@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 36d450e3b8
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 9, 12:30 UTC

Adds FunAudioChat multimodal audio model support (#2) (#33058)

Signed-off-by: ramos <49182011+nemoramo@users.noreply.github.com> Signed-off-by: mayufeng <mayufeng@example.com> Co-authored-by: mayufeng <mayufeng@example.com>

ramos committed Jan 28, 2026 at 13:18 UTC 36d450e3b88476c07e2dde9297692386f7174175
8 files changed +1264
examples/offline_inference/audio_language.py
+37
@@ -117,6 +117,31 @@ def run_glmasr(question: str, audio_count: int) -> ModelRequestData:
117 )
118
119
120 +# FunAudioChat
121 +def run_funaudiochat(question: str, audio_count: int) -> ModelRequestData:
122 + # NOTE: FunAudioChat is not available on the HuggingFace Hub at the time of
123 + # writing. Pass a local model path via `--model`.
124 + model_name = "funaudiochat"
125 +
126 + engine_args = EngineArgs(
127 + model=model_name,
128 + max_model_len=4096,
129 + max_num_seqs=2,
130 + limit_mm_per_prompt={"audio": audio_count},
131 + enforce_eager=True,
132 + )
133 +
134 + audio_in_prompt = "".join(
135 + ["<|audio_bos|><|AUDIO|><|audio_eos|>\n" for _ in range(audio_count)]
136 + )
137 + prompt = f"{audio_in_prompt}{question}"
138 +
139 + return ModelRequestData(
140 + engine_args=engine_args,
141 + prompt=prompt,
142 + )
143 +
144 +
145 # Granite Speech
146 def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:
147 # NOTE - the setting in this example are somewhat different from what is
@@ -410,6 +435,7 @@ model_example_map = {
435 "audioflamingo3": run_audioflamingo3,
436 "gemma3n": run_gemma3n,
437 "glmasr": run_glmasr,
438 + "funaudiochat": run_funaudiochat,
439 "granite_speech": run_granite_speech,
440 "midashenglm": run_midashenglm,
441 "minicpmo": run_minicpmo,
@@ -435,6 +461,12 @@ def parse_args():
461 choices=model_example_map.keys(),
462 help='Huggingface "model_type".',
463 )
464 + parser.add_argument(
465 + "--model",
466 + type=str,
467 + default=None,
468 + help="Model ID or local path override. Required for funaudiochat.",
469 + )
470 parser.add_argument(
471 "--num-prompts", type=int, default=1, help="Number of prompts to run."
472 )
@@ -467,6 +499,9 @@ def main(args):
499 if model not in model_example_map:
500 raise ValueError(f"Model type {model} is not supported.")
501
502 + if model == "funaudiochat" and not args.model:
503 + raise ValueError("--model is required when --model-type=funaudiochat")
504 +
505 if args.tensor_parallel_size is not None and args.tensor_parallel_size < 1:
506 raise ValueError(
507 f"tensor_parallel_size must be a positive integer, "
@@ -477,6 +512,8 @@ def main(args):
512 req_data = model_example_map[model](
513 question_per_audio_count[audio_count], audio_count
514 )
515 + if model == "funaudiochat":
516 + req_data.engine_args.model = args.model
517
518 # Disable other modalities to save memory
519 default_limits = {"image": 0, "video": 0, "audio": 0}
tests/models/registry.py
+3
@@ -692,6 +692,9 @@ _MULTIMODAL_EXAMPLE_MODELS = {
692 "baidu/ERNIE-4.5-VL-28B-A3B-PT",
693 trust_remote_code=True,
694 ),
695 + "FunAudioChatForConditionalGeneration": _HfExamplesInfo(
696 + "funaudiochat", is_available_online=False
697 + ),
698 "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
699 "Gemma3ForConditionalGeneration": _HfExamplesInfo("google/gemma-3-4b-it"),
700 "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it"),
vllm/model_executor/models/funaudiochat.py new
+1083
@@ -0,0 +1,1083 @@
1 +# SPDX-License-Identifier: Apache-2.0
2 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3 +
4 +"""Inference-only FunAudioChat model compatible with HuggingFace weights.
5 +
6 +FunAudioChat is a Qwen3 text model augmented with:
7 + - a continuous audio encoder (Whisper-mel frontend + transformer)
8 + - a discrete audio encoder (speech tokenizer + projector)
9 +
10 +In the HF implementation, audio features are scattered into `<|AUDIO|>` token
11 +positions via `inputs_embeds`, while `position_ids` (RoPE) remains standard 1D.
12 +"""
13 +
14 +from __future__ import annotations
15 +
16 +import os
17 +from collections.abc import Iterable, Mapping, Sequence
18 +from functools import cached_property
19 +from typing import Any
20 +
21 +import numpy as np
22 +import torch
23 +import torch.nn as nn
24 +from transformers import PreTrainedTokenizerFast, WhisperFeatureExtractor
25 +from transformers.activations import get_activation
26 +from transformers.feature_extraction_utils import BatchFeature
27 +from transformers.modeling_outputs import BaseModelOutput
28 +
29 +from vllm.config import VllmConfig
30 +from vllm.config.multimodal import BaseDummyOptions
31 +from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
32 +from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
33 +from vllm.model_executor.model_loader.weight_utils import default_weight_loader
34 +from vllm.multimodal import MULTIMODAL_REGISTRY
35 +from vllm.multimodal.inputs import (
36 + MultiModalDataDict,
37 + MultiModalFieldConfig,
38 + MultiModalKwargsItems,
39 +)
40 +from vllm.multimodal.parse import (
41 + AudioProcessorItems,
42 + MultiModalDataItems,
43 + MultiModalDataParser,
44 +)
45 +from vllm.multimodal.processing import (
46 + BaseDummyInputsBuilder,
47 + BaseMultiModalProcessor,
48 + BaseProcessingInfo,
49 + PromptReplacement,
50 + PromptUpdate,
51 + PromptUpdateDetails,
52 +)
53 +from vllm.sequence import IntermediateTensors
54 +from vllm.utils.import_utils import _has_module
55 +
56 +from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
57 +from .utils import AutoWeightsLoader, init_vllm_registered_model, maybe_prefix
58 +
59 +
60 +class _SinusoidsPositionEmbedding(nn.Module):
61 + def __init__(self, length: int, channels: int, max_timescale: float = 10000.0):
62 + super().__init__()
63 + if channels % 2 != 0:
64 + raise ValueError("SinusoidsPositionEmbedding needs even channels input")
65 +
66 + log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
67 + inv_timescales = torch.exp(
68 + -log_timescale_increment * torch.arange(channels // 2).float()
69 + )
70 + scaled_time = (
71 + torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
72 + )
73 + self.register_buffer(
74 + "positional_embedding",
75 + torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1),
76 + persistent=False,
77 + )
78 +
79 +
80 +class FunAudioChatAudioAttention(nn.Module):
81 + """Multi-headed attention used inside the continuous audio tower."""
