Leoy d5ea9eb6f9 fix(glm_asr): warn when vLLM dtype=fp16 (degraded output) (#2993) ...

The GLM-ASR vLLM engine accepts dtype='fp16' but, like Fun-ASR-Nano, fp16
can produce degraded or garbage transcription due to numerical overflow in
the audio embedding path. Fun-ASR-Nano already warns about this; GLM-ASR did
not, so users hitting it got silently poor output with no hint to switch to
bf16/fp32.

Add a small dependency-free helper that emits a one-time warning when fp16 is
requested (the value is still honoured for GPUs that only support fp16), wire
it into GLMASRVLLMEngine.__init__, and cover it with unit tests that run
without a GPU or vLLM.

2026-06-20 22:40:03 +08:00

..

run_test.py

fix: repair KWS symbol cleanup regex

2026-05-21 13:05:51 +00:00

test_asr_inference_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00

test_asr_vad_punc_inference_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00

test_auto_model.py

Improve auto-registration import diagnostics

2026-05-20 12:51:56 +00:00

test_fsmn_vad_dynamic_silence.py

fix: respect explicit VAD silence threshold

2026-06-11 15:16:24 +00:00

test_fun_asr_nano_repetition_penalty.py

fix: avoid CUDA crash from repetition_penalty in Fun-ASR-Nano vLLM prompt-embeds mode (#2948) (#2974)

2026-06-15 09:31:04 -07:00

test_glm_asr_vllm_dtype.py

fix(glm_asr): warn when vLLM dtype=fp16 (degraded output) (#2993)

2026-06-20 22:40:03 +08:00

test_kws_utils.py

fix: repair KWS symbol cleanup regex

2026-05-21 13:05:51 +00:00

test_lm_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00

test_punc_model_none.py

Fix #2839: initialize punc_res before conditional to prevent UnboundLocalError (#2840)

2026-05-19 16:30:00 +08:00

test_punctuation_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00

test_sv_inference_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00

test_tp_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00

test_vad_inference_pipeline.py

Dev gzf exp (#1654)

2024-04-24 16:03:38 +08:00