+        audio_top_p: float = 0.8,
+        audio_top_k: int = 25,
+        audio_repetition_penalty: float = 1.0,
+        use_kv_cache: bool = True,
+        return_dict_in_generate: bool = True,
+    ):
+        if input_ids.ndim == 2:
+            input_ids = input_ids.unsqueeze(0)
+        if input_ids.ndim != 3:
+            raise ValueError(f"Expected input_ids with 3 dims, got shape {tuple(input_ids.shape)}")
+        if attention_mask is None:
+            attention_mask = torch.ones(input_ids.shape[:2], dtype=torch.bool, device=input_ids.device)
+        elif attention_mask.ndim == 1:
+            attention_mask = attention_mask.unsqueeze(0)
+
+        effective_nq = self._resolve_inference_nq(nq)
+        batch_size = input_ids.shape[0]
+        current_input_ids = input_ids
+        current_attention_mask = attention_mask.to(device=input_ids.device)
+        current_model_input_ids = current_input_ids
+        generated_frames = []
+        finished = torch.zeros(batch_size, dtype=torch.bool, device=input_ids.device)
+        past_key_values = None
+        local_dtype = self.local_transformer.ln_f.weight.dtype
+
+        for _ in range(max_new_frames):
+            generated_audio_history = torch.stack(generated_frames, dim=1) if generated_frames else None
+            global_inputs_embeds = self._build_inputs_embeds(current_model_input_ids)
+            global_outputs = self.transformer(
+                input_ids=None,
+                past_key_values=past_key_values,
+                attention_mask=current_attention_mask,
+                position_ids=None,
+                inputs_embeds=global_inputs_embeds,
+                use_cache=use_kv_cache,
+                output_attentions=False,
+                output_hidden_states=False,
+                return_dict=True,
+                cu_seqlens=None,
+                num_sequences=None,
+            )
+            global_hidden_states = global_outputs.last_hidden_state[:, -1, :].to(dtype=local_dtype)
+
+            local_inputs_embeds = global_hidden_states.unsqueeze(1)
+            local_hidden_states = self._decode_local_last_hidden_state(local_inputs_embeds)
+            text_logits = self.text_lm_head(local_hidden_states)
+            self._ensure_finite_generation_logits(text_logits, "text logits")
+            next_text_tokens = self._sample_next_assistant_text_token(
+                logits=text_logits,
+                do_sample=do_sample,
+                temperature=text_temperature,
+                top_k=text_top_k,
+                top_p=text_top_p,
+            )
+            should_continue = next_text_tokens.eq(self.config.audio_assistant_slot_token_id) & ~finished
+            finished = finished | next_text_tokens.eq(self.config.audio_end_token_id)
+            if not should_continue.any():
+                break
+
+            next_frame_tokens = []
+            current_local_input = self.transformer.wte(next_text_tokens).to(dtype=local_dtype)
+            for channel_index in range(effective_nq):
+                local_inputs_embeds = torch.cat([local_inputs_embeds, current_local_input.unsqueeze(1)], dim=1)
+                local_hidden_states = self._decode_local_last_hidden_state(local_inputs_embeds)
+                channel_logits = self.audio_lm_heads[channel_index](local_hidden_states)
+                self._ensure_finite_generation_logits(channel_logits, f"audio logits[{channel_index}]")
+                channel_token = self._sample_next_token(
+                    logits=channel_logits,
+                    do_sample=do_sample,
+                    temperature=audio_temperature,
+                    top_k=audio_top_k,
+                    top_p=audio_top_p,
+                    previous_token_ids=(
+                        None if generated_audio_history is None else generated_audio_history[:, :, channel_index]
+                    ),
+                    repetition_penalty=audio_repetition_penalty,
+                )
+                next_frame_tokens.append(channel_token)
+                current_local_input = self.audio_embeddings[channel_index](channel_token).to(dtype=local_dtype)
+
+            next_frame_prefix = torch.stack(next_frame_tokens, dim=-1)
+            if effective_nq < self.config.n_vq:
+                next_frame = torch.full(
+                    (batch_size, self.config.n_vq),
+                    self.config.audio_pad_token_id,
+                    dtype=next_frame_prefix.dtype,
+                    device=next_frame_prefix.device,
+                )
+                next_frame[:, :effective_nq] = next_frame_prefix
+            else:
+                next_frame = next_frame_prefix