82 +
83 + def __init__(self, config: Any):
84 + super().__init__()
85 + self.embed_dim = int(config.d_model)
86 + self.total_num_heads = int(config.encoder_attention_heads)
87 + self.dropout = float(getattr(config, "attention_dropout", 0.0))
88 + self.head_dim = self.embed_dim // self.total_num_heads
89 + self.num_key_value_groups = 1 # needed for eager attention
90 + self.config = config
91 +
92 + if self.head_dim * self.total_num_heads != self.embed_dim:
93 + raise ValueError(
94 + "embed_dim must be divisible by num_heads "
95 + f"(got embed_dim={self.embed_dim}, "
96 + f"num_heads={self.total_num_heads})."
97 + )
98 + self.scaling = self.head_dim**-0.5
99 + self.attention_dropout = 0.0
100 + self.is_decoder = False
101 + self.is_causal = False
102 +
103 + self.qkv_proj = QKVParallelLinear(
104 + self.embed_dim,
105 + self.head_dim,
106 + self.total_num_heads,
107 + bias=True,
108 + )
109 + self.num_heads = self.qkv_proj.num_heads
110 + self.num_kv_heads = self.qkv_proj.num_kv_heads
111 + self.q_size = self.num_heads * self.head_dim
112 + self.kv_size = self.num_kv_heads * self.head_dim
113 +
114 + self.attn = MMEncoderAttention(
115 + num_heads=self.num_heads,
116 + head_size=self.head_dim,
117 + scale=self.scaling,
118 + num_kv_heads=self.num_kv_heads,
119 + prefix="funaudiochat_audio_tower.attn",
120 + )
121 + self.out_proj = RowParallelLinear(
122 + self.embed_dim,
123 + self.embed_dim,
124 + bias=True,
125 + )
126 +
127 + def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
128 + stacked_params_mapping = [
129 + ("qkv_proj", "q_proj", "q"),
130 + ("qkv_proj", "k_proj", "k"),
131 + ("qkv_proj", "v_proj", "v"),
132 + ]
133 +
134 + params_dict = dict(self.named_parameters())
135 + with torch.no_grad():
136 + if self.qkv_proj.bias is not None:
137 + # HF FunAudioChat uses bias=False for k_proj. Ensure the missing
138 + # shard starts as zeros, while allowing q/v shards to load.
139 + self.qkv_proj.bias.zero_()
140 +
141 + loaded_params: set[str] = set()
142 + for name, loaded_weight in weights:
143 + for param_name, shard_name, shard_id in stacked_params_mapping:
144 + if shard_name not in name:
145 + continue
146 + name = name.replace(shard_name, param_name)
147 + param = params_dict[name]
148 + weight_loader = getattr(param, "weight_loader", default_weight_loader)
149 + weight_loader(param, loaded_weight, shard_id)
150 + break
151 + else:
152 + # Skip loading extra bias for GPTQ models.
153 + if name.endswith(".bias") and name not in params_dict:
154 + continue
155 +
156 + param = params_dict[name]
157 + weight_loader = getattr(param, "weight_loader", default_weight_loader)
158 + weight_loader(param, loaded_weight)
159 +
160 + loaded_params.add(name)
161 +
162 + return loaded_params
163 +
164 + def forward(
165 + self,
166 + hidden_states: torch.Tensor,
167 + cu_seqlens: torch.Tensor | None = None,
168 + attention_mask: torch.Tensor | None = None,
169 + **kwargs: object,
170 + ) -> torch.Tensor:
171 + del kwargs
172 + del attention_mask
173 + seq_length, _ = hidden_states.size()
174 +
175 + qkv, _ = self.qkv_proj(hidden_states)
176 + query_states, key_states, value_states = qkv.split(
177 + [self.q_size, self.kv_size, self.kv_size], dim=-1
178 + )
179 +
180 + max_seqlen: torch.Tensor | None = None
181 + if cu_seqlens is not None:
182 + max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
183 +
184 + attn_output = self.attn(
185 + query_states.reshape(1, seq_length, self.q_size),
186 + key_states.reshape(1, seq_length, self.kv_size),
187 + value_states.reshape(1, seq_length, self.kv_size),
188 + cu_seqlens=cu_seqlens,
189 + max_seqlen=max_seqlen,
190 + ).reshape(seq_length, -1)
191 +
192 + output, _ = self.out_proj(attn_output)
193 + return output
194 +
195 +
196 +class FunAudioChatAudioEncoderLayer(nn.Module):
197 + def __init__(self, config: Any):
198 + super().__init__()
199 + self.embed_dim = int(config.d_model)
200 + self.self_attn = FunAudioChatAudioAttention(config)
201 + self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
202 + self.dropout = float(config.dropout)
203 + self.activation_fn = get_activation(str(config.activation_function))
204 + self.activation_dropout = float(config.activation_dropout)
205 + self.fc1 = nn.Linear(self.embed_dim, int(config.encoder_ffn_dim))
206 + self.fc2 = nn.Linear(int(config.encoder_ffn_dim), self.embed_dim)
207 + self.final_layer_norm = nn.LayerNorm(self.embed_dim)
208 +
209 + def forward(
210 + self,
211 + hidden_states: torch.Tensor,
212 + cu_seqlens: torch.Tensor,
213 + attention_mask: torch.Tensor | None = None,
214 + **kwargs: object,
215 + ) -> tuple[torch.Tensor]:
216 + residual = hidden_states
217 + hidden_states = self.self_attn_layer_norm(hidden_states)
218 + hidden_states = self.self_attn(
219 + hidden_states=hidden_states,
220 + cu_seqlens=cu_seqlens,
221 + attention_mask=attention_mask,
222 + **kwargs,
223 + )
224 + hidden_states = residual + hidden_states
225 +
226 + residual = hidden_states
227 + hidden_states = self.final_layer_norm(hidden_states)
228 + hidden_states = self.activation_fn(self.fc1(hidden_states))
229 + hidden_states = nn.functional.dropout(
230 + hidden_states, p=self.activation_dropout, training=self.training
231 + )
232 + hidden_states = self.fc2(hidden_states)
233 + hidden_states = nn.functional.dropout(
234 + hidden_states, p=self.dropout, training=self.training
235 + )
236 + hidden_states = residual + hidden_states
237 +
238 + return (hidden_states,)
239 +
240 +
241 +class FunAudioChatAudioEncoder(nn.Module):
242 + """Continuous audio tower."""
243 +
244 + def __init__(self, config: Any):
245 + super().__init__()
246 + self.config = config
247 +
248 + embed_dim = int(config.d_model)
249 + self.num_mel_bins = int(config.num_mel_bins)
250 + self.max_source_positions = int(config.max_source_positions)
251 + self.embed_scale = (embed_dim**0.5) if bool(config.scale_embedding) else 1.0
252 + self.n_window = int(config.n_window)
253 +
254 + self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
255 + self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)
256 + self.layers = nn.ModuleList(
257 + [
258 + FunAudioChatAudioEncoderLayer(config)
259 + for _ in range(int(config.encoder_layers))
260 + ]
261 + )
262 + self.ln_post = nn.LayerNorm(embed_dim)
263 + self.avg_pooler = nn.AvgPool1d(2, stride=2)
264 + self.proj = nn.Linear(embed_dim, int(config.output_dim))
265 + self.positional_embedding = _SinusoidsPositionEmbedding(
266 + self.max_source_positions, embed_dim
267 + )
268 +
269 + # Present in HF weights even if unused during S2T.
