Add MOSS music audio captioner

extensions_built_in/captioner/AceStepCaptioner.py

@@ -425,6 +425,14 @@ class AceStepCaptioner(BaseCaptioner):
             )
         return item
 
+    @staticmethod
+    def _place_inputs(model, inputs):
+        """Move a processor batch onto the model; the thinker also reads its
+        2D pad mask from the model. Called again before every decode pass."""
+        inputs = inputs.to(model.device).to(model.dtype)
+        model._pad_mask_2d = inputs.get("attention_mask", None)
+        return inputs
+
     def _generate(
         self,
         model,
@@ -434,17 +442,17 @@ class AceStepCaptioner(BaseCaptioner):
         retries: tuple = (),
         break_loops: bool = False,
         file_path: str = "",
+        max_new_tokens: int = MAX_NEW_TOKENS,
     ) -> str:
-        inputs = inputs.to(model.device).to(model.dtype)
+        inputs = self._place_inputs(model, inputs)
         # a generate that dies between static-cache creation and its first
         # forward leaves a half-built cache that breaks every later call
         stale_cache = getattr(model, "_cache", None)
         if stale_cache is not None and not stale_cache.is_initialized:
             del model._cache
-        model._pad_mask_2d = inputs.get("attention_mask", None)
         input_len = inputs["input_ids"].shape[1]
         # greedy decode loops on repeated phrases without a penalty
-        gen_kwargs = {"max_new_tokens": MAX_NEW_TOKENS, "repetition_penalty": repetition_penalty}
+        gen_kwargs = {"max_new_tokens": max_new_tokens, "repetition_penalty": repetition_penalty}
         if model.generation_config.cache_implementation == "static":
             if input_len + 16 < STATIC_MAX_LENGTH:
                 from transformers.generation import MaxLengthCriteria, StoppingCriteriaList
@@ -453,7 +461,7 @@ class AceStepCaptioner(BaseCaptioner):
                     "repetition_penalty": repetition_penalty,
                     "max_length": STATIC_MAX_LENGTH,
                     "stopping_criteria": StoppingCriteriaList(
-                        [MaxLengthCriteria(max_length=min(input_len + MAX_NEW_TOKENS, STATIC_MAX_LENGTH))]
+                        [MaxLengthCriteria(max_length=min(input_len + max_new_tokens, STATIC_MAX_LENGTH))]
                     ),
                 }
             else:
@@ -473,7 +481,7 @@ class AceStepCaptioner(BaseCaptioner):
             kwargs = dict(gen_kwargs)
             if breaker is not None:
                 kwargs["logits_processor"] = LogitsProcessorList([breaker])
-            model._pad_mask_2d = inputs.get("attention_mask", None)
+            self._place_inputs(model, inputs)
             generated = model.generate(**inputs, **kwargs)
             return strip_eos(generated[0, input_len:]), breaker
 
@@ -533,25 +541,29 @@ class AceStepCaptioner(BaseCaptioner):
             file_path=file_path,
         )
 
-    def get_audio_caption(self, inputs) -> str:
+    def get_audio_caption(self, inputs, file_path: str = "") -> str:
         if self.caption_config.low_vram and self.model.device != torch.device("cpu"):
             self.model.to("cpu")
         if self.model2.device == torch.device("cpu"):
             self.model2.to(self.device_torch)
         return self._generate(self.model2, self.processor2, inputs, CAPTIONER_REPETITION_PENALTY)
 
+    def _transcribe_item(self, item: dict) -> str:
+        """Raw transcriber text for a prepped item: '# Languages ... # Lyrics ...'."""
+        return self.get_audio_lyrics(item["lyrics_inputs"], item["file"])
+
     def _caption_item(self, item: dict) -> str:
         if self.caption_config.caption_format == "yue2":
-            lyrics = clean_lyrics(self.get_audio_lyrics(item["lyrics_inputs"], item["file"]))
+            lyrics = clean_lyrics(self._transcribe_item(item))
             if self.caption_config.fixed_caption is not None:
                 caption = self.caption_config.fixed_caption
             else:
-                caption = self.get_audio_caption(item["caption_inputs"])
+                caption = self.get_audio_caption(item["caption_inputs"], item["file"])
             caption = " ".join(caption.split())
             return f"{caption}\n[Lyrics]\n{lyrics}"
 
         analysis = item["analysis"]
-        lyrics = self.get_audio_lyrics(item["lyrics_inputs"], item["file"])
+        lyrics = self._transcribe_item(item)
 
         language = "en"
         if "# Languages" in lyrics and "# Lyrics" in lyrics:
@@ -563,7 +575,7 @@ class AceStepCaptioner(BaseCaptioner):
         if self.caption_config.fixed_caption is not None:
             caption = self.caption_config.fixed_caption
         else:
-            caption = self.get_audio_caption(item["caption_inputs"])
+            caption = self.get_audio_caption(item["caption_inputs"], item["file"])
 
         output = f"<CAPTION>\n{caption}\n</CAPTION>\n"
         output += f"<LYRICS>\n{lyrics}\n</LYRICS>\n"