+            padded_next_frame = next_frame.masked_fill(~should_continue.unsqueeze(-1), self.config.audio_pad_token_id)
+            generated_frames.append(padded_next_frame)
+
+            next_row = self._build_generation_row(
+                batch_size=batch_size,
+                device=input_ids.device,
+                audio_token_ids=padded_next_frame,
+            )
+            if (~should_continue).any():
+                next_row[~should_continue, 0, 0] = self.config.pad_token_id
+                next_row[~should_continue, 0, 1:] = self.config.audio_pad_token_id
+
+            current_input_ids = torch.cat([current_input_ids, next_row], dim=1)
+            current_attention_mask = torch.cat([current_attention_mask, should_continue.unsqueeze(1)], dim=1)
+            if use_kv_cache:
+                current_model_input_ids = next_row
+                past_key_values = global_outputs.past_key_values
+            else:
+                current_model_input_ids = current_input_ids
+
+        if generated_frames:
+            audio_token_ids = torch.stack(generated_frames, dim=1)
+        else:
+            audio_token_ids = torch.empty((batch_size, 0, self.config.n_vq), dtype=torch.long, device=input_ids.device)
+
+        if not return_dict_in_generate:
+            return audio_token_ids
+        return NanoTTSGenerationOutput(audio_token_ids=audio_token_ids, prompt_input_ids=input_ids)
+
+    @torch.no_grad()
+    def inference(
+        self,
+        text: str,
+        output_audio_path: Union[str, Path],
+        mode: str = "continuation",
+        prompt_text: Optional[str] = None,
+        prompt_audio_path: Optional[Union[str, Path]] = None,
+        reference_audio_path: Optional[Union[str, Path]] = None,
+        text_tokenizer=None,
+        text_tokenizer_path: Optional[str] = None,
+        audio_tokenizer=None,
+        audio_tokenizer_type: Optional[str] = None,
+        audio_tokenizer_pretrained_name_or_path: Optional[str] = None,
+        device: Optional[Union[str, torch.device]] = None,
+        nq: Optional[int] = None,
+        max_new_frames: int = 300,
+        do_sample: bool = False,
+        text_temperature: float = 1.5,
+        text_top_p: float = 1.0,
+        text_top_k: int = 50,
+        audio_temperature: float = 1.7,
+        audio_top_p: float = 0.8,
+        audio_top_k: int = 25,
+        audio_repetition_penalty: float = 1.0,
+        use_kv_cache: bool = True,
+        voice_clone_max_text_tokens: int = DEFAULT_VOICE_CLONE_MAX_TEXT_TOKENS,
+        voice_clone_max_memory_per_sample_gb: float = DEFAULT_VOICE_CLONE_MAX_MEMORY_PER_SAMPLE_GB,
+    ) -> dict[str, Any]:
+        resolved_device = self._resolve_device(device)
+        effective_nq = self._resolve_inference_nq(nq)
+        if next(self.parameters()).device != resolved_device:
+            self.to(resolved_device)
+
+        was_training = self.training
+        self.eval()
+
+        text_tokenizer = self._load_text_tokenizer(
+            text_tokenizer=text_tokenizer,
+            text_tokenizer_path=text_tokenizer_path,
+        )
+        audio_tokenizer = self._load_audio_tokenizer(
+            audio_tokenizer=audio_tokenizer,
+            audio_tokenizer_type=audio_tokenizer_type,
+            audio_tokenizer_pretrained_name_or_path=audio_tokenizer_pretrained_name_or_path,
+            device=resolved_device,
+        )
+
+        target_sample_rate = self._resolve_audio_tokenizer_sample_rate(audio_tokenizer)
+        target_channels = self._resolve_audio_tokenizer_channels(audio_tokenizer)
+        effective_prompt_audio_path = prompt_audio_path or reference_audio_path
+        resolved_mode = self._resolve_inference_mode(
+            mode=mode,
+            has_prompt_text=prompt_text is not None,
+            has_prompt_audio=effective_prompt_audio_path is not None,
+        )
+        if reference_audio_path is not None and prompt_audio_path is None:
+            logging.warning(
+                "reference_audio_path=%s is treated as prompt_audio_path for backward compatibility.",
+                reference_audio_path,
+            )
+
+        prompt_audio_codes = None
+        if effective_prompt_audio_path is not None:
+            waveform, sample_rate = self._load_reference_audio(
+                effective_prompt_audio_path,
+                target_sample_rate,
+                target_channels,
+            )
+            encoded = self._call_audio_encode(
+                audio_tokenizer=audio_tokenizer,
+                waveform=waveform.to(resolved_device),
+                sample_rate=sample_rate,
+            )
+            prompt_audio_codes = self._mask_unused_audio_channels(