270 + self.audio_bos_eos_token = nn.Embedding(2, int(config.output_dim))
271 +
272 + @property
273 + def dtype(self) -> torch.dtype:
274 + return self.conv1.weight.dtype
275 +
276 + def _prepare_attention_mask(
277 + self, inputs_tensor: torch.Tensor, cu_seqlens: torch.Tensor
278 + ) -> torch.Tensor | None:
279 + if getattr(self.config, "_attn_implementation", "eager") == "flash_attention_2":
280 + return None
281 +
282 + seq_length = inputs_tensor.shape[0]
283 + attention_mask = torch.full(
284 + (1, 1, seq_length, seq_length),
285 + torch.finfo(inputs_tensor.dtype).min,
286 + device=inputs_tensor.device,
287 + dtype=inputs_tensor.dtype,
288 + )
289 + for i in range(1, len(cu_seqlens)):
290 + start = int(cu_seqlens[i - 1].item())
291 + end = int(cu_seqlens[i].item())
292 + attention_mask[..., start:end, start:end] = 0
293 + return attention_mask
294 +
295 + def forward(
296 + self,
297 + input_features: torch.Tensor,
298 + feature_lens: torch.Tensor,
299 + aftercnn_lens: torch.Tensor,
300 + speech_maxlen: int,
301 + **kwargs: object,
302 + ) -> BaseModelOutput:
303 + # For max-length audio (300s => ~7500 speech frames at 25Hz), the
304 + # Torch SDPA path can be prohibitively memory hungry (~O(n^2) inside the
305 + # longest chunks). Require FlashAttention for such inputs to avoid OOM
306 + # and performance cliffs.
307 + if int(speech_maxlen) >= 7500:
308 + if not _has_module("flash_attn"):
309 + raise RuntimeError(
310 + "FunAudioChat long audio (~300s) requires FlashAttention-2 "
311 + "for the continuous audio tower, but `flash_attn` is not "
312 + "installed in the runtime environment."
313 + )
314 + if not getattr(
315 + self.layers[0].self_attn.attn, "is_flash_attn_backend", False
316 + ):
317 + raise RuntimeError(
318 + "FunAudioChat long audio (~300s) requires FlashAttention for the "
319 + "continuous audio tower, but the selected MM encoder attention "
320 + "backend is not FlashAttention."
321 + )
322 +
323 + # Handle empty / invalid items (feature_lens == 0) without crashing.
324 + original_batch_size = int(feature_lens.size(0))
325 + device = input_features.device
326 +
327 + valid_mask = feature_lens > 0
328 + valid_indices = torch.where(valid_mask)[0]
329 +
330 + if valid_indices.numel() == 0:
331 + output_dim = int(self.proj.out_features)
332 + return BaseModelOutput(
333 + last_hidden_state=torch.zeros(
334 + (original_batch_size, speech_maxlen, output_dim),
335 + device=device,
336 + dtype=self.proj.weight.dtype,
337 + )
338 + )
339 +
340 + input_features_list = input_features.split(feature_lens.tolist(), dim=1)
341 + valid_input_features_list = [input_features_list[int(i)] for i in valid_indices]
342 + valid_input_features = torch.cat(valid_input_features_list, dim=1)
343 +
344 + valid_feature_lens = feature_lens[valid_mask]
345 + valid_aftercnn_lens = aftercnn_lens[valid_mask]
346 +
347 + chunk_num = torch.ceil(valid_feature_lens / (self.n_window * 2)).long()
348 +
349 + chunk_lengths_list: list[int] = []
350 + full_chunk_len = self.n_window * 2
351 + for i, length in enumerate(valid_feature_lens):
352 + num_chunks_for_sample = int(chunk_num[i].item())
353 + if num_chunks_for_sample == 0:
354 + continue
355 + chunk_lengths_list.extend([full_chunk_len] * (num_chunks_for_sample - 1))
356 + last_chunk_len = int(length.item()) % full_chunk_len
357 + if last_chunk_len == 0:
358 + last_chunk_len = full_chunk_len
359 + chunk_lengths_list.append(last_chunk_len)
360 +
361 + chunk_lengths = torch.tensor(
362 + chunk_lengths_list, dtype=torch.long, device=device
363 + )
364 +
365 + chunk_list = valid_input_features.split(chunk_lengths.tolist(), dim=1)
366 + padded_feature, padded_mask, padded_mask_after_cnn = (
367 + self.padded_and_mask_function(
368 + chunk_list, chunk_lengths, padding_value=0, padding_side="right"
369 + )
370 + )
371 +
372 + padded_embed = nn.functional.gelu(self.conv1(padded_feature)) * padded_mask
373 + padded_embed = nn.functional.gelu(self.conv2(padded_embed)).transpose(1, 2)
374 +
375 + padded_embed = padded_embed + self.positional_embedding.positional_embedding[
376 + : padded_embed.shape[1], :
377 + ].unsqueeze(0).to(padded_embed.dtype)
378 +
379 + hidden_states = padded_embed[padded_mask_after_cnn]
380 + cu_seqlens = torch.cat(
381 + (
382 + torch.zeros(1, device=padded_mask_after_cnn.device, dtype=torch.int32),
383 + padded_mask_after_cnn.sum(1).cumsum(0),
384 + )
385 + ).to(torch.int32)
386 +
387 + for encoder_layer in self.layers:
388 + (hidden_states,) = encoder_layer(
389 + hidden_states,
390 + cu_seqlens=cu_seqlens,
391 + **kwargs,
392 + )
393 +
394 + hidden_states_list = hidden_states.split(valid_aftercnn_lens.tolist(), dim=0)
395 +
396 + pooled_list: list[torch.Tensor] = []
397 + pooled_lengths: list[int] = []
398 + for each_audio_states in hidden_states_list:
399 + seq_len = int(each_audio_states.shape[0])
400 + if seq_len >= 2:
401 + pooled = nn.functional.avg_pool1d(
402 + each_audio_states.transpose(0, 1), kernel_size=2, stride=2
403 + ).transpose(0, 1)
404 + else:
405 + pooled = each_audio_states
406 + pooled_list.append(pooled)
407 + pooled_lengths.append(int(pooled.shape[0]))
408 +
409 + pooled_concat = torch.cat(pooled_list, dim=0)
410 + processed_concat = self.proj(self.ln_post(pooled_concat))
411 + processed_audio_list = list(processed_concat.split(pooled_lengths, dim=0))
412 +
413 + output_dim = (
414 + int(processed_audio_list[0].shape[-1])
415 + if processed_audio_list
416 + else int(self.proj.out_features)
417 + )
418 + output_hidden_states = torch.zeros(
419 + (original_batch_size, speech_maxlen, output_dim),
420 + dtype=processed_audio_list[0].dtype
421 + if processed_audio_list
422 + else self.proj.weight.dtype,
423 + device=device,
424 + )
425 +
426 + for valid_idx, processed in zip(valid_indices, processed_audio_list):
427 + seq_len = min(int(processed.shape[0]), int(speech_maxlen))
428 + output_hidden_states[int(valid_idx), :seq_len] = processed[:seq_len]
429 +
430 + return BaseModelOutput(last_hidden_state=output_hidden_states)
431 +
432 + def padded_and_mask_function(
433 + self,
434 + tensor_list: Sequence[torch.Tensor],
435 + tensor_len: torch.Tensor,
436 + padding_value: float = 0.0,
437 + padding_side: str = "right",
438 + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
439 + max_len = int(tensor_len.max().item())
440 + dim = int(tensor_list[0].shape[0])
441 + padded_tensor = torch.full(
442 + size=(len(tensor_list), dim, max_len),
443 + fill_value=padding_value,
444 + dtype=self.dtype,
445 + device=tensor_list[0].device,
446 + )
447 +
448 + batch_mask = torch.zeros(
449 + (len(tensor_len), max_len), dtype=torch.long, device=padded_tensor.device
450 + )
451 + for i, length in enumerate(tensor_len):
452 + length_val = int(length.item())
453 + batch_mask[i, :length_val] = 1
454 + padded_tensor[i, :, :length_val] = tensor_list[i]
455 +
456 + feature_lens_after_cnn = (tensor_len - 1) // 2 + 1
457 + max_len_after_cnn = int(feature_lens_after_cnn.max().item())
458 + batch_mask_after_cnn = torch.zeros(
459 + (len(tensor_len), max_len_after_cnn),
460 + dtype=torch.long,
461 + device=padded_tensor.device,
462 + )
463 + for i, length in enumerate(feature_lens_after_cnn):
464 + batch_mask_after_cnn[i, : int(length.item())] = 1
465 +
466 + if padding_side != "right":
467 + raise NotImplementedError("Only right padding is supported.")