extensions_built_in/captioner/MossMusicCaptioner.py

@@ -0,0 +1,169 @@
+"""MOSS-Music captioner: one 8B audio-language model (OpenMOSS-Team/MOSS-Music-8B-Instruct)
+does both the lyric transcription and the description that AceStepCaptioner
+runs two Qwen2.5-Omni checkpoints for. Output layouts, loop breaking and the
+worker pipeline are inherited from AceStepCaptioner."""
+
+import re
+
+from optimum.quanto import freeze
+
+from toolkit.audio.moss_music import HF_REPO, load_moss_music
+from toolkit.basic import flush
+from toolkit.util.quantize import get_qtype, quantize
+
+from .AceStepCaptioner import (
+    CAPTIONER_REPETITION_PENALTY,
+    TRANSCRIBER_REPETITION_PENALTY,
+    TRANSCRIBER_RETRIES,
+    AceStepCaptionConfig,
+    AceStepCaptioner,
+    analyze_audio,
+)
+
+# timestamps are requested on purpose: each line is anchored to the input time
+# markers, so decode has to advance through the song instead of looping on a
+# repeated phrase. They are stripped from the saved caption.
+TRANSCRIBE_PROMPT = "Transcribe the lyrics of this song with timestamps."
+# the model writes multi-section essays with timings, chords and lyric quotes
+# unless framed as a generator prompt; this keeps it to one plain paragraph
+CAPTION_PROMPT = (
+    "Write a prompt that a text-to-music generator could use to recreate this track. One paragraph, "
+    "under 80 words: genre, mood, instrumentation, tempo feel, production style, vocal style. "
+    "No timestamps, section timings, chord names, or lyric quotes."
+)
+LANGUAGE_PROMPT = (
+    "What language are the lyrics sung in? Answer with only the two-letter "
+    "ISO 639-1 language code, or none if there are no lyrics."
+)
+LANGUAGE_MAX_NEW_TOKENS = 8
+
+# the model writes "[00:18.00 - 00:22.00] line"; a looping decode can run the
+# seconds field past 59 ("[00:100.00 - ...]"), so any bracketed run of time
+# characters is treated as a timestamp, plus bare mm:ss spans at a line start
+_TIME = r"(?:\d{1,2}:)?\d{1,3}:\d{1,3}(?:[.:]\d{1,3})?|\d+(?:\.\d+)?\s*s"
+_SPAN = rf"(?:{_TIME})(?:\s*(?:-{{1,3}}>?|–|—|~|to)\s*(?:{_TIME}))?"
+_TIME_CHARS = r"[\d:.\s\-–—>~s]*\d[\d:.\s\-–—>~s]*"
+TIMESTAMP_RE = re.compile(
+    rf"\[{_TIME_CHARS}\]|\({_TIME_CHARS}\)|<{_TIME_CHARS}>|^\s*{_SPAN}\s*[:\-–—|]?",
+    re.MULTILINE,
+)
+
+
+def strip_timestamps(text: str) -> str:
+    """Remove time annotations and tidy the line layout they leave behind."""
+    text = TIMESTAMP_RE.sub("", text)
+    lines = [" ".join(line.split()) for line in text.splitlines()]
+    out = []
+    for line in lines:
+        if line == "" and (not out or out[-1] == ""):
+            continue
+        out.append(line)
+    return "\n".join(out).strip()
+
+
+class MossMusicCaptionConfig(AceStepCaptionConfig):
+    def __init__(self, **kwargs):
+        kwargs.setdefault("model_name_or_path", HF_REPO)
+        super().__init__(**kwargs)
+        self.caption_prompt: str = kwargs.get("caption_prompt") or CAPTION_PROMPT
+        self.transcribe_prompt: str = kwargs.get("transcribe_prompt") or TRANSCRIBE_PROMPT
+        # MOSS transcribes the full mix better than a separated vocal stem
+        self.extract_vocals_before_transcribe = False
+        # leave the model's "[00:18.00 - 00:22.00]" line times in the saved lyrics
+        self.keep_timestamps: bool = bool(kwargs.get("keep_timestamps", False))
+
+
+class MossMusicCaptioner(AceStepCaptioner):
+    caption_config_class = MossMusicCaptionConfig
+    caption_config: MossMusicCaptionConfig
+
+    def load_model(self):
+        self.print_and_status_update("Loading MOSS-Music model")
+        self.model, self.processor = load_moss_music(
+            self.caption_config.model_name_or_path,
+            dtype=self.torch_dtype,
+            device=str(self.device_torch),
+        )
+        if self.caption_config.quantize:
+            self.print_and_status_update("Quantizing MOSS-Music model")
+            quantize(self.model, weights=get_qtype(self.caption_config.qtype))
+            freeze(self.model)
+            flush()
+        # one model does both jobs; AceStepCaptioner's second slot stays empty
+        self.model2 = None
+        self.processor2 = None
+        flush()
+
+    @staticmethod
+    def _place_inputs(model, inputs):
+        inputs = inputs.to(model.device)
+        inputs["audio_data"] = inputs["audio_data"].to(model.dtype)
+        return inputs
+
+    def _make_inputs(self, audio_data, prompt_text: str):
+        inputs = self.processor(text=prompt_text, audios=[audio_data])
+        inputs["audio_input_mask"] = inputs["input_ids"] == self.processor.audio_token_id
+        return inputs
+
+    def _prep_file(self, file_path: str) -> dict:
+        audio_data = self._load_audio(file_path)
+        is_ace_step = self.caption_config.caption_format == "ace_step"
+        item = {
+            "file": file_path,
+            "analysis": analyze_audio(file_path) if is_ace_step else None,
+            "lyrics_inputs": self._make_inputs(audio_data, self.caption_config.transcribe_prompt),
+        }
+        if is_ace_step:
+            item["language_inputs"] = self._make_inputs(audio_data, LANGUAGE_PROMPT)
+        if self.caption_config.fixed_caption is None:
+            item["caption_inputs"] = self._make_inputs(
+                audio_data, self.caption_config.caption_prompt
+            )
+        return item
+
+    def get_audio_lyrics(self, inputs, file_path: str = "") -> str:
+        text = self._generate(
+            self.model,
+            self.processor,
+            inputs,
+            TRANSCRIBER_REPETITION_PENALTY,
+            retries=TRANSCRIBER_RETRIES,
+            break_loops=True,
+            file_path=file_path,
+        )
+        return text.strip() if self.caption_config.keep_timestamps else strip_timestamps(text)
+
+    def get_audio_language(self, inputs) -> str:
+        answer = self._generate(
+            self.model,
+            self.processor,
+            inputs,
+            repetition_penalty=1.0,
+            max_new_tokens=LANGUAGE_MAX_NEW_TOKENS,
+        )
+        match = re.search(r"[a-z]{2,3}", answer.lower())
+        code = match.group(0) if match else ""
+        # instrumentals answer none; ACE-Step's default language is en
+        return code if len(code) == 2 and code != "no" else "en"
+
+    def _transcribe_item(self, item: dict) -> str:
+        lyrics = self.get_audio_lyrics(item["lyrics_inputs"], item["file"])
+        if "language_inputs" not in item:
+            return lyrics
+        # the layout AceStepCaptioner._caption_item parses from the ACE transcriber
+        language = self.get_audio_language(item["language_inputs"])
+        return f"# Languages\n{language}\n# Lyrics\n{lyrics}"
+
+    def get_audio_caption(self, inputs, file_path: str = "") -> str:
+        # a tag-list prompt can degenerate into a repeated tag; same loop
+        # breaking as transcription
+        caption = self._generate(
+            self.model,
+            self.processor,
+            inputs,
+            CAPTIONER_REPETITION_PENALTY,
+            retries=TRANSCRIBER_RETRIES,
+            break_loops=True,
+            file_path=file_path,
+        )
+        return strip_timestamps(caption)

extensions_built_in/captioner/__init__.py

@@ -13,6 +13,18 @@ class AceStepCaptionerExtension(Extension):
         return AceStepCaptioner
 
 
+class MossMusicCaptionerExtension(Extension):
+    uid = "MossMusicCaptioner"
+    name = "MOSS-Music Captioner"
+
+    @classmethod
+    def get_process(cls):
+        # import your process class here so it is only loaded when needed and return it
+        from .MossMusicCaptioner import MossMusicCaptioner
+
+        return MossMusicCaptioner
+
+
 class Qwen3VLCaptionerExtension(Extension):
     uid = "Qwen3VLCaptioner"
     name = "Qwen 3VL Captioner"
@@ -63,6 +75,7 @@ class Ideogram4CaptionerExtension(Extension):
 