+                self._normalize_audio_codes(encoded),
+                nq=effective_nq,
+            ).to(resolved_device)
+
+        if resolved_mode == "voice_clone":
+            split_voice_clone_text_chunks = self._split_text_into_best_sentences(
+                text_tokenizer=text_tokenizer,
+                text=text,
+                max_tokens=voice_clone_max_text_tokens,
+            )
+            voice_clone_text_chunks = split_voice_clone_text_chunks if len(split_voice_clone_text_chunks) > 1 else [text]
+        else:
+            voice_clone_text_chunks = [text]
+
+        generated_audio_token_chunks: list[torch.LongTensor] = []
+        decoded_waveform_chunks: list[torch.FloatTensor] = []
+        decoded_sample_rate: Optional[int] = None
+
+        if resolved_mode == "voice_clone" and len(voice_clone_text_chunks) > 1:
+            voice_clone_chunk_batch_size = self._resolve_voice_clone_chunk_batch_size(
+                resolved_device=resolved_device,
+                chunk_count=len(voice_clone_text_chunks),
+                max_memory_per_sample_gb=float(voice_clone_max_memory_per_sample_gb),
+            )
+        else:
+            voice_clone_chunk_batch_size = 1
+
+        for batch_start in range(0, len(voice_clone_text_chunks), voice_clone_chunk_batch_size):
+            batch_chunks = voice_clone_text_chunks[batch_start : batch_start + voice_clone_chunk_batch_size]
+            batch_prompt_input_ids: list[torch.LongTensor] = []
+            batch_attention_masks: list[torch.BoolTensor] = []
+            for text_chunk in batch_chunks:
+                prompt_input_ids, attention_mask = self.build_inference_input_ids(
+                    text=text_chunk,
+                    text_tokenizer=text_tokenizer,
+                    mode=resolved_mode,
+                    prompt_text=prompt_text,
+                    prompt_audio_codes=prompt_audio_codes,
+                    device=resolved_device,
+                )
+                batch_prompt_input_ids.append(prompt_input_ids)
+                batch_attention_masks.append(attention_mask)
+
+            batched_prompt_input_ids, batched_attention_mask = self._left_pad_inference_batch(
+                input_id_batches=batch_prompt_input_ids,
+                attention_mask_batches=batch_attention_masks,
+                device=resolved_device,
+            )
+            batched_audio_token_ids = self._generate_audio_token_ids_with_fallback(
+                prompt_input_ids=batched_prompt_input_ids,
+                attention_mask=batched_attention_mask,
+                effective_nq=effective_nq,
+                max_new_frames=max_new_frames,
+                do_sample=do_sample,
+                text_temperature=text_temperature,
+                text_top_p=text_top_p,
+                text_top_k=text_top_k,
+                audio_temperature=audio_temperature,
+                audio_top_p=audio_top_p,
+                audio_top_k=audio_top_k,
+                audio_repetition_penalty=audio_repetition_penalty,
+                use_kv_cache=use_kv_cache,
+                resolved_device=resolved_device,
+            )
+
+            for sample_index in range(len(batch_chunks)):
+                audio_token_ids = self._trim_generated_audio_token_ids(
+                    batched_audio_token_ids[sample_index],
+                    effective_nq=effective_nq,
+                )
+                generated_audio_token_chunks.append(audio_token_ids)
+
+                if resolved_mode == "voice_clone" and len(voice_clone_text_chunks) > 1:
+                    decoded_waveform, current_sample_rate = self._decode_audio_token_ids_to_waveform(
+                        audio_tokenizer=audio_tokenizer,
+                        audio_token_ids=audio_token_ids,
+                        target_sample_rate=target_sample_rate,
+                        effective_nq=effective_nq,
+                        resolved_device=resolved_device,
+                    )
+                    if decoded_sample_rate is None:
+                        decoded_sample_rate = current_sample_rate
+                    elif decoded_sample_rate != current_sample_rate:
+                        raise ValueError(
+                            f"Decoded sample rates differ across voice_clone chunks: {decoded_sample_rate} vs {current_sample_rate}"
+                        )
+                    decoded_waveform_chunks.append(decoded_waveform)
+
+        if generated_audio_token_chunks:
+            audio_token_ids = torch.cat(generated_audio_token_chunks, dim=0)
+        else:
+            audio_token_ids = torch.empty((0, self.config.n_vq), dtype=torch.long, device=resolved_device)