468 +
469 + return (
470 + padded_tensor,
471 + batch_mask.unsqueeze(1).to(padded_tensor.dtype),
472 + batch_mask_after_cnn.bool(),
473 + )
474 +
475 + # From the HF FunAudioChat implementation.
476 + def _get_feat_extract_output_lengths(
477 + self, input_lengths: torch.LongTensor
478 + ) -> tuple[torch.LongTensor, torch.LongTensor]:
479 + input_lengths = (input_lengths - 1) // 2 + 1
480 + output_lengths = (input_lengths - 2) // 2 + 1
481 + return input_lengths, output_lengths
482 +
483 +
484 +class FunAudioChatDiscreteEncoder(nn.Module):
485 + """Discrete audio encoder (speech tokenizer -> grouped embeddings)."""
486 +
487 + def __init__(self, config: Any):
488 + super().__init__()
489 + self.padding_idx = int(config.pad_token_id)
490 + self.group_size = int(config.group_size)
491 + self.hidden_size = int(config.output_dim)
492 + self.continuous_features_mode = getattr(
493 + config, "continuous_features_mode", "add"
494 + )
495 + self.embed_tokens = nn.Embedding(
496 + int(config.codebook_size), self.hidden_size, self.padding_idx
497 + )
498 + self.output_matching = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
499 + self.continual_output_matching = nn.Linear(
500 + self.hidden_size, self.hidden_size, bias=False
501 + )
502 +
503 + def forward(
504 + self,
505 + audio_ids: torch.Tensor,
506 + continuous_audio_features: torch.Tensor | None = None,
507 + continuous_audio_output_lengths: torch.Tensor | None = None,
508 + feature_exist_mask: torch.Tensor | None = None,
509 + ) -> torch.Tensor:
510 + del continuous_audio_output_lengths
511 +
512 + inputs_embeds = self.embed_tokens(audio_ids)
513 + hidden_states = inputs_embeds.reshape(
514 + inputs_embeds.shape[0], -1, self.group_size * self.hidden_size
515 + )
516 + hidden_states = hidden_states.reshape(
517 + hidden_states.shape[0], -1, self.group_size, self.hidden_size
518 + ).mean(dim=2)
519 + hidden_states = self.output_matching(hidden_states)
520 +
521 + if continuous_audio_features is not None:
522 + continuous_audio_features = continuous_audio_features.reshape(
523 + continuous_audio_features.shape[0],
524 + -1,
525 + self.group_size,
526 + self.hidden_size,
527 + ).mean(dim=2)
528 + continuous_audio_hidden_states = self.continual_output_matching(
529 + continuous_audio_features
530 + )
531 +
532 + if feature_exist_mask is None:
533 + feature_exist_mask = torch.ones(
534 + (hidden_states.shape[0],),
535 + dtype=torch.bool,
536 + device=hidden_states.device,
537 + )
538 + if self.continuous_features_mode == "add":
539 + hidden_states[feature_exist_mask] += continuous_audio_hidden_states
540 + else:
541 + hidden_states[feature_exist_mask] = continuous_audio_hidden_states
542 +
543 + return hidden_states
544 +
545 + def _get_feat_extract_output_lengths(
546 + self, input_lengths: torch.LongTensor
547 + ) -> tuple[torch.LongTensor, torch.LongTensor]:
548 + output_lengths = (input_lengths + self.group_size - 1) // self.group_size
549 + return input_lengths, output_lengths
550 +
551 +
552 +class FunAudioChatProcessingInfo(BaseProcessingInfo):
553 + token_fps: int = 25
554 +
555 + def get_supported_mm_limits(self) -> Mapping[str, int | None]:
556 + return {"audio": None}
557 +
558 + def get_target_channels(self) -> int:
559 + return 1
560 +
561 + def get_mm_max_tokens_per_item(
562 + self,
563 + seq_len: int,
564 + mm_counts: Mapping[str, int],
565 + ) -> Mapping[str, int] | None:
566 + # The discrete audio encoder downsamples 25Hz frames with group_size=5,
567 + # so for a 300s clip the max number of `<|AUDIO|>` placeholders is 1500.
568 + cfg = self.get_hf_config()
569 + audio_cfg = getattr(cfg, "audio_config", None)
570 + max_audio_tokens = int(getattr(audio_cfg, "max_source_positions", 1500))
571 + return {"audio": max_audio_tokens}
572 +
573 + @cached_property
574 + def feature_extractor(self) -> WhisperFeatureExtractor:
575 + return WhisperFeatureExtractor.from_pretrained(self.model_id)
576 +
577 + @cached_property
578 + def speech_tokenizer(self) -> PreTrainedTokenizerFast:
579 + return PreTrainedTokenizerFast.from_pretrained(
580 + self.model_id, subfolder="speech_tokenizer"
581 + )
582 +
583 + def get_feature_extractor(self) -> WhisperFeatureExtractor:
584 + return self.feature_extractor
585 +
586 + def get_speech_tokenizer(self) -> PreTrainedTokenizerFast:
587 + return self.speech_tokenizer
588 +
589 + def get_audio_group_size(self) -> int:
590 + cfg = self.get_hf_config()
591 + audio_cfg = getattr(cfg, "audio_config", None)
592 + return int(getattr(audio_cfg, "group_size", 5))
593 +
594 +
595 +class FunAudioChatDummyInputsBuilder(
596 + BaseDummyInputsBuilder[FunAudioChatProcessingInfo]
597 +):
598 + def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
599 + num_audios = mm_counts.get("audio", 0)
600 + return "<|audio_bos|><|AUDIO|><|audio_eos|>" * int(num_audios)
601 +
602 + def get_dummy_mm_data(
603 + self,
604 + seq_len: int,
605 + mm_counts: Mapping[str, int],
606 + mm_options: Mapping[str, BaseDummyOptions] | None = None,
607 + ) -> MultiModalDataDict:
608 + feature_extractor = self.info.get_feature_extractor()
609 + sampling_rate = int(feature_extractor.sampling_rate)
610 +
611 + # Dummy inputs are used for profiling; construct the worst-case audio
612 + # length that maximizes the number of encoder tokens.
