Spaces:
Paused
Paused
Commit History
-
add 2
7ab065c -
sync local state with remote
Verified signature1c7b2ff -
remove usage instructions and example text from UI
Verified signature2af14be -
fix: use torchaudio.info with suppressed deprecation warnings instead of torchcodec
Verified signature9e44af9 -
fix: convert non-WAV audio to WAV before encoding, use torchcodec for duration
Verified signature9508f19 -
fix: use torchaudio.info for audio duration (supports m4a without deprecation warnings)
Verified signature3a6a578 -
fix: use librosa.get_duration for audio duration check (supports m4a and more formats)
Verified signatured99970c -
fix: only unpatch ZeroGPU in nanovllm child processes, skip @spaces.GPU workers
Verified signature0443284 -
fix: forcibly clear ZeroGPU dispatch/function mode stacks in child processes
Verified signature3e0652f -
fix: replace ZeroGPU dispatch stubs with pass-through in child processes instead of trying to pop modes
Verified signature41ea91f -
fix: properly unpatch ZeroGPU dispatch modes and force torch.load map_location=cpu in child processes
Verified signatured0f587a -
fix: exclude assets/ and examples/ from LFS tracking
Verified signature99506e9 -
fix: use functools.partial for pickleable multiprocessing wrapper
71f1902 -
fix: make spaces torch mode __exit__ no-op to avoid PyTorch 2.8.0 IndexError
dd79864 -
fix: unpatch ZeroGPU torch stubs in every multiprocessing child process
033d78c -
fix: import os before using os.environ in top-level patches
06f5b3b -
fix: patch spaces ZeroGPU workers to non-daemon to allow child processes
2ca879f -
fix: use direct env check to disable prewarm on ZeroGPU
9f85f57 -
fix: wrap TTS inference in @spaces.GPU for ZeroGPU compatibility
f170d0c -
fix: suppress harmless asyncio BaseEventLoop __del__ errors during GC
619c24e -
fix: close asyncio event loop in worker thread to suppress __del__ errors
2da9298 -
fix: upgrade torch to 2.8.0 for ZeroGPU compatibility
39a0ffd -
fix: import spaces before other packages to avoid CUDA init error
6757fca
-
fix: harden async nanovllm demo stabilityβ
ει«Move the Gradio demo to an async nano-vLLM bridge, limit denoise concurrency, and improve user-facing error handling so bad requests do not disrupt other active sessions.44d9bcd -
fix: revert concurrency limit to 1 β nanovllm event loop is not thread-safeβ
ει«nanovllm-voxcpm uses a single asyncio event loop internally (loop.run_until_complete). With concurrency > 1 multiple Gradio worker threads hit the same loop simultaneously, causing "RuntimeError: This event loop is already running" on every request. Made-with: Cursor364d86e
-
fix: remove unsupported inference_timesteps parameter and dit_steps sliderβ
ει«nanovllm-voxcpm does not support inference_timesteps; remove the parameter to prevent accidental server rebuilds when users adjust the slider. Made-with: Cursor0203b68
-
fix: add cookbook guide heading linkβ
ει«Surface the VoxCPM2 cookbook as a linked heading after the usage overview so users can discover the best-practices guide without cluttering the top of the demo. Made-with: Cursor5351045 -
fix: store daily request logs on persistent diskβ
ει«Write request logs to daily JSONL files under persistent storage while leaving model and dependency caches on ephemeral paths so a small /data volume can be used safely. Made-with: Cursor39d1990 -
fix: show localized upload alert for long audioβ
ει«Validate reference audio duration on upload and surface a localized Gradio error immediately so users see the limit before ASR, denoising, or generation starts. Made-with: Cursor1cf9f24 -
fix: zipenhancer for denoise
ει«96f588d -
fix: add denoiser and reference audio guardsβ
ει«Enable ZipEnhancer preprocessing for reference audio, move SenseVoice downloads to Hugging Face caching, and reject overly long reference clips before ASR or NanoVLLM encoding. Made-with: Cursorffbf382
-
fix: harden Gradio NanoVLLM startup pathβ
ει«Move flash-attn and nanovllm-voxcpm installation into the build requirements and add cache, prewarm, and retry safeguards so the Space starts more predictably and recovers better from transient backend failures. Made-with: Cursor7d8de04 -
fix: install flash-attn from prebuilt wheelβ
ει«Use a resolved flash-attn release wheel instead of building from source at runtime so the Gradio Space avoids CUDA toolchain mismatches during startup. Made-with: Cursord138332 -
fix: pin Gradio NanoVLLM runtime dependenciesβ
ει«Avoid the torch 2.11 CUDA 13 stack that breaks flash-attn builds on Spaces by pinning a more compatible torch/torchaudio/triton set and installing a pinned flash-attn wheel without dependency upgrades. Made-with: Cursored56d7a -
fix: switch Gradio demo backend to nanovllm_voxcpmβ
ει«Keep the existing VoxCPM demo UI while routing inference through Nano-vLLM on persistent GPU Spaces. Add runtime/system dependency setup so the managed Gradio environment can attempt the non-Docker backend path. Made-with: Cursor09dc185 -
add torch version && torch compile == true
ει«29b439e -
add torch version
ει«9dc8c20
-
fix: dynamically install torchaudio matching ZeroGPU torch+CUDA versionβ
ει«Made-with: Cursord3b2b16 -
fix: install CPU torchaudio to avoid CUDA 13 mismatch on ZeroGPUβ
ει«Made-with: Cursorf15ea05 -
fix: install voxcpm with --no-deps at runtime to preserve ZeroGPU torch/torchaudioβ
ει«Made-with: Cursor59f3c75 -
fix: remove torch/torchaudio pin, rely on ZeroGPU pre-installed versionβ
ει«Made-with: Cursor60d19ef -
fix: pin torch/torchaudio to 2.5.0 and move import voxcpm to module levelβ
ει«Made-with: Cursor1253e36 -
fix: move SDPA backend config into GPU function contextβ
ει«On ZeroGPU, CUDA is only available inside @spaces.GPU functions. Module-level torch.backends.cuda calls may interfere with CUDA init. Made-with: Cursorb84906b -
fix: disable flash/mem-efficient SDPA backends for ZeroGPU compatβ
ει«The step-by-step attention in minicpm4 triggers an IndexError on ZeroGPU with flash/mem-efficient SDPA. Fall back to math backend. Made-with: Cursor8f661ec -
fix: disable optimize to avoid ZeroGPU torch.compile incompatibilityβ
ει«Made-with: Cursord45dc51 -
upgrade to VoxCPM2 + Gradio 6 with full i18n UIβ
ει«- Switch default model from VoxCPM1.5 to VoxCPM2 - Rewrite UI to match VoxCPM/app.py: i18n (en/zh-CN), three generation modes (Voice Design, Controllable Cloning, Ultimate Cloning), Control Instruction support, and rich voice examples - Upgrade to Gradio 6 (theme/css/i18n in launch()) - Keep HF Spaces @spaces.GPU and model pre-download at startup - Rename logo asset to voxcpm_logo.png Made-with: Cursor7d9f729
-
Update README.md
Verified signatureced8c00
-
set zero gpu inference
ει«f6f653d -
set zero gpu inference
ει«1700cda -
set zero gpu inference
ει«1488551