audio_language.py 14.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""
4
This example shows how to use vLLM for running offline inference
5
with the correct prompt format on audio language models.
6
7
8
9

For most models, the prompt format should follow corresponding examples
on HuggingFace model repository.
"""
10

11
import os
12
from dataclasses import asdict
13
from typing import Any, NamedTuple
14
15

from huggingface_hub import snapshot_download
16
17
from transformers import AutoTokenizer

18
from vllm import LLM, EngineArgs, SamplingParams
19
from vllm.assets.audio import AudioAsset
20
from vllm.lora.request import LoRARequest
21
22
from vllm.utils import FlexibleArgumentParser

23
audio_assets = [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")]
24
25
26
question_per_audio_count = {
    0: "What is 1+1?",
    1: "What is recited in the audio?",
27
    2: "What sport and what nursery rhyme are referenced?",
28
}
29

30
31
32

class ModelRequestData(NamedTuple):
    engine_args: EngineArgs
33
34
35
36
37
    prompt: str | None = None
    prompt_token_ids: dict[str, list[int]] | None = None
    multi_modal_data: dict[str, Any] | None = None
    stop_token_ids: list[int] | None = None
    lora_requests: list[LoRARequest] | None = None
38
39


40
41
42
43
# NOTE: The default `max_num_seqs` and `max_model_len` may result in OOM on
# lower-end GPUs.
# Unless specified, these settings have been tested to work on a single L4.

44

Patrick von Platen's avatar
Patrick von Platen committed
45
46
47
# Voxtral
def run_voxtral(question: str, audio_count: int) -> ModelRequestData:
    from mistral_common.audio import Audio
48
    from mistral_common.protocol.instruct.chunk import (
Patrick von Platen's avatar
Patrick von Platen committed
49
50
51
        AudioChunk,
        RawAudio,
        TextChunk,
52
53
    )
    from mistral_common.protocol.instruct.messages import (
Patrick von Platen's avatar
Patrick von Platen committed
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
        UserMessage,
    )
    from mistral_common.protocol.instruct.request import ChatCompletionRequest
    from mistral_common.tokens.tokenizers.mistral import MistralTokenizer

    model_name = "mistralai/Voxtral-Mini-3B-2507"
    tokenizer = MistralTokenizer.from_hf_hub(model_name)

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=8192,
        max_num_seqs=2,
        limit_mm_per_prompt={"audio": audio_count},
        config_format="mistral",
        load_format="mistral",
        tokenizer_mode="mistral",
        enforce_eager=True,
        enable_chunked_prefill=False,
    )

    text_chunk = TextChunk(text=question)
    audios = [
        Audio.from_file(str(audio_assets[i].get_local_path()), strict=False)
        for i in range(audio_count)
    ]
    audio_chunks = [
        AudioChunk(input_audio=RawAudio.from_audio(audio)) for audio in audios
    ]

    messages = [UserMessage(content=[*audio_chunks, text_chunk])]

    req = ChatCompletionRequest(messages=messages, model=model_name)

    tokens = tokenizer.encode_chat_completion(req)
    prompt_ids, audios = tokens.tokens, tokens.audios

    audios_and_sr = [(au.audio_array, au.sampling_rate) for au in audios]

    multi_modal_data = {"audio": audios_and_sr}

    return ModelRequestData(
        engine_args=engine_args,
        prompt_token_ids=prompt_ids,
        multi_modal_data=multi_modal_data,
    )


Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
# Gemma3N
def run_gemma3n(question: str, audio_count: int) -> ModelRequestData:
    model_name = "google/gemma-3n-E2B-it"
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=2048,
        max_num_batched_tokens=2048,
        max_num_seqs=2,
        limit_mm_per_prompt={"audio": audio_count},
        enforce_eager=True,
    )
    prompt = f"<start_of_turn>user\n<audio_soft_token>{question}"
    "<end_of_turn>\n<start_of_turn>model\n"
    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
    )


120
121
# Granite Speech
def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:
122
    # NOTE - the setting in this example are somewhat different from what is
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
    # optimal for granite speech, and it is generally recommended to use beam
    # search. Check the model README for suggested settings.
    # https://huggingface.co/ibm-granite/granite-speech-3.3-8b
    model_name = "ibm-granite/granite-speech-3.3-8b"

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=2048,
        max_num_seqs=2,
        enable_lora=True,
        max_lora_rank=64,
        limit_mm_per_prompt={"audio": audio_count},
    )

    # The model has an audio-specific lora directly in its model dir;
    # it should be enabled whenever you pass audio inputs to the model.
    speech_lora_path = model_name
    audio_placeholder = "<|audio|>" * audio_count
    prompts = f"<|start_of_role|>system<|end_of_role|>Knowledge Cutoff Date: April 2024.\nToday's Date: December 19, 2024.\nYou are Granite, developed by IBM. You are a helpful AI assistant<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>{audio_placeholder}{question}<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>"  # noqa: E501

    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompts,
        lora_requests=[LoRARequest("speech", 1, speech_lora_path)],
    )


