Commit History
17 days ago
-
Fix streaming decode with use_cuda_graph=True: first frame never runs, K-frame calls repeat the last frame
Verified signatureHi! `batch_decode(streaming=True, use_cuda_graph=True)` gives wrong audio for me. I traced it to `_graphed_decode_frame`: 1. On the first call the graph is captured but not replayed. Capture doesn't run anything, so frame 0 is never decoded and the decoder state skips it. 2. It returns the graph's static output buffers. `step()` keeps a view per frame and only concatenates after the loop, so when a call decodes more than one frame, all of them end up as the last one. This PR replays right after capture and returns copies. The non-graph path is unchanged. Tested on an RTX 4070 Ti SUPER (torch 2.8, fp32), 64 frames of random codes, 4 frames per call, compared with `use_cuda_graph=False`: | | frames out | frames matching plain decode | max abs diff | ms / frame | |---|---|---|---|---| | plain decode | 64 | – | – | 19.2 | | graph, before | 63 | 0% | – | 3.8 | | graph, after | 64 | 100% | 0.0 | 3.7 | The fixed version also matches exactly with 1 and 3 frames per call, and the same patc…2d88660
-
Update MOSS-Audio-Tokenizer-Nano model card
KuangWei Chen6aa02b0
-
optimize flash attention import
Kuangwei Chen8ee35eb -
Update readme
Kuangwei Chen737e5b1 -
Update for realtime
Kuangwei Chen1c2bf4d
-
update nano
--replace-alld4a3b2c -
initial commit
Verified signature61f845e