+
+        if resolved_mode == "voice_clone" and len(voice_clone_text_chunks) > 1:
+            waveform = (
+                self._concat_voice_clone_waveform_chunks(
+                    waveform_chunks=decoded_waveform_chunks,
+                    text_chunks=voice_clone_text_chunks,
+                    sample_rate=decoded_sample_rate,
+                )
+                if decoded_waveform_chunks
+                else torch.zeros((target_channels, 0), dtype=torch.float32)
+            )
+        else:
+            decode_audio_token_ids = audio_token_ids
+            prompt_waveform_prefix_samples = 0
+            if resolved_mode == "continuation" and prompt_audio_codes is not None:
+                decode_audio_token_ids = torch.cat([prompt_audio_codes, audio_token_ids], dim=0)
+                prompt_waveform_prefix_samples = (
+                    int(prompt_audio_codes.shape[0]) * self._resolve_audio_tokenizer_downsample_rate(audio_tokenizer)
+                )
+
+            waveform, decoded_sample_rate = self._decode_audio_token_ids_to_waveform(
+                audio_tokenizer=audio_tokenizer,
+                audio_token_ids=decode_audio_token_ids,
+                target_sample_rate=target_sample_rate,
+                effective_nq=effective_nq,
+                resolved_device=resolved_device,
+            )
+            if prompt_waveform_prefix_samples > 0:
+                if decoded_sample_rate != target_sample_rate:
+                    prompt_waveform_prefix_samples = int(
+                        round(prompt_waveform_prefix_samples * float(decoded_sample_rate) / float(target_sample_rate))
+                    )
+                prompt_waveform_prefix_samples = min(prompt_waveform_prefix_samples, int(waveform.shape[-1]))
+                waveform = waveform[:, prompt_waveform_prefix_samples:]
+
+        assert decoded_sample_rate is not None
+
+        output_path = Path(output_audio_path)
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        torchaudio.save(str(output_path), waveform, decoded_sample_rate)
+
+        if was_training:
+            self.train()
+
+        return {
+            "audio_path": str(output_path),
+            "sample_rate": decoded_sample_rate,
+            "audio_token_ids": audio_token_ids.detach().cpu(),
+            "waveform": waveform,
+            "reference_audio_token_ids": None if prompt_audio_codes is None else prompt_audio_codes.detach().cpu(),
+        }

prompting.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+from typing import List, Sequence
+
+from .configuration_nanotts import NanoTTSConfig
+
+
+USER_ROLE_PREFIX = "user\n"
+USER_TEMPLATE_REFERENCE_PREFIX = (
+    "<user_inst>\n"
+    "- Reference(s):\n"
+)
+USER_TEMPLATE_AFTER_REFERENCE = (
+    "\n- Instruction:\nNone\n"
+    "- Tokens:\nNone\n"
+    "- Quality:\nNone\n"
+    "- Sound Event:\nNone\n"
+    "- Ambient Sound:\nNone\n"
+    "- Language:\nNone\n"
+    "- Text:\n"
+)
+USER_TEMPLATE_PREFIX = USER_TEMPLATE_REFERENCE_PREFIX + "None" + USER_TEMPLATE_AFTER_REFERENCE
+USER_TEMPLATE_SUFFIX = "\n</user_inst>"
+ASSISTANT_TURN_PREFIX = "\n"
+ASSISTANT_ROLE_PREFIX = "assistant\n"
+
+
+def encode_text(tokenizer, text: str) -> List[int]:
+    try:
+        return list(tokenizer.encode(text, add_special_tokens=False))
+    except TypeError:
+        return list(tokenizer.encode(text))
+
+
+def decode_text(tokenizer, token_ids: Sequence[int]) -> str:
+    try:
+        return str(
+            tokenizer.decode(
+                list(token_ids),
+                skip_special_tokens=False,
+                clean_up_tokenization_spaces=False,
+            )
+        )
+    except TypeError:
+        try:
+            return str(tokenizer.decode(list(token_ids), skip_special_tokens=False))
+        except TypeError:
+            return str(tokenizer.decode(list(token_ids)))
+
+
+def build_user_prompt_prefix(tokenizer, config: NanoTTSConfig) -> List[int]:
+    return [config.im_start_token_id] + encode_text(tokenizer, USER_ROLE_PREFIX) + encode_text(
+        tokenizer,
+        USER_TEMPLATE_REFERENCE_PREFIX,
+    )
+
+
+def build_user_prompt_after_reference(tokenizer) -> List[int]:
+    return encode_text(tokenizer, USER_TEMPLATE_AFTER_REFERENCE)
+
+
+def build_assistant_prompt_prefix(tokenizer, config: NanoTTSConfig) -> List[int]:
+    return encode_text(tokenizer, USER_TEMPLATE_SUFFIX) + [config.im_end_token_id] + encode_text(
+        tokenizer,
+        ASSISTANT_TURN_PREFIX,
+    ) + [config.im_start_token_id] + encode_text(
+        tokenizer,
+        ASSISTANT_ROLE_PREFIX,
+    )
+
+
+def build_prompt_prefix(tokenizer, config: NanoTTSConfig) -> List[int]:
