Repository Issue Activity (beta)

#213 transformers backend: audio encoder attention window (cu_seqlens) is not applied on sdpa/eager; clips > 8 s are encoded with full attention

#197 Qwen3-ForcedAligner-0.6B (qwen-asr) emits zero-duration word spans [t, t] on source audio at high rates

#203 Add Bengali (bn) language support to Qwen-Audio-3.0-ASR

#201 attn_implementation is accepted but silently ignored — never propagated to nested configs (eager request runs sdpa)

#188 VLLM加速推理与直接加载推理效果不一致

#190 Can not download the weights from https://www.modelscope.cn during the execution

#211 [Reproducibility] Unable to reproduce the reported FLEURS-zh CER 2.41 for Qwen3-ASR-1.7B

#186 提示词，热词，效果不理想

#43 TypeError in `check_model_inputs`

#163 中文混合被识别出全英

#165 一些噪音或回声会识别为'嗯.'.'哎.','嗨.'等内容,影响使用

#193 Can qwen-asr be upgraded to vllm 0.21.0 version

#87 Fine-tuning for New Languages (Persian, Urdu, Hebrew) – Best Practices & Aligner Fine-Tuning

#168 Qwen3-ASR-1.7B模型支持闽南语识别吗？需要修改哪些地方呢

#210 Qwen3-ASR-0.6B/1.7B: transformers cannot save the shipped generation_config.json

#129 Endless repetition in output

#199 流式耗时问题

#207 [Transformers] Heterogeneous-length audio batch corrupts the shorter sample

#174 我们有一套Qwen3-ASR的Lora和RL代码可以贡献给社区。

#161 GSPO

#176 Fine-tuning callback silently skips preprocessor / chat-template copy when --model_path is an HF Hub ID (root cause + fix)

#148 Serious Issues with Real-Time Streaming

#94 Qwen3-ASR Streaming result does not contain punctuations for Arabic

#138 Support transformers >= 5.3 , vllm >= 0.17 and some PR.

#110 Why can't vllm be found after installation?