 AI_TOOLKIT_EXTENSIONS = [
     AceStepCaptionerExtension,
+    MossMusicCaptionerExtension,
     Qwen3VLCaptionerExtension,
     Qwen3OmniCaptionerExtension,
     Qwen25OmniCaptionerExtension,

toolkit/audio/moss_music/LICENSE

@@ -0,0 +1,21 @@
+The code in this directory is adapted from MOSS-Music by the OpenMOSS Team
+(https://github.com/OpenMOSS/MOSS-Music), licensed under the Apache License,
+Version 2.0. Modifications for ai-toolkit: transformers 5.x compatibility
+(WhisperEncoderLayer tensor return, generate input preparation), removal of the
+chat-template loader, and direct Qwen2Tokenizer loading.
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

toolkit/audio/moss_music/__init__.py

@@ -0,0 +1,35 @@
+import torch
+
+from .configuration import MossMusicConfig, MossMusicEncoderConfig
+from .modeling import MossMusicModel
+from .processing import MelConfig, MossMusicProcessor
+
+HF_REPO = "OpenMOSS-Team/MOSS-Music-8B-Instruct"
+
+
+def load_moss_music(
+    name_or_path: str = HF_REPO,
+    dtype: torch.dtype = torch.bfloat16,
+    device="cpu",
+    enable_time_marker: bool = True,
+):
+    """Model + processor from a hub id or local checkpoint dir. Time markers
+    (elapsed seconds interleaved into the audio tokens) are what the model was
+    trained with and keep long transcriptions from looping; leave them on."""
+    model = MossMusicModel.from_pretrained(name_or_path, dtype=dtype, device_map=device)
+    model.eval()
+    processor = MossMusicProcessor.from_pretrained(
+        name_or_path, enable_time_marker=enable_time_marker
+    )
+    return model, processor
+
+
+__all__ = [
+    "HF_REPO",
+    "MelConfig",
+    "MossMusicConfig",
+    "MossMusicEncoderConfig",
+    "MossMusicModel",
+    "MossMusicProcessor",
+    "load_moss_music",
+]

toolkit/audio/moss_music/configuration.py

@@ -0,0 +1,135 @@
+from typing import List, Optional
+from dataclasses import dataclass, field
+
+from transformers import PretrainedConfig, Qwen3Config
+
+
+@dataclass
+class MossMusicEncoderConfig:
+    d_model: int = 1280
+    output_dim: int = 1280
+    num_mel_bins: int = 128
+    encoder_layers: int = 32
+    encoder_attention_heads: int = 20
+    encoder_ffn_dim: int = 5120
+    downsample_rate: int = 8
+    downsample_hidden_size: int = 480
+    encoder_attention_window_size: int = 100
+    max_source_positions: int = 1500
+    dropout: float = 0.1
+    attention_dropout: float = 0.1
+    activation_dropout: float = 0.0
+    activation_function: str = "gelu"
+    layer_norm_eps: float = 1e-5
+    _attn_implementation: str = "eager"
+    pretrained_path: str = ""
+
+    n_window: int = 200
+    conv_chunksize: int = 64
+
+    deepstack_encoder_layer_indexes: List[int] = field(
+        default_factory=lambda: [8, 16, 24]
+    )
+
+    @classmethod
+    def from_dict(cls, config_dict):
+        if config_dict is None:
+            return cls()
+        allowed_keys = set(cls.__dataclass_fields__.keys())
+        filtered = {k: v for k, v in config_dict.items() if k in allowed_keys}
+        return cls(**filtered)
+
+    def to_dict(self):
+        return {
+            "d_model": self.d_model,
+            "output_dim": self.output_dim,
+            "num_mel_bins": self.num_mel_bins,
+            "encoder_layers": self.encoder_layers,
+            "encoder_attention_heads": self.encoder_attention_heads,
+            "encoder_ffn_dim": self.encoder_ffn_dim,
+            "downsample_rate": self.downsample_rate,
+            "downsample_hidden_size": self.downsample_hidden_size,
+            "encoder_attention_window_size": self.encoder_attention_window_size,
+            "max_source_positions": self.max_source_positions,
+            "dropout": self.dropout,
+            "attention_dropout": self.attention_dropout,
+            "activation_dropout": self.activation_dropout,
+            "activation_function": self.activation_function,
+            "layer_norm_eps": self.layer_norm_eps,
+            "_attn_implementation": self._attn_implementation,
+            "pretrained_path": self.pretrained_path,
+            "n_window": self.n_window,
+            "conv_chunksize": self.conv_chunksize,
+            "deepstack_encoder_layer_indexes": list(
+                self.deepstack_encoder_layer_indexes or []
+            ),
+        }
+
+
+class MossMusicConfig(PretrainedConfig):
+    model_type = "moss_music"
+    is_composition = True
+
+    def __init__(
+        self,
+        audio_config=None,
+        language_config=None,
+        adapter_hidden_size=8192,
+        ignore_index=-100,
+        deepstack_num_inject_layers: Optional[int] = None,
+        **kwargs,
+    ):
+        if isinstance(audio_config, dict):
+            audio_config = MossMusicEncoderConfig.from_dict(audio_config)
+        elif audio_config is None:
+            audio_config = MossMusicEncoderConfig()
+
+        if isinstance(language_config, dict):
+            language_config = Qwen3Config(**language_config)
+        elif language_config is None:
+            language_config = Qwen3Config()
+
+        self.audio_config = audio_config
+        self.language_config = language_config
+        self.adapter_hidden_size = adapter_hidden_size
+        self.ignore_index = ignore_index
+        self.deepstack_num_inject_layers = deepstack_num_inject_layers
+
+        _propagate_keys = {
+            "num_hidden_layers",
+            "eos_token_id",
+            "bos_token_id",
+            "vocab_size",
+            "tie_word_embeddings",
+        }
+        for key in ("num_hidden_layers", "eos_token_id", "bos_token_id", "vocab_size"):
+            kwargs.setdefault(key, getattr(language_config, key, None))
+        kwargs.setdefault("tie_word_embeddings", False)
+
+        if hasattr(language_config, "to_dict"):
+            _lang_keys = set(language_config.to_dict().keys())
+            for key in list(kwargs.keys()):
+                if key in _lang_keys and key not in _propagate_keys:
+                    kwargs.pop(key)
+
+        super().__init__(**kwargs)
+
+    def to_dict(self):
+        output = super().to_dict()
+        output["audio_config"] = (
+            self.audio_config.to_dict()
+            if hasattr(self.audio_config, "to_dict")
+            else self.audio_config
+        )
+        output["language_config"] = (
+            self.language_config.to_dict()
+            if hasattr(self.language_config, "to_dict")
+            else self.language_config
+        )
+        output["adapter_hidden_size"] = self.adapter_hidden_size
+        output["ignore_index"] = self.ignore_index
+        output["deepstack_num_inject_layers"] = self.deepstack_num_inject_layers
+        return output
+
+
+__all__ = ["MossMusicEncoderConfig", "MossMusicConfig"]