613 + cfg = self.info.get_hf_config()
614 + audio_cfg = getattr(cfg, "audio_config", None)
615 + max_audio_tokens = int(getattr(audio_cfg, "max_source_positions", 1500))
616 + group_size = self.info.get_audio_group_size()
617 + token_fps = int(getattr(self.info, "token_fps", 25))
618 + target_num_frames = max(1, max_audio_tokens) * max(1, group_size)
619 + audio_len = max(
620 + 1,
621 + (target_num_frames * sampling_rate + token_fps - 1) // token_fps,
622 + )
623 + num_audios = int(mm_counts.get("audio", 0))
624 +
625 + audio_overrides = mm_options.get("audio") if mm_options else None
626 + return {
627 + "audio": self._get_dummy_audios(
628 + length=audio_len,
629 + num_audios=num_audios,
630 + overrides=audio_overrides,
631 + )
632 + }
633 +
634 +
635 +class FunAudioChatMultiModalProcessor(
636 + BaseMultiModalProcessor[FunAudioChatProcessingInfo]
637 +):
638 + def _get_data_parser(self) -> MultiModalDataParser:
639 + feature_extractor = self.info.get_feature_extractor()
640 + return MultiModalDataParser(
641 + target_sr=int(feature_extractor.sampling_rate),
642 + target_channels=self.info.get_target_channels(),
643 + )
644 +
645 + def _call_hf_processor(
646 + self,
647 + prompt: str,
648 + mm_data: Mapping[str, object],
649 + mm_kwargs: Mapping[str, object],
650 + tok_kwargs: Mapping[str, object],
651 + ) -> BatchFeature:
652 + tokenizer = self.info.get_tokenizer()
653 + input_ids = torch.tensor([tokenizer.encode(prompt, **tok_kwargs)])
654 +
655 + audios = mm_data.get("audios", [])
656 + if not audios:
657 + return BatchFeature({"input_ids": input_ids})
658 +
659 + feature_extractor = self.info.get_feature_extractor()
660 + sr = int(feature_extractor.sampling_rate)
661 + min_samples = int(getattr(feature_extractor, "n_fft", 400) or 400)
662 +
663 + wavs: list[np.ndarray] = []
664 + speech_strs: list[str] = []
665 +
666 + speech_tokenizer = self.info.get_speech_tokenizer()
667 + pad_token = speech_tokenizer.pad_token or "<|audio_pad|>"
668 + for audio in audios:
669 + if isinstance(audio, torch.Tensor):
670 + audio = audio.detach().cpu().numpy()
671 + audio_np = np.asarray(audio, dtype=np.float32)
672 +
673 + if min_samples > 0 and audio_np.shape[0] < min_samples:
674 + audio_np = np.pad(
675 + audio_np, (0, min_samples - audio_np.shape[0]), mode="constant"
676 + )
677 +
678 + wavs.append(audio_np)
679 + num_frames = int(
680 + (float(audio_np.shape[0]) / float(sr)) * float(self.info.token_fps)
681 + )
682 + speech_strs.append(pad_token * max(1, int(num_frames)))
683 +
684 + audio_group_size = self.info.get_audio_group_size()
685 + speech_inputs = speech_tokenizer(
686 + speech_strs,
687 + return_attention_mask=True,
688 + return_token_type_ids=False,
689 + padding=True,
690 + pad_to_multiple_of=audio_group_size,
691 + return_tensors="pt",
692 + )
693 +
694 + wav_inputs = feature_extractor(
695 + wavs,
696 + sampling_rate=sr,
697 + return_attention_mask=True,
698 + padding="max_length",
699 + return_tensors="pt",
700 + )
701 +
702 + mm_inputs: dict[str, torch.Tensor] = {
703 + "speech_ids": speech_inputs["input_ids"],
704 + "speech_attention_mask": speech_inputs["attention_mask"],
705 + "input_features": wav_inputs["input_features"],
706 + "feature_attention_mask": wav_inputs["attention_mask"],
707 + "feature_exist_mask": torch.ones((len(wavs),), dtype=torch.bool),
708 + }
709 +
710 + return BatchFeature({"input_ids": input_ids, **mm_inputs})
711 +
712 + def _hf_processor_applies_updates(
713 + self,
714 + prompt_text: str,
715 + mm_items: MultiModalDataItems,
716 + hf_processor_mm_kwargs: Mapping[str, object],
717 + tokenization_kwargs: Mapping[str, object],
718 + ) -> bool:
719 + return False
720 +
721 + def _get_mm_fields_config(
722 + self,
723 + hf_inputs: BatchFeature,
724 + hf_processor_mm_kwargs: Mapping[str, object],
725 + ) -> Mapping[str, MultiModalFieldConfig]:
726 + return {
727 + "speech_ids": MultiModalFieldConfig.batched("audio"),
728 + "speech_attention_mask": MultiModalFieldConfig.batched("audio"),
729 + "input_features": MultiModalFieldConfig.batched("audio"),
730 + "feature_attention_mask": MultiModalFieldConfig.batched("audio"),
731 + "feature_exist_mask": MultiModalFieldConfig.batched("audio"),
732 + }
733 +
734 + def _get_prompt_updates(
735 + self,
736 + mm_items: MultiModalDataItems,
737 + hf_processor_mm_kwargs: Mapping[str, object],
738 + out_mm_kwargs: MultiModalKwargsItems,
739 + ) -> Sequence[PromptUpdate]:
740 + tokenizer = self.info.get_tokenizer()
741 + vocab = tokenizer.get_vocab()
742 +
743 + audio_token = "<|AUDIO|>"
744 + audio_token_id = vocab[audio_token]
745 +
746 + out_mm_data = out_mm_kwargs.get_data()
747 + speech_attention_mask = out_mm_data.get("speech_attention_mask")
748 + if speech_attention_mask is None:
749 + audio_output_lengths: list[int] = []
750 + else:
751 + assert isinstance(speech_attention_mask, torch.Tensor)
752 + speech_lengths = speech_attention_mask.sum(-1)
753 + group_size = self.info.get_audio_group_size()
754 + audio_output_lengths = (
755 + (speech_lengths + group_size - 1) // group_size
756 + ).tolist()
757 +
758 + def get_replacement_funaudiochat(item_idx: int):
759 + num_features = (
760 + int(audio_output_lengths[item_idx]) if audio_output_lengths else 1
761 + )
762 + if num_features <= 0:
763 + audios = mm_items.get_items("audio", AudioProcessorItems)
764 + audio_len = audios.get_audio_length(item_idx)
765 + raise ValueError(
766 + f"The audio (len={audio_len}) is too short to be "
767 + "represented inside the model"
768 + )
769 +
770 + audio_tokens = [audio_token_id] * num_features
771 + return PromptUpdateDetails.select_token_id(
772 + audio_tokens,
773 + embed_token_id=audio_token_id,
774 + )
775 +
776 + return [
777 + PromptReplacement(
778 + modality="audio",
779 + target=audio_token,
780 + replacement=get_replacement_funaudiochat,
781 + )
782 + ]
783 +
784 +
785 +@MULTIMODAL_REGISTRY.register_processor(
786 + FunAudioChatMultiModalProcessor,
787 + info=FunAudioChatProcessingInfo,
788 + dummy_inputs=FunAudioChatDummyInputsBuilder,
789 +)
790 +class FunAudioChatForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
791 + @classmethod
792 + def get_placeholder_str(cls, modality: str, i: int) -> str | None:
793 + if modality.startswith("audio"):
794 + return "<|audio_bos|><|AUDIO|><|audio_eos|>"
795 +
796 + raise ValueError("Only audio modality is supported")
