LauraGPT commited on
docs: link tested native CUDA transcription recipe
Browse filesSigned-off-by: LauraGPT <18321252+LauraGPT@users.noreply.github.com>
README.md
CHANGED
|
@@ -103,6 +103,24 @@ Hotwords are hints, not an enforced vocabulary. In the paired Chinese sample che
|
|
| 103 |
|
| 104 |
`language` accepts Chinese, English, and Japanese as ISO codes, full English names, or the checkpoint Chinese names (`中文`, `英文`, `日文`). Keep batch inputs and decoded outputs in order; use `processor_kwargs={"text_kwargs": {"padding": True}}` together with the other processor options for different-length batches. Reject empty batches before calling the processor. An upstream `IndexError` for `audio=[]` was observed in the historical `fc501` source environment; that observation is not a new 5.17.0 test result.
|
| 105 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
## Capability boundary
|
| 107 |
|
| 108 |
This Transformers checkpoint is for generation-based transcription. It does not include the native CTC branch, so this export does not provide CTC-dependent timestamps or speaker diarization. Native model support does not add an HTTP server, realtime audio transport, or known-speaker identification.
|
|
|
|
| 103 |
|
| 104 |
`language` accepts Chinese, English, and Japanese as ISO codes, full English names, or the checkpoint Chinese names (`中文`, `英文`, `日文`). Keep batch inputs and decoded outputs in order; use `processor_kwargs={"text_kwargs": {"padding": True}}` together with the other processor options for different-length batches. Reject empty batches before calling the processor. An upstream `IndexError` for `audio=[]` was observed in the historical `fc501` source environment; that observation is not a new 5.17.0 test result.
|
| 105 |
|
| 106 |
+
## CUDA with the canonical CLI
|
| 107 |
+
|
| 108 |
+
The [isolated CUDA recipe](https://github.com/QwenAudio/Fun-ASR/tree/main/examples/transformers#cuda-an-isolated-tested-recipe)
|
| 109 |
+
provides separate GPU requirements and explicit `--device cuda --dtype bfloat16`
|
| 110 |
+
selection. Keep the CPU environment above separate. CUDA requests fail when
|
| 111 |
+
unavailable rather than silently falling back; BF16 requires device support.
|
| 112 |
+
|
| 113 |
+
On 2026-09-10, Linux/Python 3.12, H100 80 GB, driver 550.127.08,
|
| 114 |
+
Transformers 5.17.0 and matching torch/torchaudio 2.11.0+cu128 passed eight
|
| 115 |
+
float32/BF16 English, Chinese, keyword and padded Chinese/English batch cases
|
| 116 |
+
using the pinned public snapshot. Exact CLI GPU single-file/batch and default
|
| 117 |
+
CPU regression also passed. The Chinese keyword error remains; these are not
|
| 118 |
+
accuracy, throughput, minimum-VRAM or concurrency benchmarks. Other GPUs,
|
| 119 |
+
float16 and hosted GPU Space/Colab execution were not validated.
|
| 120 |
+
|
| 121 |
+
An attention-dispatch warning was observed. Functional success does not verify
|
| 122 |
+
every component's attention kernel or imply Flash Attention performance.
|
| 123 |
+
|
| 124 |
## Capability boundary
|
| 125 |
|
| 126 |
This Transformers checkpoint is for generation-based transcription. It does not include the native CTC branch, so this export does not provide CTC-dependent timestamps or speaker diarization. Native model support does not add an HTTP server, realtime audio transport, or known-speaker identification.
|