151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
# MiDashengLM
def run_midashenglm(question: str, audio_count: int):
    model_name = "mispeech/midashenglm-7b"

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        max_num_seqs=5,
        limit_mm_per_prompt={"audio": audio_count},
    )

    audio_in_prompt = "".join(
        ["<|audio_bos|><|AUDIO|><|audio_eos|>" for idx in range(audio_count)]
    )

    default_system = "You are a helpful language and speech assistant."

    prompt = (
        f"<|im_start|>system\n{default_system}<|im_end|>\n"
        "<|im_start|>user\n"
        f"{audio_in_prompt}{question}<|im_end|>\n"
        "<|im_start|>assistant\n"
    )
    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
    )


181
# MiniCPM-O
182
def run_minicpmo(question: str, audio_count: int) -> ModelRequestData:
183
    model_name = "openbmb/MiniCPM-o-2_6"
184
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
185
186
187
188
    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
189
        max_num_seqs=2,
190
191
        limit_mm_per_prompt={"audio": audio_count},
    )
192

193
    stop_tokens = ["<|im_end|>", "<|endoftext|>"]
194
195
196
197
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]

    audio_placeholder = "(<audio>./</audio>)" * audio_count
    audio_chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n<|spk_bos|><|spk|><|spk_eos|><|tts_bos|>' }}{% endif %}"  # noqa: E501
198
199
200
201
202
203
204
    messages = [{"role": "user", "content": f"{audio_placeholder}\n{question}"}]
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        chat_template=audio_chat_template,
    )
205
206
207
208
209
210

    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
        stop_token_ids=stop_token_ids,
    )
211
212


213
# Phi-4-multimodal-instruct
214
def run_phi4mm(question: str, audio_count: int) -> ModelRequestData:
215
216
217
218
219
220
221
222
    """
    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
    show how to process audio inputs.
    """
    model_path = snapshot_download("microsoft/Phi-4-multimodal-instruct")
    # Since the vision-lora and speech-lora co-exist with the base model,
    # we have to manually specify the path of the lora weights.
    speech_lora_path = os.path.join(model_path, "speech-lora")
223
    placeholders = "".join([f"<|audio_{i + 1}|>" for i in range(audio_count)])
224

225
    prompts = f"<|user|>{placeholders}{question}<|end|><|assistant|>"
226

227
    engine_args = EngineArgs(
228
229
        model=model_path,
        trust_remote_code=True,
230
        max_model_len=12800,
231
232
233
        max_num_seqs=2,
        enable_lora=True,
        max_lora_rank=320,
234
        limit_mm_per_prompt={"audio": audio_count},
235
236
    )

237
238
239
240
241
    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompts,
        lora_requests=[LoRARequest("speech", 1, speech_lora_path)],
    )
242
243


244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
def run_phi4_multimodal(question: str, audio_count: int) -> ModelRequestData:
    """
    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
    show how to process audio inputs.
    """
    model_path = snapshot_download(
        "microsoft/Phi-4-multimodal-instruct", revision="refs/pr/70"
    )
    # Since the vision-lora and speech-lora co-exist with the base model,
    # we have to manually specify the path of the lora weights.
    speech_lora_path = os.path.join(model_path, "speech-lora")
    placeholders = "<|audio|>" * audio_count

    prompts = f"<|user|>{placeholders}{question}<|end|><|assistant|>"

    engine_args = EngineArgs(
        model=model_path,
        max_model_len=12800,
        max_num_seqs=2,
        enable_lora=True,
        max_lora_rank=320,
        limit_mm_per_prompt={"audio": audio_count},
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompts,
        lora_requests=[LoRARequest("speech", 1, speech_lora_path)],
    )


275
# Qwen2-Audio
276
def run_qwen2_audio(question: str, audio_count: int) -> ModelRequestData:
277
278
    model_name = "Qwen/Qwen2-Audio-7B-Instruct"

279
280
281
282
283
284
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        limit_mm_per_prompt={"audio": audio_count},
    )
285

286
287
288
289
290
291
    audio_in_prompt = "".join(
        [
            f"Audio {idx + 1}: <|audio_bos|><|AUDIO|><|audio_eos|>\n"
            for idx in range(audio_count)
        ]
    )
292

293
294
295
296
297
298
    prompt = (
        "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
        "<|im_start|>user\n"
        f"{audio_in_prompt}{question}<|im_end|>\n"
        "<|im_start|>assistant\n"
    )
299
300
301
302
303

    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
    )
304
305


306
307
308
309
310
311
312
313
314
315
316
# Qwen2.5-Omni
def run_qwen2_5_omni(question: str, audio_count: int):
    model_name = "Qwen/Qwen2.5-Omni-7B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        limit_mm_per_prompt={"audio": audio_count},
    )

