Qwen3-ASR-1.7B

from modal_training_gym.common.models.qwen3_asr_1_7b import Qwen3_ASR_1_7B

Alibaba Qwen3-ASR-1.7B speech recognition model.

Attributes

model_name str

Hugging Face repository ID. Default: "Qwen/Qwen3-ASR-1.7B"

architecture ModelArchitecture | None

Megatron architecture parameters for the text backbone. Default: ModelArchitecture(...)

response_parser Callable[[str], ParsedResponse] | None

Parser for generated text. Default: parse_qwen3_response

requires_bshd bool

Requires padded BSHD batches during training. Default: True

audio_placeholder str

Token sequence that marks audio input. Default: "<|audio_start|><|audio_pad|><|audio_end|>"

download

download() -> None

Download the model and add the tokenizer.json required by SGLang.

parse_response

parse_response(text: str) -> ParsedResponse

Parse model text with response_parser.

Without a configured parser, the model text becomes ParsedResponse.content.

Returns

Parsed model output.