797 +
798 + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
799 + super().__init__()
800 + config = vllm_config.model_config.hf_config
801 + quant_config = vllm_config.quant_config
802 + multimodal_config = vllm_config.model_config.multimodal_config
803 + self.config = config
804 + self.multimodal_config = multimodal_config
805 + self.quant_config = quant_config
806 +
807 + with self._mark_tower_model(vllm_config, "audio"):
808 + self.continuous_audio_tower = FunAudioChatAudioEncoder(config.audio_config)
809 + self.audio_tower = FunAudioChatDiscreteEncoder(config.audio_config)
810 +
811 + with self._mark_language_model(vllm_config):
812 + self.language_model = init_vllm_registered_model(
813 + vllm_config=vllm_config,
814 + hf_config=config.text_config,
815 + prefix=maybe_prefix(prefix, "language_model"),
816 + architectures=["Qwen3ForCausalLM"],
817 + )
818 +
819 + self.make_empty_intermediate_tensors = (
820 + self.language_model.make_empty_intermediate_tensors
821 + )
822 +
823 + def get_language_model(self) -> torch.nn.Module:
824 + return self.language_model
825 +
826 + def _get_continuous_audio_features(
827 + self,
828 + input_features: torch.Tensor,
829 + feature_attention_mask: torch.Tensor,
830 + speech_maxlen: int,
831 + ) -> tuple[torch.Tensor, torch.Tensor]:
832 + # Align mask and features to avoid indexing errors when padding differs.
833 + if (
834 + input_features.dim() == 3
835 + and feature_attention_mask.shape[1] != input_features.shape[-1]
836 + ):
837 + min_len = min(
838 + int(feature_attention_mask.shape[1]), int(input_features.shape[-1])
839 + )
840 + feature_attention_mask = feature_attention_mask[:, :min_len]
841 + input_features = input_features[:, :, :min_len]
842 +
843 + feature_lens = torch.sum(feature_attention_mask, dim=1)
844 +
845 + flat_features = input_features.permute(0, 2, 1)[
846 + feature_attention_mask.bool()
847 + ].permute(1, 0)
848 +
849 + audio_feat_lengths, audio_output_lengths = (
850 + self.continuous_audio_tower._get_feat_extract_output_lengths(feature_lens)
851 + )
852 +
853 + audio_outputs = self.continuous_audio_tower(
854 + flat_features,
855 + feature_lens=feature_lens,
856 + aftercnn_lens=audio_feat_lengths,
857 + speech_maxlen=speech_maxlen,
858 + )
859 + return audio_outputs.last_hidden_state, audio_output_lengths
860 +
861 + def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
862 + speech_ids = kwargs.get("speech_ids")
863 + speech_attention_mask = kwargs.get("speech_attention_mask")
864 + input_features = kwargs.get("input_features")
865 + feature_attention_mask = kwargs.get("feature_attention_mask")
866 + feature_exist_mask = kwargs.get("feature_exist_mask")
867 +
868 + if speech_ids is None:
869 + return []
870 +
871 + pad_id = int(getattr(self.audio_tower, "padding_idx", 0))
872 +
873 + if not isinstance(speech_ids, torch.Tensor):
874 + if (
875 + isinstance(speech_ids, (list, tuple))
876 + and len(speech_ids) > 0
877 + and all(isinstance(t, torch.Tensor) for t in speech_ids)
878 + ):
879 + speech_ids_tensors = []
880 + for t in speech_ids:
881 + if t.dim() == 2 and t.shape[0] == 1:
882 + t = t.squeeze(0)
883 + if t.dim() != 1:
884 + raise TypeError(
885 + "FunAudioChat speech_ids must be a 1D tensor per item "
886 + f"(got shape={tuple(t.shape)})"
887 + )
888 + speech_ids_tensors.append(t)
889 + speech_ids = nn.utils.rnn.pad_sequence(
890 + speech_ids_tensors,
891 + batch_first=True,
892 + padding_value=pad_id,
893 + )
894 + else:
895 + raise TypeError(
896 + "FunAudioChat speech_ids must be a Tensor or a sequence of Tensors "
897 + f"(got {type(speech_ids)})"
898 + )
899 +
900 + if speech_attention_mask is None:
901 + speech_attention_mask = speech_ids.ne(pad_id).to(dtype=torch.int64)
902 +
903 + if not isinstance(speech_attention_mask, torch.Tensor):
904 + if (
905 + isinstance(speech_attention_mask, (list, tuple))
906 + and len(speech_attention_mask) > 0
907 + and all(isinstance(t, torch.Tensor) for t in speech_attention_mask)
908 + ):
909 + mask_tensors = []
910 + for t in speech_attention_mask:
911 + if t.dim() == 2 and t.shape[0] == 1:
912 + t = t.squeeze(0)
913 + if t.dim() != 1:
914 + raise TypeError(
915 + "FunAudioChat speech_attention_mask must be a 1D tensor "
916 + f"per item (got shape={tuple(t.shape)})"
917 + )
918 + mask_tensors.append(t)
919 + speech_attention_mask = nn.utils.rnn.pad_sequence(
920 + mask_tensors,
921 + batch_first=True,
922 + padding_value=0,
923 + )
924 + else:
925 + raise TypeError(
926 + "FunAudioChat speech_attention_mask must be a Tensor or a "
927 + f"sequence of Tensors (got {type(speech_attention_mask)})"
928 + )
929 +
930 + debug = os.getenv("VLLM_FUN_AUDIOCHAT_DEBUG", "") == "1"
931 + if debug:
932 + print(
933 + f"[FunAudioChat] embed_multimodal speech_ids={tuple(speech_ids.shape)} "
934 + f"speech_attention_mask={tuple(speech_attention_mask.shape)}",
935 + flush=True,
936 + )
937 + attn_impl = getattr(
938 + self.continuous_audio_tower.config, "_attn_implementation", None
939 + )
940 + print(
941 + f"[FunAudioChat] audio_attn_impl={attn_impl}",
942 + flush=True,
943 + )
944 + if hasattr(self.continuous_audio_tower, "conv1"):
945 + conv1_w = self.continuous_audio_tower.conv1.weight
946 + print(
947 + f"[FunAudioChat] conv1_w_norm={float(conv1_w.norm().item()):.6g}",
948 + flush=True,
949 + )
950 + try:
951 + attn0 = self.continuous_audio_tower.layers[0].self_attn
952 + q_norm = float(attn0.q_proj.weight.norm().item())
953 + k_norm = float(attn0.k_proj.weight.norm().item())
954 + v_norm = float(attn0.v_proj.weight.norm().item())
955 + o_norm = float(attn0.out_proj.weight.norm().item())
956 + print(
957 + f"[FunAudioChat] attn0_q_norm={q_norm:.6g} "
958 + f"k_norm={k_norm:.6g} "
959 + f"v_norm={v_norm:.6g} "
960 + f"o_norm={o_norm:.6g}",
961 + flush=True,
962 + )
963 + except Exception:
964 + pass
965 + if isinstance(input_features, torch.Tensor):
966 + print(
967 + f"[FunAudioChat] input_features={tuple(input_features.shape)}",
968 + flush=True,
969 + )
970 + if isinstance(feature_attention_mask, torch.Tensor):
971 + print(
972 + "[FunAudioChat] feature_attention_mask="
973 + f"{tuple(feature_attention_mask.shape)}",
974 + flush=True,
975 + )
976 +
977 + group_size = int(self.audio_tower.group_size)
978 + speech_maxlen = int(speech_ids.shape[-1])
979 +
980 + # Ensure token length is divisible by group_size.
