cortiq 0.8.54

Added

• Unlimited-OCR-2 (unlimited-ocr, uocr.rs): multi-page document parsing. The
  DeepEncoder (SAM ViT-B feeding CLIP-L), the MoE decoder and its DASH attention:
  a per-layer router scores 32-token blocks through a learned summary, restricts
  them to the query's page, takes entmax-1.5 over the best 16 and reads the best
  12 with a centered log-prior bias plus the last 128 tokens, at most 512 rows
  per query. Two-axis page2d rotary, the page-count guard, Pillow's fixed-point
  bicubic. On wgpu the router and attend ride the token graph and a 64-row batch
  graph chunk (gpu_wgpu/uocr.rs), SAM and CLIP stay resident (vit::sam_encode).
  An RTX PRO 4500 decodes at 320 tok/s against 20 in Transformers float32 and holds
  it over 24 pages; a tiled page's encoder takes 0.84 s. q8_2f matches float32
  on the first 300 greedy tokens of a page.
• MiniCPM-OCR-1.0 (minicpm_ocr, minicpm_ocr.rs): the SigLIP tower with nearest
  position buckets, the 2×2 merger, the slicing image processor and its prompt
  layout, on the resident ViT kernels. 254–267 tok/s against 35 in Transformers
  float32; q8_2f matches float32 on 300 greedy tokens of a page.
• cortiq run --image and /v1/chat/completions image parts for both models.
• Native Metal for both models. gpu_metal/vit.rs keeps a vision tower resident
  (SigLIP, CLIP-L, SAM's windows, global blocks and neck) on the encoders' GEMMs;
  gpu_metal/uocr.rs is the DASH router and attend on the token graph and the
  chunk prefill. On a 10-core M4, Unlimited-OCR decodes at 95–112 tok/s against 71
  on its CPU, a tiled page's encoder takes 3.5 s against 20 s, its prompt 1.4 s
  against 12.6 s; MiniCPM-OCR decodes at 73–102 tok/s with 0.5 s a view against 1.2 s.

Changed

• A MoE shared expert wider than its routed bank rides the batch graph, and in
  q8_2f the token graph too; on Metal the token graph and the chunk prefill too.
• Metal's rows graph takes a q8_2f prompt chunk through the f32-tile GEMM instead
  of a matvec per position (MiniCPM-OCR q8_2f, 2511 rows: 19.4 s to 2.5 s).

Fixed

• vision_gemm_nt on Metal refuses a k off the kernel's grid of 32. SigLIP's
  588-wide patch embedding read past each row, last row into unwritten memory.
• The decision server installs its signal listeners before it binds the port: a
  SIGTERM right after the first response could kill it without the ledger flush.