VibeVoice/vllm_plugin · vib/VibeVoice

PpengzhiliangFix streaming release: default model id, doc links, ffmpeg step

文件	最后提交记录	最后更新时间
scripts	Fix streaming release: default model id, doc links, ffmpeg step The vLLM streaming server defaulted to microsoft/VibeVoice-ASR-Streaming, which does not exist on the Hub — following the docs verbatim (the docker command passes no --model) failed with a 401 RepositoryNotFoundError. Point it at the released microsoft/VibeVoice-ASR-Streaming-7B. Also fix two links to the same non-existent repo, and move the ffmpeg install out of Usage 1 into Installation, since the file-inference usage needs it too and otherwise dies on FileNotFoundError: 'ffmpeg'. Refresh the technical report PDF.	12 天前
tests	Serve streaming ASR through the vLLM plugin A streaming prompt interleaves one audio window per chunk, so a single prompt carries N audio items rather than one. The plugin now handles that shape: - Streaming checkpoints declare VibeVoiceForASRStreamingTraining; the decoder is identical, so the architecture name alone selects the interleaved prompt layout. Deriving it from the name rather than a config flag keeps a checkpoint from being served under the wrong protocol. - The multimodal budget is sized from the window the checkpoint was trained on, read off the checkpoint. Profiling the 61-minute non-streaming bound per item would VAE-encode ~88M samples N times over and OOM at startup. - asr_streaming.py carries the chunk geometry; asr_streaming_server.py the incremental session on top of it. Adds a launcher, a browser demo API, an end-to-end API test, and docs/vibevoice-vllm-asr-streaming.md.	18 天前
tools	Serve streaming ASR through the vLLM plugin A streaming prompt interleaves one audio window per chunk, so a single prompt carries N audio items rather than one. The plugin now handles that shape: - Streaming checkpoints declare VibeVoiceForASRStreamingTraining; the decoder is identical, so the architecture name alone selects the interleaved prompt layout. Deriving it from the name rather than a config flag keeps a checkpoint from being served under the wrong protocol. - The multimodal budget is sized from the window the checkpoint was trained on, read off the checkpoint. Profiling the 61-minute non-streaming bound per item would VAE-encode ~88M samples N times over and OOM at startup. - asr_streaming.py carries the chunk geometry; asr_streaming_server.py the incremental session on top of it. Adds a launcher, a browser demo API, an end-to-end API test, and docs/vibevoice-vllm-asr-streaming.md.	18 天前
__init__.py	Serve streaming ASR through the vLLM plugin A streaming prompt interleaves one audio window per chunk, so a single prompt carries N audio items rather than one. The plugin now handles that shape: - Streaming checkpoints declare VibeVoiceForASRStreamingTraining; the decoder is identical, so the architecture name alone selects the interleaved prompt layout. Deriving it from the name rather than a config flag keeps a checkpoint from being served under the wrong protocol. - The multimodal budget is sized from the window the checkpoint was trained on, read off the checkpoint. Profiling the 61-minute non-streaming bound per item would VAE-encode ~88M samples N times over and OOM at startup. - asr_streaming.py carries the chunk geometry; asr_streaming_server.py the incremental session on top of it. Adds a launcher, a browser demo API, an end-to-end API test, and docs/vibevoice-vllm-asr-streaming.md.	18 天前
asr_streaming.py	Serve streaming ASR through the vLLM plugin A streaming prompt interleaves one audio window per chunk, so a single prompt carries N audio items rather than one. The plugin now handles that shape: - Streaming checkpoints declare VibeVoiceForASRStreamingTraining; the decoder is identical, so the architecture name alone selects the interleaved prompt layout. Deriving it from the name rather than a config flag keeps a checkpoint from being served under the wrong protocol. - The multimodal budget is sized from the window the checkpoint was trained on, read off the checkpoint. Profiling the 61-minute non-streaming bound per item would VAE-encode ~88M samples N times over and OOM at startup. - asr_streaming.py carries the chunk geometry; asr_streaming_server.py the incremental session on top of it. Adds a launcher, a browser demo API, an end-to-end API test, and docs/vibevoice-vllm-asr-streaming.md.	18 天前
asr_streaming_server.py	Serve streaming ASR through the vLLM plugin A streaming prompt interleaves one audio window per chunk, so a single prompt carries N audio items rather than one. The plugin now handles that shape: - Streaming checkpoints declare VibeVoiceForASRStreamingTraining; the decoder is identical, so the architecture name alone selects the interleaved prompt layout. Deriving it from the name rather than a config flag keeps a checkpoint from being served under the wrong protocol. - The multimodal budget is sized from the window the checkpoint was trained on, read off the checkpoint. Profiling the 61-minute non-streaming bound per item would VAE-encode ~88M samples N times over and OOM at startup. - asr_streaming.py carries the chunk geometry; asr_streaming_server.py the incremental session on top of it. Adds a launcher, a browser demo API, an end-to-end API test, and docs/vibevoice-vllm-asr-streaming.md.	18 天前
inputs.py	fix: add audio duration validation to prevent OOM in vLLM plugin Add a configurable duration check to vibevoice_audio_input_mapper that rejects audio exceeding VIBEVOICE_MAX_AUDIO_DURATION (default 3660s / 61 minutes) before tensor conversion. Returns a clear error message instead of an opaque CUDA OOM crash. Refs: https://github.com/microsoft/VibeVoice/issues/210	5 个月前
model.py	Serve streaming ASR through the vLLM plugin A streaming prompt interleaves one audio window per chunk, so a single prompt carries N audio items rather than one. The plugin now handles that shape: - Streaming checkpoints declare VibeVoiceForASRStreamingTraining; the decoder is identical, so the architecture name alone selects the interleaved prompt layout. Deriving it from the name rather than a config flag keeps a checkpoint from being served under the wrong protocol. - The multimodal budget is sized from the window the checkpoint was trained on, read off the checkpoint. Profiling the 61-minute non-streaming bound per item would VAE-encode ~88M samples N times over and OOM at startup. - asr_streaming.py carries the chunk geometry; asr_streaming_server.py the incremental session on top of it. Adds a launcher, a browser demo API, an end-to-end API test, and docs/vibevoice-vllm-asr-streaming.md.	18 天前