981 + target_len = ((speech_maxlen + group_size - 1) // group_size) * group_size
982 + if target_len > speech_maxlen:
983 + pad_id = int(self.audio_tower.padding_idx)
984 + pad_len = target_len - speech_maxlen
985 + speech_ids = nn.functional.pad(speech_ids, (0, pad_len), value=pad_id)
986 + speech_attention_mask = nn.functional.pad(
987 + speech_attention_mask, (0, pad_len), value=0
988 + )
989 + speech_maxlen = int(speech_ids.shape[-1])
990 +
991 + continuous_audio_features = None
992 + continuous_audio_output_lengths = None
993 + if input_features is not None and feature_attention_mask is not None:
994 + assert isinstance(input_features, torch.Tensor)
995 + assert isinstance(feature_attention_mask, torch.Tensor)
996 + continuous_audio_features, continuous_audio_output_lengths = (
997 + self._get_continuous_audio_features(
998 + input_features=input_features,
999 + feature_attention_mask=feature_attention_mask,
1000 + speech_maxlen=speech_maxlen,
1001 + )
1002 + )
1003 +
1004 + if feature_exist_mask is None:
1005 + feature_exist_mask = torch.ones(
1006 + (speech_ids.shape[0],), dtype=torch.bool, device=speech_ids.device
1007 + )
1008 + assert isinstance(feature_exist_mask, torch.Tensor)
1009 +
1010 + audio_features = self.audio_tower(
1011 + speech_ids,
1012 + continuous_audio_features=continuous_audio_features,
1013 + continuous_audio_output_lengths=continuous_audio_output_lengths,
1014 + feature_exist_mask=feature_exist_mask,
1015 + )
1016 +
1017 + _, audio_output_lengths = self.audio_tower._get_feat_extract_output_lengths(
1018 + speech_attention_mask.sum(-1)
1019 + )
1020 + lengths = audio_output_lengths.tolist()
1021 +
1022 + embeds = tuple(
1023 + audio_features[i, : int(length)] for i, length in enumerate(lengths)
1024 + )
1025 + if debug:
1026 + embed_lens = [int(t.shape[0]) for t in embeds]
1027 + print(f"[FunAudioChat] embed_multimodal out_lens={embed_lens}", flush=True)
1028 + if embeds:
1029 + t0 = embeds[0]
1030 + print(
1031 + f"[FunAudioChat] embed0 dtype={t0.dtype} device={t0.device} "
1032 + f"nan={bool(torch.isnan(t0).any())} "
1033 + f"norm={float(t0.norm().item()):.6g}",
1034 + flush=True,
1035 + )
1036 + dump_path = os.getenv("VLLM_FUN_AUDIOCHAT_DUMP_PATH", "")
1037 + if (
1038 + dump_path
1039 + and speech_ids.shape[0] == 1
1040 + and len(embeds) == 1
1041 + and embed_lens[0] > 10
1042 + ):
1043 + if not os.path.exists(dump_path):
1044 + np.save(dump_path, embeds[0].detach().float().cpu().numpy())
1045 + print(f"[FunAudioChat] dumped embeds to {dump_path}", flush=True)
1046 + cont_path = dump_path.replace(".npy", "_cont.npy")
1047 + if continuous_audio_features is not None and not os.path.exists(
1048 + cont_path
1049 + ):
1050 + np.save(
1051 + cont_path,
1052 + continuous_audio_features.detach().float().cpu().numpy(),
1053 + )
1054 + print(
1055 + f"[FunAudioChat] dumped continuous to {cont_path}", flush=True
1056 + )
1057 + return embeds
1058 +
1059 + def forward(
1060 + self,
1061 + input_ids: torch.Tensor,
1062 + positions: torch.Tensor,
1063 + intermediate_tensors: IntermediateTensors | None = None,
1064 + inputs_embeds: torch.Tensor | None = None,
1065 + **kwargs: object,
1066 + ) -> torch.Tensor | IntermediateTensors:
1067 + del kwargs
1068 + if intermediate_tensors is not None:
1069 + inputs_embeds = None
1070 +
1071 + return self.language_model.model(
1072 + input_ids,
1073 + positions,
1074 + intermediate_tensors,
1075 + inputs_embeds=inputs_embeds,
1076 + )
1077 +
1078 + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None:
1079 + return self.language_model.compute_logits(hidden_states)
1080 +
1081 + def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1082 + loader = AutoWeightsLoader(self, skip_prefixes=["audio_invert_tower."])
