[model] support FireRedLID (#39290)

Signed-off-by: PatchouliTaisa <patchychen@tencent.com>
Co-authored-by: PatchouliTaisa <patchychen@tencent.com>
This commit is contained in:
PatchyTIS
2026-04-10 08:43:58 +00:00
committed by GitHub
co-authored by PatchouliTaisa
parent 8e8a3becd1
commit 967146e7bd
14 changed files with 1803 additions and 347 deletions
@@ -537,9 +537,30 @@ def run_whisper(question: str, audio_count: int) -> ModelRequestData:
)
# FireRedLID
def run_fireredlid(question: str, audio_count: int) -> ModelRequestData:
assert audio_count == 1, "FireRedLID only supports single audio input per prompt"
model_name = "PatchyTisa/FireRedLID-vllm"
prompt = "<sos>"
engine_args = EngineArgs(
model=model_name,
max_model_len=8,
max_num_seqs=5,
limit_mm_per_prompt={"audio": audio_count},
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompt,
)
model_example_map = {
"audioflamingo3": run_audioflamingo3,
"cohere_asr": run_cohere_asr,
"fireredlid": run_fireredlid,
"funaudiochat": run_funaudiochat,
"gemma3n": run_gemma3n,
"glmasr": run_glmasr,
@@ -55,7 +55,91 @@ def run_whisper():
)
def run_fireredasr2():
"""
FireRedASR2 – Automatic Speech Recognition model.
This model uses a Conformer encoder + Qwen2 LLM decoder architecture
for speech-to-text transcription. Audio is passed via the implicit
prompt format with the ``<|AUDIO|>`` placeholder token.
"""
engine_args = EngineArgs(
model="allendou/FireRedASR2-LLM-vllm",
max_model_len=448,
max_num_seqs=16,
limit_mm_per_prompt={"audio": 1},
)
prompt_str = (
"<|im_start|>user\n<|AUDIO|>请转写音频为文字<|im_end|>\n<|im_start|>assistant\n"
)
prompts = [
{ # Implicit prompt with audio
"prompt": prompt_str,
"multi_modal_data": {
"audio": AudioAsset("mary_had_lamb").audio_and_sample_rate,
},
},
{ # Another audio sample
"prompt": prompt_str,
"multi_modal_data": {
"audio": AudioAsset("winning_call").audio_and_sample_rate,
},
},
]
return ModelRequestData(
engine_args=engine_args,
prompts=prompts,
)
def run_fireredlid():
"""
FireRedLID – Language Identification model.
This encoder-decoder model identifies the spoken language of an audio
clip. It outputs at most 2 tokens representing the detected language
(e.g. "en", "zh mandarin").
"""
engine_args = EngineArgs(
model="PatchyTisa/FireRedLID-vllm",
max_model_len=8,
max_num_seqs=16,
limit_mm_per_prompt={"audio": 1},
)
prompts = [
{ # Test explicit encoder/decoder prompt
"encoder_prompt": {
"prompt": "",
"multi_modal_data": {
"audio": AudioAsset("mary_had_lamb").audio_and_sample_rate,
},
},
"decoder_prompt": "<sos>",
},
{ # Another audio sample
"encoder_prompt": {
"prompt": "",
"multi_modal_data": {
"audio": AudioAsset("winning_call").audio_and_sample_rate,
},
},
"decoder_prompt": "<sos>",
},
]
return ModelRequestData(
engine_args=engine_args,
prompts=prompts,
)
model_example_map = {
"fireredasr2": run_fireredasr2,
"fireredlid": run_fireredlid,
"whisper": run_whisper,
}