+    return (
+        build_user_prompt_prefix(tokenizer, config)
+        + encode_text(tokenizer, "None")
+        + build_user_prompt_after_reference(tokenizer)
+    )
+
+
+def build_prompt_suffix(tokenizer, config: NanoTTSConfig) -> List[int]:
+    return build_assistant_prompt_prefix(tokenizer, config)
+
+
+def build_prompt_token_ids(
+    tokenizer,
+    config: NanoTTSConfig,
+    text_token_ids: Sequence[int],
+) -> List[int]:
+    return build_prompt_prefix(tokenizer, config) + [int(token_id) for token_id in text_token_ids] + build_prompt_suffix(
+        tokenizer,
+        config,
+    )

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d19e63fdc6a35f61a7d1c27e06bfacaba7c1ed40ea3c619c86efc64bcd50a496
+size 234693095

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "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
+  }
+}

tokenization_nanotts_sentencepiece.py

@@ -0,0 +1,103 @@
+from __future__ import annotations
+
+import shutil
+from pathlib import Path
+from typing import Any
+
+import sentencepiece as spm
+from transformers import PreTrainedTokenizer
+
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+
+class NanoTTSSentencePieceTokenizer(PreTrainedTokenizer):
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file: str,
+        unk_token: str = "<unk>",
+        bos_token: str = "<s>",
+        eos_token: str = "</s>",
+        pad_token: str = "<pad>",
+        sp_model_kwargs: dict[str, Any] | None = None,
+        **kwargs,
+    ) -> None:
+        self.vocab_file = str(vocab_file)
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else dict(sp_model_kwargs)
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+        super().__init__(
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return int(self.sp_model.get_piece_size())
+
+    def get_vocab(self) -> dict[str, int]:
+        vocab = {self.sp_model.id_to_piece(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text: str) -> list[str]:
+        return list(self.sp_model.encode(text, out_type=str))
+
+    def _convert_token_to_id(self, token: str) -> int:
+        token_id = int(self.sp_model.piece_to_id(token))
+        return token_id
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return str(self.sp_model.id_to_piece(int(index)))
+
+    def convert_tokens_to_string(self, tokens: list[str]) -> str:
+        return str(self.sp_model.decode(tokens))
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: str | None = None) -> tuple[str]:
+        save_dir = Path(save_directory)
+        save_dir.mkdir(parents=True, exist_ok=True)
+        out_name = "tokenizer.model" if filename_prefix is None else f"{filename_prefix}-tokenizer.model"
+        out_path = save_dir / out_name
+        if Path(self.vocab_file).resolve() != out_path.resolve():
+            shutil.copyfile(self.vocab_file, out_path)
+        return (str(out_path),)
+
+    def build_inputs_with_special_tokens(
+        self,
+        token_ids_0: list[int],
+        token_ids_1: list[int] | None = None,
+    ) -> list[int]:
+        if token_ids_1 is None:
+            return list(token_ids_0)
+        return list(token_ids_0) + list(token_ids_1)
+
+    def get_special_tokens_mask(
+        self,
+        token_ids_0: list[int],
+        token_ids_1: list[int] | None = None,
+        already_has_special_tokens: bool = False,
+    ) -> list[int]:
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0,
+                token_ids_1=token_ids_1,
+                already_has_special_tokens=True,
+            )
+        if token_ids_1 is None:
+            return [0] * len(token_ids_0)
+        return [0] * (len(token_ids_0) + len(token_ids_1))
+
+    def create_token_type_ids_from_sequences(
+        self,
+        token_ids_0: list[int],
+        token_ids_1: list[int] | None = None,
+    ) -> list[int]:
+        if token_ids_1 is None:
+            return [0] * len(token_ids_0)
+        return [0] * (len(token_ids_0) + len(token_ids_1))

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c353ee1479b536bf414c1b247f5542b6607fb8ae91320e5af1781fee200fddff
+size 470897

tokenizer_config.json

@@ -0,0 +1,52 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "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": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [],
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_nanotts_sentencepiece.NanoTTSSentencePieceTokenizer",
+      null
+    ]
+  },
+  "backend": "custom",
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "extra_special_tokens": {},
+  "model_max_length": 16384,
+  "pad_token": "<pad>",
+  "tokenizer_class": "NanoTTSSentencePieceTokenizer",
+  "unk_token": "<unk>"
+}