toolkit/audio/moss_music/modeling.py

@@ -0,0 +1,568 @@
+from typing import Any, List, Optional, Tuple
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers.generation.utils import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.qwen3.modeling_qwen3 import Qwen3DecoderLayer, Qwen3Model
+from transformers.models.whisper.modeling_whisper import WhisperEncoderLayer
+
+from .configuration import MossMusicConfig, MossMusicEncoderConfig
+
+
+class SinusoidsPositionEmbedding(nn.Module):
+    def __init__(self, num_positions: int, embedding_dim: int):
+        super().__init__()
+        max_timescale = 10000.0
+        log_timescale_increment = math.log(max_timescale) / (embedding_dim // 2 - 1)
+        inv_timescales = torch.exp(
+            -log_timescale_increment * torch.arange(embedding_dim // 2).float()
+        )
+        self.register_buffer("inv_timescales", inv_timescales, persistent=False)
+
+    def forward(self, seq_len: int, device: torch.device):
+        scaled_time = torch.arange(
+            seq_len, device=device, dtype=self.inv_timescales.dtype
+        ).unsqueeze(1) * self.inv_timescales.unsqueeze(0)
+        sin_emb = torch.sin(scaled_time)
+        cos_emb = torch.cos(scaled_time)
+        pos_emb = torch.cat([sin_emb, cos_emb], dim=1)
+        return pos_emb.unsqueeze(0)
+
+
+class MossMusicEncoder(nn.Module):
+    """Audio encoder with conv-stem downsampling and Whisper transformer layers."""
+
+    def __init__(self, config: MossMusicEncoderConfig):
+        super().__init__()
+        self.config = config
+        self.gelu = nn.GELU()
+
+        self.conv1 = nn.Conv2d(
+            1,
+            config.downsample_hidden_size,
+            kernel_size=(3, 3),
+            stride=(2, 2),
+            padding=(1, 1),
+        )
+        self.conv2 = nn.Conv2d(
+            config.downsample_hidden_size,
+            config.downsample_hidden_size,
+            kernel_size=(3, 3),
+            stride=(2, 2),
+            padding=(1, 1),
+        )
+        self.conv3 = nn.Conv2d(
+            config.downsample_hidden_size,
+            config.downsample_hidden_size,
+            kernel_size=(3, 3),
+            stride=(2, 2),
+            padding=(1, 1),
+        )
+
+        # 128 mel bins / 8 = 16 after 3 convs with stride=2
+        self.stem_proj = nn.Linear(config.downsample_hidden_size * 16, config.d_model)
+        self.embed_positions = SinusoidsPositionEmbedding(
+            config.max_source_positions, config.d_model
+        )
+        self.layers = nn.ModuleList(
+            [WhisperEncoderLayer(config) for _ in range(config.encoder_layers)]
+        )
+        self.layer_norm = nn.LayerNorm(config.d_model, eps=config.layer_norm_eps)
+        self.out_proj = (
+            nn.Linear(config.d_model, config.output_dim, bias=False)
+            if config.output_dim != config.d_model
+            else nn.Identity()
+        )
+
+        self.deepstack_encoder_layer_indexes = list(
+            config.deepstack_encoder_layer_indexes or []
+        )
+        self._deepstack_capture_map = {
+            layer_idx: capture_idx
+            for capture_idx, layer_idx in enumerate(self.deepstack_encoder_layer_indexes)
+        }
+
+        self.n_window = int(config.n_window)
+        self.chunk_frames = int(self.n_window * 2)
+        self.conv_chunksize = int(config.conv_chunksize)
+
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.conv1.weight.dtype
+
+    @staticmethod
+    def _compute_downsampled_length(lengths: torch.Tensor) -> torch.Tensor:
+        def conv_out_len(L):
+            return (L - 1) // 2 + 1
+
+        return conv_out_len(conv_out_len(conv_out_len(lengths)))
+
+    def _encode_chunk_batch(
+        self,
+        input_features: torch.Tensor,
+        seq_lengths: torch.Tensor,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Encode a batch of (already padded) chunks through the conv stem and
+        transformer layers. Returns (last_hidden, ordered_deepstack_hidden_states).
+        """
+        if input_features.dim() == 2:
+            input_features = input_features.unsqueeze(0)
+
+        downsampled_lengths = self._compute_downsampled_length(seq_lengths)
+
+        # [B, n_mels, T] -> [B, 1, n_mels, T]
+        x = input_features.unsqueeze(1)
+        x = self.gelu(self.conv1(x))
+        x = self.gelu(self.conv2(x))
+        x = self.gelu(self.conv3(x))
+
+        # [B, C, F, T] -> [B, T, C*F]
+        x = x.permute(0, 3, 1, 2).contiguous().flatten(2)
+        x = self.stem_proj(x)
+
+        max_len = int(downsampled_lengths.max().item())
+        if x.size(1) > max_len:
+            x = x[:, :max_len, :]
+
+        positions = self.embed_positions(x.shape[1], x.device)
+        x = x + positions.to(x.dtype)
+
+        padding_mask = (
+            torch.arange(x.size(1), device=x.device)[None, :]
+            >= downsampled_lengths[:, None]
+        )
+        attention_mask = (1.0 - (~padding_mask).to(dtype=x.dtype)) * torch.finfo(
+            x.dtype
+        ).min
+        attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)
+
+        deepstack_hidden_states: List[Optional[torch.Tensor]] = [None] * len(
+            self.deepstack_encoder_layer_indexes
+        )
+        for layer_idx, layer in enumerate(self.layers):
+            x = layer(x, attention_mask)
+            capture_idx = self._deepstack_capture_map.get(layer_idx)
+            if capture_idx is not None:
+                deepstack_hidden_states[capture_idx] = x
+
+        x = self.layer_norm(x)
+        x = self.out_proj(x)
+
+        ordered_deepstack_hidden_states = [
+            h for h in deepstack_hidden_states if h is not None
+        ]
+        if not isinstance(self.out_proj, nn.Identity):
+            ordered_deepstack_hidden_states = [
+                self.out_proj(h) for h in ordered_deepstack_hidden_states
+            ]
+        return x, ordered_deepstack_hidden_states
+
+    def forward(
+        self,
+        input_features: torch.Tensor,
+        feature_lens: Optional[torch.Tensor] = None,
+        output_deepstack_hidden_states: bool = True,
+    ) -> BaseModelOutputWithPast:
+        if input_features.dim() == 3:
+            if feature_lens is None:
+                feature_lens = torch.full(
+                    (input_features.size(0),),
+                    input_features.size(-1),
+                    dtype=torch.long,
+                    device=input_features.device,
+                )
+            else:
+                feature_lens = feature_lens.to(
+                    device=input_features.device, dtype=torch.long
+                )
+            valid_chunks = [
