w2v-bert-fast open-sources exact 3x speedup for w2v-BERT 2.0 training and inference
kastnerkyle · x · 2026-11-23
Developer Wataru-Nakata released w2v-bert-fast, an accelerated implementation of Hugging Face's Wav2Vec2BertModel (w2v-BERT 2.0) that delivers roughly 3x speedup for both training (forward+backward) and inference (forward-only).
Key points:
- Exact outputs: same weights, same state-dict keys, numerically identical results — drop-in replacement.
- Techniques: fused Triton relative-position attention, channels-last convolutions, GPU-side SeamlessM4T feature extraction, and per-layer torch.compile.
- Usage: install from GitHub and load with w2vbertfast.load(..., dtype=torch.bfloat16, compile=True).
Apache-2.0 licensed; useful for speech pretraining, fine-tuning, and inference pipelines.
More from Multimodal
- AI reads 2,000 historical illustrations: individual errors, near-perfect aggregate patterns — _akpiper · 2026-11-23
- Free Claude plugin turns your MCP-connected video models into an MV factory — AIandDesign · 2026-11-23
- AI short film 'The Wow Lunch': Midjourney images animated by Seedance 2.5 via Claude MCP — azed_ai · 2026-11-22
- Wan 2.2 Image-to-Video on 8GB VRAM: 4-Step Clips With Auto Last-Frame Saving — Equivalent-Ad1959 · 2026-11-22
- Zero Custom Nodes: A ComfyUI Pipeline That Puts a Model Into Any Reference Video's Motion — lmcdesign · 2026-11-22
- Open-source Short-Drama Skill helped a creator ship 20+ AI dramas with 50M+ views — vista8 · 2026-11-22