1083 + return loader.load_weights(weights)
vllm/model_executor/models/registry.py
+4
@@ -312,6 +312,10 @@ _MULTIMODAL_MODELS = {
312 "ernie45_vl",
313 "Ernie4_5_VLMoeForConditionalGeneration",
314 ),
315 + "FunAudioChatForConditionalGeneration": (
316 + "funaudiochat",
317 + "FunAudioChatForConditionalGeneration",
318 + ),
319 "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
320 "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"), # noqa: E501
321 "Gemma3nForConditionalGeneration": (
vllm/multimodal/audio.py
+8
@@ -1,5 +1,6 @@
1 # SPDX-License-Identifier: Apache-2.0
2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3 +import math
4 from dataclasses import dataclass
5 from enum import Enum
6 from typing import Literal
@@ -195,6 +196,13 @@ class AudioResampler:
196 raise RuntimeError(
197 "Audio resampling is not supported when `target_sr` is not provided"
198 )
199 + if math.isclose(
200 + float(orig_sr),
201 + float(self.target_sr),
202 + rel_tol=0.0,
203 + abs_tol=1e-6,
204 + ):
205 + return audio
206 if self.method == "librosa":
207 return resample_audio_librosa(
208 audio, orig_sr=orig_sr, target_sr=self.target_sr
vllm/transformers_utils/config.py
+1
@@ -77,6 +77,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
77 deepseek_vl_v2="DeepseekVLV2Config",
78 deepseek_v32="DeepseekV3Config",
79 flex_olmo="FlexOlmoConfig",
80 + funaudiochat="FunAudioChatConfig",
81 hunyuan_vl="HunYuanVLConfig",
82 isaac="IsaacConfig",
83 kimi_linear="KimiLinearConfig",
vllm/transformers_utils/configs/__init__.py
+4
@@ -22,6 +22,8 @@ _CLASS_TO_MODULE: dict[str, str] = {
22 "DotsOCRConfig": "vllm.transformers_utils.configs.dotsocr",
23 "EAGLEConfig": "vllm.transformers_utils.configs.eagle",
24 "FlexOlmoConfig": "vllm.transformers_utils.configs.flex_olmo",
25 + "FunAudioChatConfig": "vllm.transformers_utils.configs.funaudiochat",
26 + "FunAudioChatAudioEncoderConfig": "vllm.transformers_utils.configs.funaudiochat",
27 "HunYuanVLConfig": "vllm.transformers_utils.configs.hunyuan_vl",
28 "HunYuanVLTextConfig": "vllm.transformers_utils.configs.hunyuan_vl",
29 "HunYuanVLVisionConfig": "vllm.transformers_utils.configs.hunyuan_vl",
@@ -65,6 +67,8 @@ __all__ = [
67 "DotsOCRConfig",
68 "EAGLEConfig",
69 "FlexOlmoConfig",
70 + "FunAudioChatConfig",
71 + "FunAudioChatAudioEncoderConfig",
72 "HunYuanVLConfig",
73 "HunYuanVLTextConfig",
74 "HunYuanVLVisionConfig",
vllm/transformers_utils/configs/funaudiochat.py new
+124
@@ -0,0 +1,124 @@
1 +# SPDX-License-Identifier: Apache-2.0
2 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3 +
4 +from __future__ import annotations
5 +
6 +from transformers import PretrainedConfig
7 +
8 +# NOTE: Temporary shim for FunAudioChat checkpoints.
9 +# These checkpoints use `model_type="funaudiochat"`, which is not currently
10 +# recognized by released Transformers, and the public checkpoint does not
11 +# provide an `auto_map` to enable `trust_remote_code=True`.
12 +# Remove this file once Transformers adds native support (or the checkpoint
13 +# provides an `auto_map`) and vLLM can rely on `AutoConfig.from_pretrained()`.
14 +
15 +
16 +class FunAudioChatAudioEncoderConfig(PretrainedConfig):
17 + model_type = "funaudiochat_audio_encoder"
18 +
19 + def __init__(
20 + self,
21 + _attn_implementation: str | None = None,
22 + num_mel_bins: int = 128,
23 + encoder_layers: int = 32,
24 + encoder_attention_heads: int = 20,
25 + encoder_ffn_dim: int = 5120,
26 + d_model: int = 1280,
27 + dropout: float = 0.0,
28 + attention_dropout: float = 0.0,
29 + activation_function: str = "gelu",
30 + activation_dropout: float = 0.0,
31 + scale_embedding: bool = False,
32 + initializer_range: float = 0.02,
33 + max_source_positions: int = 1500,
34 + n_window: int = 100,
35 + output_dim: int = 3584,
36 + bos_token_id: int | None = None,
37 + codebook_size: int | None = None,
38 + continuous_features_mode: str = "replace",
39 + crq_transformer_config: dict | None = None,
40 + eos_token_id: int | None = None,
41 + group_size: int = 5,
42 + enable_audio_invert_tower: bool = True,
43 + pad_token_id: int | None = None,
44 + **kwargs,
45 + ) -> None:
46 + attn_impl = kwargs.pop("_attn_implementation", None) or _attn_implementation
47 + super().__init__(**kwargs)
48 + # Match HF default for attention implementation selection.
49 + self._attn_implementation = attn_impl or "sdpa"
50 +
51 + self.num_mel_bins = num_mel_bins
52 + self.d_model = d_model
53 + self.encoder_layers = encoder_layers
54 + self.encoder_attention_heads = encoder_attention_heads
55 + self.encoder_ffn_dim = encoder_ffn_dim
56 + self.dropout = dropout
57 + self.attention_dropout = attention_dropout
58 + self.activation_function = activation_function
59 + self.activation_dropout = activation_dropout
60 + self.num_hidden_layers = encoder_layers
61 + self.initializer_range = initializer_range
62 + self.scale_embedding = scale_embedding
63 + self.max_source_positions = max_source_positions
64 + self.n_window = n_window
65 + self.output_dim = output_dim
66 +
67 + self.bos_token_id = bos_token_id
68 + self.codebook_size = codebook_size
69 + self.continuous_features_mode = continuous_features_mode
70 + self.crq_transformer_config = crq_transformer_config
71 + self.eos_token_id = eos_token_id
72 + self.group_size = group_size
73 + self.enable_audio_invert_tower = enable_audio_invert_tower
74 + self.pad_token_id = pad_token_id
75 +
76 +
77 +class FunAudioChatConfig(PretrainedConfig):
78 + model_type = "funaudiochat"
79 + attribute_map = {
80 + "audio_token_id": "audio_token_index",
81 + }
82 +
83 + def __init__(
84 + self,
85 + audio_config: PretrainedConfig | dict | None = None,
86 + text_config: PretrainedConfig | dict | None = None,
87 + audio_token_index: int = 151646,
88 + ignore_index: int = -100,
89 + hidden_size: int | None = None,
90 + **kwargs,
91 + ) -> None:
92 + self.audio_token_index = audio_token_index
93 + self.ignore_index = ignore_index
94 +
95 + if isinstance(audio_config, dict):
96 + audio_config.setdefault(
97 + "model_type", FunAudioChatAudioEncoderConfig.model_type
98 + )
99 + audio_config = FunAudioChatAudioEncoderConfig(**audio_config)
100 + elif audio_config is None:
101 + audio_config = FunAudioChatAudioEncoderConfig()
102 + self.audio_config = audio_config
103 +
104 + if isinstance(text_config, dict):
105 + # Default to qwen2 for backwards compatibility; FunAudioChat uses
106 + # qwen3 in practice for recent checkpoints.
107 + text_config.setdefault("model_type", "qwen2")
108 + import transformers
109 +
110 + text_cls = transformers.CONFIG_MAPPING[text_config["model_type"]]
111 + text_config = text_cls(**text_config)
112 + elif text_config is None:
113 + import transformers
114 +
115 + text_config = transformers.CONFIG_MAPPING["qwen2"]()
116 + self.text_config = text_config
117 +
118 + self.hidden_size = (
119 + int(self.text_config.hidden_size)
120 + if hidden_size is None
121 + else int(hidden_size)
122 + )
123 +
124 + super().__init__(**kwargs)