+                input_features[i, :, : int(feature_lens[i].item())]
+                for i in range(int(input_features.shape[0]))
+            ]
+            input_features = torch.cat(valid_chunks, dim=1)
+        elif input_features.dim() != 2:
+            raise ValueError(
+                f"Expected [n_mels, T] or [B, n_mels, T], got {tuple(input_features.shape)}."
+            )
+
+        if feature_lens is None:
+            feature_lens = torch.tensor(
+                [int(input_features.shape[1])],
+                device=input_features.device,
+                dtype=torch.long,
+            )
+        else:
+            feature_lens = feature_lens.to(
+                device=input_features.device, dtype=torch.long
+            )
+
+        chunk_frames = int(self.chunk_frames)
+        chunk_num = torch.ceil(
+            feature_lens.to(torch.float32) / float(chunk_frames)
+        ).long()
+        chunk_lengths = torch.full(
+            (int(chunk_num.sum().item()),),
+            chunk_frames,
+            dtype=torch.long,
+            device=feature_lens.device,
+        )
+        tail_chunk_index = F.pad(chunk_num, (1, 0), value=-1).cumsum(0)[1:]
+        chunk_lengths[tail_chunk_index] = feature_lens % chunk_frames
+        chunk_lengths[chunk_lengths == 0] = chunk_frames
+
+        chunk_list = input_features.T.split(chunk_lengths.tolist(), dim=0)
+        padded_feature = nn.utils.rnn.pad_sequence(
+            chunk_list, batch_first=True
+        ).transpose(1, 2)
+
+        feature_lens_after_cnn = self._compute_downsampled_length(chunk_lengths)
+        t_down_max = (
+            int(feature_lens_after_cnn.max().item())
+            if feature_lens_after_cnn.numel() > 0
+            else 0
+        )
+        padded_mask_after_cnn = nn.utils.rnn.pad_sequence(
+            [
+                torch.ones(int(L.item()), dtype=torch.bool, device=padded_feature.device)
+                for L in feature_lens_after_cnn
+            ],
+            batch_first=True,
+        )
+        if padded_mask_after_cnn.shape[1] < t_down_max:
+            padded_mask_after_cnn = F.pad(
+                padded_mask_after_cnn,
+                (0, t_down_max - padded_mask_after_cnn.shape[1]),
+                value=False,
+            )
+
+        num_deepstack = len(self.deepstack_encoder_layer_indexes)
+        padded_embeds: List[torch.Tensor] = []
+        deepstack_padded_embeds: List[List[torch.Tensor]] = [
+            [] for _ in range(num_deepstack)
+        ]
+        for feat_chunk, len_chunk in zip(
+            padded_feature.split(self.conv_chunksize, dim=0),
+            chunk_lengths.split(self.conv_chunksize, dim=0),
+        ):
+            out, deepstack_outs = self._encode_chunk_batch(feat_chunk, len_chunk)
+            if out.shape[1] < t_down_max:
+                out = F.pad(out, (0, 0, 0, t_down_max - out.shape[1]))
+            padded_embeds.append(out)
+            if output_deepstack_hidden_states and num_deepstack > 0:
+                if len(deepstack_outs) != num_deepstack:
+                    raise RuntimeError(
+                        "Deepstack output count does not match configured layer indexes."
+                    )
+                for capture_idx, ds in enumerate(deepstack_outs):
+                    if ds.shape[1] < t_down_max:
+                        ds = F.pad(ds, (0, 0, 0, t_down_max - ds.shape[1]))
+                    deepstack_padded_embeds[capture_idx].append(ds)
+
+        if padded_embeds:
+            padded_embed = torch.cat(padded_embeds, dim=0)
+        else:
+            padded_embed = torch.empty(
+                (0, t_down_max, self.config.output_dim),
+                device=padded_feature.device,
+            )
+
+        valid_tokens = padded_embed[padded_mask_after_cnn]  # [N_valid, D]
+        last_hidden_state = valid_tokens.unsqueeze(0)  # [1, N_valid, D]
+
+        deepstack_states: Optional[Tuple[torch.Tensor, ...]] = None
+        if output_deepstack_hidden_states and num_deepstack > 0:
+            collected: List[torch.Tensor] = []
+            for chunks_list in deepstack_padded_embeds:
+                if chunks_list:
+                    ds = torch.cat(chunks_list, dim=0)
+                    collected.append(ds[padded_mask_after_cnn].unsqueeze(0))
+                else:
+                    collected.append(
+                        torch.empty(
+                            (1, 0, self.config.output_dim),
+                            device=padded_feature.device,
+                            dtype=padded_embed.dtype,
+                        )
+                    )
+            deepstack_states = tuple(collected)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=last_hidden_state,
+            hidden_states=deepstack_states,
+        )
+
+
+class GatedMLP(nn.Module):
+    def __init__(self, input_size, hidden_size, output_size):
+        super().__init__()
+        self.gate_proj = nn.Linear(input_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(input_size, hidden_size, bias=False)
+        self.down_proj = nn.Linear(hidden_size, output_size, bias=False)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class MossMusicPreTrainedModel(PreTrainedModel):
+    config_class = MossMusicConfig
+    config: MossMusicConfig
+    base_model_prefix = ""
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen3DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _can_compile_fullgraph = False
+    _supports_attention_backend = True
+    _can_record_outputs = {"hidden_states": Qwen3DecoderLayer}
+
+
+class MossMusicModel(MossMusicPreTrainedModel, GenerationMixin):
+    config_class = MossMusicConfig
+    _tied_weights_keys: List[str] = []
+
+    def __init__(self, config: MossMusicConfig):
+        super().__init__(config)
+
+        self.audio_encoder = MossMusicEncoder(config.audio_config)
+        self.language_model = Qwen3Model(config.language_config)
+
+        self.audio_adapter = GatedMLP(
+            input_size=config.audio_config.output_dim,
+            hidden_size=config.adapter_hidden_size,
+            output_size=config.language_config.hidden_size,
+        )
+
+        deepstack_k = len(
+            getattr(config.audio_config, "deepstack_encoder_layer_indexes", []) or []
+        )
+        if config.deepstack_num_inject_layers is not None:
+            deepstack_k = min(deepstack_k, int(config.deepstack_num_inject_layers))
+        self.deepstack_audio_merger_list = nn.ModuleList(
+            [
+                GatedMLP(
+                    input_size=config.audio_config.output_dim,
+                    hidden_size=config.adapter_hidden_size,
+                    output_size=config.language_config.hidden_size,
