-
Notifications
You must be signed in to change notification settings - Fork 5
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
- Status: Open.#1792 In ohdearquant/lattice;
- Status: Open.#1791 In ohdearquant/lattice;
- Status: Open.#1790 In ohdearquant/lattice;
Speculative softmax and step logprobs still scale logits before centering
bugSomething isn't workingSomething isn't workinglattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1774 In ohdearquant/lattice;inference: gdn_chunked_state_vs_serial_state_diff fails at main (one element at 5.1% vs 5% bound)
bugSomething isn't workingSomething isn't workinglattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1773 In ohdearquant/lattice;inference: CPU prefill normalizes every row and calls no-op LoRA hooks
enhancementNew feature or requestNew feature or requestlattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1767 In ohdearquant/lattice;inference: standalone Q8 CPU path dequantizes tiles to f32 for M=1 GEMM calls
enhancementNew feature or requestNew feature or requestlattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1766 In ohdearquant/lattice;inference: CPU prefill scratch sizes every phase family per request
enhancementNew feature or requestNew feature or requestlattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1765 In ohdearquant/lattice;inference: CPU decode attention reduces V dim-outer (production path and four copies)
enhancementNew feature or requestNew feature or requestlattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1764 In ohdearquant/lattice;inference: CPU batched prefill attention re-reads K/V per query row with a scalar dot
enhancementNew feature or requestNew feature or requestlattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1763 In ohdearquant/lattice;inference: PaddleOCR-VL greedy loops run one unused decode step at the token budget
lattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)Status: Open.#1762 In ohdearquant/lattice;inference: DecoderSession finish contract is not enforced by the driver or the CPU sessions
lattice-inferenceAffects the lattice-inference crate (transformer inference)Affects the lattice-inference crate (transformer inference)tech-debtTech-debt consolidation work from the 2026-07 auditTech-debt consolidation work from the 2026-07 auditStatus: Open.#1761 In ohdearquant/lattice;