317
318
319
    audio_in_prompt = "".join(
        ["<|audio_bos|><|AUDIO|><|audio_eos|>\n" for idx in range(audio_count)]
    )
320
321
322
323

    default_system = (
        "You are Qwen, a virtual human developed by the Qwen Team, Alibaba "
        "Group, capable of perceiving auditory and visual inputs, as well as "
324
325
        "generating text and speech."
    )
326

327
328
329
330
331
332
    prompt = (
        f"<|im_start|>system\n{default_system}<|im_end|>\n"
        "<|im_start|>user\n"
        f"{audio_in_prompt}{question}<|im_end|>\n"
        "<|im_start|>assistant\n"
    )
333
334
335
336
337
338
    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
    )


339
# Ultravox 0.5-1B
340
def run_ultravox(question: str, audio_count: int) -> ModelRequestData:
341
    model_name = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
342

343
    tokenizer = AutoTokenizer.from_pretrained(model_name)
344
345
346
347
    messages = [{"role": "user", "content": "<|audio|>\n" * audio_count + question}]
    prompt = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
348

349
350
351
352
353
354
355
356
357
358
359
360
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        trust_remote_code=True,
        limit_mm_per_prompt={"audio": audio_count},
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
    )
361
362
363


# Whisper
364
def run_whisper(question: str, audio_count: int) -> ModelRequestData:
365
    assert audio_count == 1, "Whisper only support single audio input per prompt"
366
367
368
369
    model_name = "openai/whisper-large-v3-turbo"

    prompt = "<|startoftranscript|>"

370
371
372
373
374
375
376
377
378
379
380
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=448,
        max_num_seqs=5,
        limit_mm_per_prompt={"audio": audio_count},
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompt=prompt,
    )
381
382
383


model_example_map = {
Patrick von Platen's avatar
Patrick von Platen committed
384
    "voxtral": run_voxtral,
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
385
    "gemma3n": run_gemma3n,
386
    "granite_speech": run_granite_speech,
387
    "midashenglm": run_midashenglm,
388
    "minicpmo": run_minicpmo,
389
    "phi4_mm": run_phi4mm,
390
    "phi4_multimodal": run_phi4_multimodal,
391
    "qwen2_audio": run_qwen2_audio,
392
    "qwen2_5_omni": run_qwen2_5_omni,
393
394
    "ultravox": run_ultravox,
    "whisper": run_whisper,
395
}
396
397


398
399
def parse_args():
    parser = FlexibleArgumentParser(
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
        description="Demo on using vLLM for offline inference with "
        "audio language models"
    )
    parser.add_argument(
        "--model-type",
        "-m",
        type=str,
        default="ultravox",
        choices=model_example_map.keys(),
        help='Huggingface "model_type".',
    )
    parser.add_argument(
        "--num-prompts", type=int, default=1, help="Number of prompts to run."
    )
    parser.add_argument(
        "--num-audios",
        type=int,
        default=1,
        choices=[0, 1, 2],
        help="Number of audio items per prompt.",
    )
    parser.add_argument(
        "--seed",
        type=int,
        default=None,
        help="Set the seed when initializing `vllm.LLM`.",
    )
427
428
429
430

    return parser.parse_args()


431
432
433
434
435
def main(args):
    model = args.model_type
    if model not in model_example_map:
        raise ValueError(f"Model type {model} is not supported.")

436
    audio_count = args.num_audios
437
438
439
    req_data = model_example_map[model](
        question_per_audio_count[audio_count], audio_count
    )
440

441
442
443
    # Disable other modalities to save memory
    default_limits = {"image": 0, "video": 0, "audio": 0}
    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
444
445
        req_data.engine_args.limit_mm_per_prompt or {}
    )
446

447
448
449
    engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
    llm = LLM(**engine_args)

450
451
    # We set temperature to 0.2 so that outputs can be different
    # even when all prompts are identical when running batch inference.
452
453
454
    sampling_params = SamplingParams(
        temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
    )
455

Patrick von Platen's avatar
Patrick von Platen committed
456
457
458
459
460
461
462
463
464
    mm_data = req_data.multi_modal_data
    if not mm_data:
        mm_data = {}
        if audio_count > 0:
            mm_data = {
                "audio": [
                    asset.audio_and_sample_rate for asset in audio_assets[:audio_count]
                ]
            }
465
466

    assert args.num_prompts > 0
Patrick von Platen's avatar
Patrick von Platen committed
467
468
469
470
471
472
473
    inputs = {"multi_modal_data": mm_data}

    if req_data.prompt:
        inputs["prompt"] = req_data.prompt
    else:
        inputs["prompt_token_ids"] = req_data.prompt_token_ids

474
    if args.num_prompts > 1:
475
        # Batch inference
476
        inputs = [inputs] * args.num_prompts
477
    # Add LoRA request if applicable
478
479
480
    lora_request = (
        req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
    )
481
482
483
484
485
486

    outputs = llm.generate(
        inputs,
        sampling_params=sampling_params,
        lora_request=lora_request,
    )
487
488
489
490
491
492
493

    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)


if __name__ == "__main__":
494
    args = parse_args()
495
    main(args)