+                )
+                for _ in range(deepstack_k)
+            ]
+        )
+
+        self.vocab_size = config.language_config.vocab_size
+        self.lm_head = nn.Linear(
+            config.language_config.hidden_size, self.vocab_size, bias=False
+        )
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_audio_features(self, input_features, feature_lens):
+        audio_outputs = self.audio_encoder(
+            input_features=input_features,
+            feature_lens=feature_lens,
+            output_deepstack_hidden_states=True,
+        )
+        deepstack = (
+            list(audio_outputs.hidden_states)
+            if audio_outputs.hidden_states is not None
+            else None
+        )
+        return audio_outputs.last_hidden_state, deepstack
+
+    def _apply_deepstack_to_hidden_states(
+        self,
+        hidden_states: torch.Tensor,
+        audio_input_mask: torch.Tensor,
+        deepstack_embeds: torch.Tensor,
+    ) -> torch.Tensor:
+        audio_input_mask = audio_input_mask.to(hidden_states.device)
+        deepstack_embeds = deepstack_embeds.to(hidden_states.device, hidden_states.dtype)
+        flat = deepstack_embeds.reshape(-1, deepstack_embeds.shape[-1])
+        hs = hidden_states.clone()
+        hs[audio_input_mask] = hs[audio_input_mask] + flat
+        return hs
+
+    def _register_llm_deepstack_hooks(
+        self,
+        audio_input_mask: torch.Tensor,
+        deepstack_audio_embeds: List[torch.Tensor],
+    ):
+        if deepstack_audio_embeds is None or len(deepstack_audio_embeds) == 0:
+            return []
+
+        layers = getattr(self.language_model, "layers", None)
+        if layers is None:
+            raise RuntimeError(
+                "Qwen3Model does not expose `.layers`; cannot register DeepStack hooks."
+            )
+
+        num_inject = len(deepstack_audio_embeds)
+        handles = []
+
+        for layer_idx, layer in enumerate(layers):
+            if layer_idx >= num_inject:
+                break
+
+            def _make_llm_hook(k: int):
+                def _hook(_module, _inputs, _output):
+                    if isinstance(_output, (tuple, list)):
+                        hs = _output[0]
+                        new_hs = self._apply_deepstack_to_hidden_states(
+                            hs, audio_input_mask, deepstack_audio_embeds[k]
+                        )
+                        return (new_hs,) + tuple(_output[1:])
+                    else:
+                        return self._apply_deepstack_to_hidden_states(
+                            _output, audio_input_mask, deepstack_audio_embeds[k]
+                        )
+
+                return _hook
+
+            handles.append(layer.register_forward_hook(_make_llm_hook(layer_idx)))
+
+        return handles
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        audio_data: Optional[torch.FloatTensor] = None,
+        audio_data_seqlens: Optional[torch.Tensor] = None,
+        audio_input_mask: Optional[torch.Tensor] = None,
+        **kwargs: Any,
+    ) -> CausalLMOutputWithPast:
+        # generate() still passes return_dict; transformers 5 models ignore it
+        kwargs.pop("return_dict", None)
+        if inputs_embeds is None:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        hook_handles = []
+        if audio_data is not None:
+            if audio_input_mask is None:
+                raise ValueError(
+                    "audio_input_mask is required when audio_data is provided."
+                )
+
+            audio_embeds, deepstack = self.get_audio_features(audio_data, audio_data_seqlens)
+            audio_embeds = self.audio_adapter(audio_embeds)
+
+            audio_token_count = int(audio_input_mask.to(torch.int32).sum().item())
+            if audio_token_count != int(audio_embeds.shape[1]):
+                raise ValueError(
+                    f"Audio token count mismatch: audio_input_mask has {audio_token_count} audio tokens, "
+                    f"but audio_embeds has length {int(audio_embeds.shape[1])}."
+                )
+
+            mask_expanded = audio_input_mask.unsqueeze(-1).expand_as(inputs_embeds)
+            inputs_embeds = inputs_embeds.clone()
+            inputs_embeds.masked_scatter_(mask_expanded, audio_embeds)
+
+            if deepstack is not None and len(self.deepstack_audio_merger_list) > 0:
+                deepstack_audio_embeds = []
+                for i, x in enumerate(deepstack[: len(self.deepstack_audio_merger_list)]):
+                    ds = self.deepstack_audio_merger_list[i](x)
+                    if int(ds.shape[1]) != audio_token_count:
+                        raise ValueError(
+                            f"DeepStack audio seq_len mismatch at index {i}: "
+                            f"expected {audio_token_count}, got {int(ds.shape[1])}."
+                        )
+                    deepstack_audio_embeds.append(ds)
+
+                try:
+                    hook_handles = self._register_llm_deepstack_hooks(
+                        audio_input_mask, deepstack_audio_embeds
+                    )
+                except Exception:
+                    for h in hook_handles:
+                        h.remove()
+                    raise
+
+        try:
+            outputs = self.language_model(
+                input_ids=None,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds,
+                use_cache=use_cache,
+                **kwargs,
+            )
+        finally:
+            for h in hook_handles:
+                h.remove()
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = nn.CrossEntropyLoss(ignore_index=self.config.ignore_index)
+            shift_logits = shift_logits.view(-1, self.config.language_config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        next_sequence_length=None,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        is_first_iteration=False,
+        **kwargs,
+    ):
+        # audio is encoded once into the prompt embeds; re-sending it on decode
+        # steps would re-run the encoder against a stale audio_input_mask
+        if not is_first_iteration:
+            for key in ("audio_data", "audio_data_seqlens", "audio_input_mask"):
+                kwargs.pop(key, None)
+        return super().prepare_inputs_for_generation(
+            input_ids,
+            next_sequence_length=next_sequence_length,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            is_first_iteration=is_first_iteration,
+            **kwargs,
+        )
+
+
+__all__ = [
+    "MossMusicEncoderConfig",
+    "MossMusicConfig",
+    "MossMusicModel",
+]

