- chore: refresh standard CPU repacks (30f4695c09f5ae234754606205a92bdb24d04278)
- feat: publish explicit CPU weight variants (00a9e25a9056f54a97b6e4957361071a16313b38)
9 days ago
Publish verification fixtures: the FP32 GPU smoke record and the original GPTQ overlay
Lets the weights-backed gates run off the machine that recorded them (amazingvince/falcon-ocr-cpu issue #8). Inputs to the checks, not inference weights.
Add the trained draft head (speculative decoding) and update the model card
EAGLE-3-style draft head for the Rust CPU runner (github.com/amazingvince/falcon-ocr-cpu, main): used automatically when in the model folder; never changes the output; 21-26% faster pages. Model card: current accuracy and speed, draft-head training data and terms, links to main.
Kernel-ready near-exact and fast files for the Falcon-OCR CPU runner