toolkit/audio/moss_music/processing.py

@@ -0,0 +1,250 @@
+import re
+from dataclasses import dataclass
+from typing import List, Optional, Sequence, Union
+
+import numpy as np
+import torch
+from transformers import BatchEncoding, Qwen2Tokenizer
+
+
+@dataclass
+class MelConfig:
+    mel_sr: int = 16000
+    mel_dim: int = 128
+    mel_n_fft: int = 400
+    mel_hop_length: int = 160
+    mel_dtype: torch.dtype = torch.bfloat16
+
+
+class MossMusicProcessor:
+    """Builds MOSS-Music inputs: Whisper log-mel features plus a Qwen chat
+    prompt whose audio span is one placeholder token per 80ms of audio, with
+    the elapsed seconds written in as digit tokens every 2s (time markers)."""
+
+    _AUDIO_SPAN_RE = re.compile(r"<\|audio_bos\|>(?:<\|AUDIO\|>)+<\|audio_eos\|>")
+
+    def __init__(
+        self,
+        tokenizer,
+        *,
+        mel_config: Optional[MelConfig] = None,
+        enable_time_marker: bool = True,
+        audio_token_id: int = 151654,
+        audio_start_id: int = 151669,
+        audio_end_id: int = 151670,
+    ):
+        self.tokenizer = tokenizer
+        self.audio_token_id = int(audio_token_id)
+        self.audio_start_id = int(audio_start_id)
+        self.audio_end_id = int(audio_end_id)
+        self.enable_time_marker = bool(enable_time_marker)
+        self.config = mel_config or MelConfig()
+        self._whisper_feature_extractor = None
+
+        self._digit_token_ids = {str(d): 15 + d for d in range(10)}
+        self.audio_tokens_per_second = 12.5
+        self.time_marker_every_seconds = 2
+        self.time_marker_every_audio_tokens = int(
+            self.audio_tokens_per_second * self.time_marker_every_seconds
+        )
+        self.model_input_names = [
+            "input_ids",
+            "attention_mask",
+            "audio_data",
+            "audio_data_seqlens",
+        ]
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        tokenizer_kwargs = {}
+        for key in ["cache_dir", "revision", "token", "local_files_only"]:
+            if key in kwargs:
+                tokenizer_kwargs[key] = kwargs[key]
+        # the checkpoint's tokenizer_config names Qwen2Tokenizer; loading it
+        # directly skips AutoTokenizer's remote-code prompt for the auto_map
+        tokenizer = Qwen2Tokenizer.from_pretrained(
+            pretrained_model_name_or_path, **tokenizer_kwargs
+        )
+        return cls(
+            tokenizer,
+            mel_config=kwargs.pop("mel_config", None),
+            enable_time_marker=kwargs.pop("enable_time_marker", True),
+            audio_token_id=kwargs.pop("audio_token_id", 151654),
+            audio_start_id=kwargs.pop("audio_start_id", 151669),
+            audio_end_id=kwargs.pop("audio_end_id", 151670),
+        )
+
+    @staticmethod
+    def _conv3_downsample_len(raw_mel_len: int) -> int:
+        def conv_out_len(length: int) -> int:
+            return (length - 1) // 2 + 1
+
+        return conv_out_len(conv_out_len(conv_out_len(int(raw_mel_len))))
+
+    def _get_whisper_feature_extractor(self):
+        if self._whisper_feature_extractor is not None:
+            return self._whisper_feature_extractor
+
+        from transformers.models.whisper.feature_extraction_whisper import (
+            WhisperFeatureExtractor,
+        )
+
+        self._whisper_feature_extractor = WhisperFeatureExtractor(
+            feature_size=int(self.config.mel_dim),
+            sampling_rate=int(self.config.mel_sr),
+            hop_length=int(self.config.mel_hop_length),
+            n_fft=int(self.config.mel_n_fft),
+        )
+        return self._whisper_feature_extractor
+
+    def _extract_mel(self, audio: Union[np.ndarray, torch.Tensor]) -> torch.Tensor:
+        if isinstance(audio, torch.Tensor):
+            audio = audio.detach().cpu().numpy()
+        wav_np = np.ascontiguousarray(audio, dtype=np.float32)
+        if wav_np.ndim == 2:
+            wav_np = wav_np[0]
+        # the raw fbank pass has no 30s chunking, so full songs come through whole
+        fe = self._get_whisper_feature_extractor()
+        feats = fe._np_extract_fbank_features(wav_np[None, ...], device="cpu")
+        return torch.from_numpy(feats[0]).to(dtype=self.config.mel_dtype)
+
+    def _get_time_marker_token_ids(self, second: int) -> List[int]:
+        return [self._digit_token_ids[digit] for digit in str(second)]
+
+    def _build_audio_tokens_with_time_markers(self, audio_seq_len: int) -> List[int]:
+        total_duration_seconds = audio_seq_len / self.audio_tokens_per_second
+        num_full_seconds = int(total_duration_seconds)
+
+        token_ids: List[int] = []
+        audio_tokens_consumed = 0
+        for second in range(
+            self.time_marker_every_seconds,
+            num_full_seconds + 1,
+            self.time_marker_every_seconds,
+        ):
+            marker_pos = (
+                second // self.time_marker_every_seconds
+            ) * self.time_marker_every_audio_tokens
+            audio_segment_len = marker_pos - audio_tokens_consumed
+            if audio_segment_len > 0:
+                token_ids.extend([self.audio_token_id] * audio_segment_len)
+                audio_tokens_consumed += audio_segment_len
+            token_ids.extend(self._get_time_marker_token_ids(second))
+
+        remaining = audio_seq_len - audio_tokens_consumed
+        if remaining > 0:
+            token_ids.extend([self.audio_token_id] * remaining)
+        return token_ids
+
+    def _build_audio_placeholder_ids(self, num_audio_tokens: int) -> List[int]:
+        if self.enable_time_marker:
+            return self._build_audio_tokens_with_time_markers(num_audio_tokens)
+        return [self.audio_token_id] * num_audio_tokens
+
+    def _build_default_prompt(self, text: str, has_audio: bool) -> str:
+        if has_audio:
+            return (
+                "<|im_start|>system\n"
+                "You are a helpful assistant.<|im_end|>\n"
+                "<|im_start|>user\n"
+                "<|audio_bos|><|AUDIO|><|audio_eos|>\n"
+                f"{text}<|im_end|>\n"
+                "<|im_start|>assistant\n"
+            )
+        return (
+            "<|im_start|>system\n"
+            "You are a helpful assistant.<|im_end|>\n"
+            "<|im_start|>user\n"
+            f"{text}<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+
+    def _build_input_from_prompt(self, prompt: str, token_lens: List[int]) -> List[int]:
+        spans = list(self._AUDIO_SPAN_RE.finditer(prompt))
+        if len(spans) != len(token_lens):
+            raise ValueError(
+                f"Audio placeholder count mismatch: found {len(spans)} spans in text, "
+                f"but got {len(token_lens)} audio inputs."
+            )
+
+        input_ids: List[int] = []
+        cursor = 0
+        for index, match in enumerate(spans):
+            prefix = prompt[cursor : match.start()]
+            if prefix:
+                input_ids.extend(self.tokenizer.encode(prefix, add_special_tokens=False))
+
+            input_ids.append(self.audio_start_id)
+            input_ids.extend(self._build_audio_placeholder_ids(int(token_lens[index])))
+            input_ids.append(self.audio_end_id)
+            cursor = match.end()
+
+        suffix = prompt[cursor:]
+        if suffix:
+            input_ids.extend(self.tokenizer.encode(suffix, add_special_tokens=False))
+        return input_ids
+
+    def __call__(
+        self,
+        *,
+        text: Union[str, Sequence[str]],
+        audios: Optional[Sequence[Union[np.ndarray, torch.Tensor]]] = None,
+        audio: Optional[Sequence[Union[np.ndarray, torch.Tensor]]] = None,
+        return_tensors: str = "pt",
+        **kwargs,
+    ):
+        if isinstance(text, (list, tuple)):
+            if len(text) != 1:
+                raise ValueError(f"Expected text batch size 1, got {len(text)}")
+            prompt_text = text[0]
+        else:
+            prompt_text = text
+
+        audio_list = audios if audios is not None else audio
+        audio_list = [] if audio_list is None else list(audio_list)
+
+        mels: List[torch.Tensor] = []
+        raw_lengths: List[int] = []
+        token_lens: List[int] = []
+        for one_audio in audio_list:
+            mel = self._extract_mel(one_audio)
+            raw_len = int(mel.shape[-1])
+            mels.append(mel)
+            raw_lengths.append(raw_len)
+            token_lens.append(self._conv3_downsample_len(raw_len))
+
+        if mels:
+            max_length = max(raw_lengths)
+            audio_batch = torch.zeros(
+                (len(mels), self.config.mel_dim, max_length),
+                dtype=self.config.mel_dtype,
+            )
+            for index, mel in enumerate(mels):
+                audio_batch[index, :, : mel.shape[-1]] = mel
+            seqlens_tensor = torch.tensor(raw_lengths, dtype=torch.long)
+        else:
+            audio_batch = None
+            seqlens_tensor = None
+
+        if self._AUDIO_SPAN_RE.search(prompt_text) is None:
+            prompt_text = self._build_default_prompt(prompt_text, has_audio=bool(audio_list))
+        input_ids_list = self._build_input_from_prompt(prompt_text, token_lens)
+
+        input_ids_tensor = torch.tensor([input_ids_list], dtype=torch.long)
+        data = {
+            "input_ids": input_ids_tensor,
+            "attention_mask": torch.ones_like(input_ids_tensor),
+        }
+        if audio_batch is not None:
+            data["audio_data"] = audio_batch
+            data["audio_data_seqlens"] = seqlens_tensor
+        return BatchEncoding(data=data, tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        return self.tokenizer.decode(*args, **kwargs)
+
+
+__all__ = ["MelConfig", "MossMusicProcessor"]

ui/src/components/CaptionSimpleJob.tsx

@@ -283,6 +283,13 @@ const CaptionSimpleJob: React.FC<Props> = ({ jobConfig, setJobConfig, gpuIDs, se
                 onChange={value => setJobConfig(value, 'config.process[0].caption.extract_vocals_before_transcribe')}
               />
             )}
+            {additionalSections.includes('caption.keep_timestamps') && (
+              <Checkbox
+                label="Keep Lyric Timestamps"
+                checked={jobConfig.config.process[0].caption.keep_timestamps || false}
+                onChange={value => setJobConfig(value, 'config.process[0].caption.keep_timestamps')}
+              />
+            )}
             {additionalSections.includes('caption.thinking') && (
               <Checkbox
                 label="Thinking"

ui/src/helpers/captionOptions.ts

@@ -1,7 +1,7 @@
 import { CloudLora, GroupedSelectOption, SelectOption } from "@/types";
 
 type CaptionGroup = 'image' | 'music' | 'video' | 'image/video/sound';
-type AdditionalSections = 'caption.model_name_or_path2' | 'caption.caption_prompt' | 'caption.max_res' | 'caption.max_new_tokens' | 'caption.fixed_caption' | 'caption.caption_format' | 'caption.extract_vocals_before_transcribe' | 'caption.thinking' | 'caption.batch_size' | 'caption.layer_offloading';
+type AdditionalSections = 'caption.model_name_or_path2' | 'caption.caption_prompt' | 'caption.max_res' | 'caption.max_new_tokens' | 'caption.fixed_caption' | 'caption.caption_format' | 'caption.extract_vocals_before_transcribe' | 'caption.keep_timestamps' | 'caption.thinking' | 'caption.batch_size' | 'caption.layer_offloading';
 
 export interface CaptionOption {
     name: string;
@@ -84,6 +84,13 @@ Part 2: on the next line write [Lyrics] and then the complete lyrics transcribed
 
 Transcribe only what is actually sung. Be decisive. No preamble, no explanations, no markdown - output only the tag line and the lyrics block.`;
 
+// MOSS-Music description prompts. The lyric transcription prompt is fixed in the
+// captioner (it asks for timestamps as loop anchors and strips them afterwards).
+// Framing the description as a generator prompt is what keeps the model to one
+// plain paragraph instead of a sectioned essay with timings and chord names.
+const mossMusicCaptionPrompt = "Write a prompt that a text-to-music generator could use to recreate this track. One paragraph, under 80 words: genre, mood, instrumentation, tempo feel, production style, vocal style. No timestamps, section timings, chord names, or lyric quotes.";
+const mossMusicTagsPrompt = "Describe this music as a single line of comma-separated style tags: genre, mood, vocal type and delivery, lead instruments, production style, tempo feel, era. Lowercase tags only, no sentences.";
+
 // Editable ADDITIONAL INSTRUCTIONS block injected into the Ideogram system prompt.
 // Users can tweak this for dataset-specific guidance without altering the fixed
 // output contract, element/background rules, or bbox format.
@@ -114,6 +121,32 @@ export const captionerTypes: CaptionOption[] = [
             'caption.extract_vocals_before_transcribe',
         ],
     },
+    {
+        name: 'MossMusicCaptioner',
+        label: 'MOSS-Music',
+        group: 'music',
+        defaults: {
+            'config.process[0].caption.model_name_or_path': ['OpenMOSS-Team/MOSS-Music-8B-Instruct', defaultNameOrPath],
+            'config.process[0].caption.extensions': [extensionsAudio, defaultExtensions],
+            'config.process[0].caption.caption_format': ['ace_step', undefined],
+            'config.process[0].caption.caption_prompt': [mossMusicCaptionPrompt, undefined],
+            'config.process[0].caption.keep_timestamps': [false, undefined],
+            'config.process[0].caption.compile': [true, false],
+        },
+        name_or_path_options: [
+            { value: 'OpenMOSS-Team/MOSS-Music-8B-Instruct', label: 'OpenMOSS-Team/MOSS-Music-8B-Instruct' },
+        ],
+        captionPrompts: {
+            'Description (ACE-Step)': mossMusicCaptionPrompt,
+            'Style tags (YuE2)': mossMusicTagsPrompt,
+        },
+        additionalSections: [
+            'caption.fixed_caption',
+            'caption.caption_format',
+            'caption.keep_timestamps',
+            'caption.caption_prompt',
+        ],
+    },
     {
         name: 'Qwen3VLCaptioner',
         label: 'Qwen3-VL',

ui/src/types.ts

@@ -339,6 +339,7 @@ export interface CaptionProcessConfig {
     fixed_caption?: string;
     caption_format?: string;
     extract_vocals_before_transcribe?: boolean;
+    keep_timestamps?: boolean;
     caption_extension?: string;
     thinking?: boolean;
     batch_size?: number;