Issues

• spyre: oc rsync carries the Mac checkout mtimes, so a warm pod build can link the previous binary

  bugtarget/spyre
  AI-native-Systems-Research/scratchy#1027 · Filter by author starpit opened on Nov 1, 2026
  1

• perf(spyre): the decode batch carried maximum spills to HBM on a planner corner mismatch

  enhancementtarget/spyre
  AI-native-Systems-Research/scratchy#1026 · Filter by author starpit opened on Nov 1, 2026
  1

• perf(spyre): split the prompt trip page gather 8 ways, staged on the host

  enhancementtarget/spyre
  AI-native-Systems-Research/scratchy#1025 · Filter by author starpit opened on Nov 1, 2026
  2

• perf(spyre): short-context TTFT is 1.30x main on granite-3.1-2b fp8 (~128 tokens)

  enhancementtarget/spyre
  AI-native-Systems-Research/scratchy#1024 · Filter by author starpit opened on Nov 1, 2026
  1

• spyre: a ladder rung that fails to lower is dropped silently — make it a build failure

  bugtarget/spyre
  AI-native-Systems-Research/scratchy#1023 · Filter by author starpit opened on Nov 1, 2026
  2

• perf(spyre): decode ITL at depth is 1.9x main on 2b, 1.3x on 8b — route one request to the batched kernel

  enhancementtarget/spyre
  AI-native-Systems-Research/scratchy#1022 · Filter by author starpit opened on Nov 1, 2026
  4

• spyre: a one-chunk decode batch decodes wrong on the card in every lowered form

  bugtarget/spyre
  AI-native-Systems-Research/scratchy#1021 · Filter by author starpit opened on Nov 1, 2026
  2

• spyre(tp): granite-2b's long-prompt reply formats differently at tp=2 (coherent)

  target/spyre
  AI-native-Systems-Research/scratchy#1020 · Filter by author starpit opened on Nov 1, 2026

• spyre TP: how to build it — the sharding scheme, the on-card P2P protocol, and the proof standard

  ready-for-agenttarget/spyre
  AI-native-Systems-Research/scratchy#1007 · Filter by author starpit opened on Oct 30, 2026
  2

• perf(metal): the 8-row decode bucket pays the SmallM twin where it loses — gpt-oss batch decode 22% above the union roof

  target/metal
  AI-native-Systems-Research/scratchy#1006 · Filter by author starpit opened on Oct 30, 2026
  2

• bench: the Metal matrix's fixed 64-token answers lock its 4-user cells into lockstep groups, which splits scratchy's TTFT in two

  needs-triagetarget/metal
  AI-native-Systems-Research/scratchy#994 · Filter by author starpit opened on Oct 29, 2026

• perf(metal): prefill speed sets TPOT in the base M5's long-prompt multi-user cells (67-83% of server time is prefill)

  needs-triagetarget/metal
  AI-native-Systems-Research/scratchy#993 · Filter by author starpit opened on Oct 29, 2026

• perf(metal): batched decode pays ~1 ms per 1K tokens of context per row on gpt-oss-20b (base M5), about 30% of a 4-row step at 4K

  needs-triagetarget/metal
  AI-native-Systems-Research/scratchy#992 · Filter by author starpit opened on Oct 29, 2026

• perf(metal): gpt-oss-20b single-stream decode trails oMLX by 3-6% on the base M5, and the gap grows with context

  needs-triagetarget/metal
  AI-native-Systems-Research/scratchy#991 · Filter by author starpit opened on Oct 29, 2026
  1

• feat(models): compile the generation config into the model crate

  AI-native-Systems-Research/scratchy#974 · Filter by author starpit opened on Oct 28, 2026

• feat(sampling): a general, device-based degeneracy defense for decode repetition loops

  AI-native-Systems-Research/scratchy#973 · Filter by author starpit opened on Oct 28, 2026
  1

• spyre: fuse-nothing fused twins — emission bakes byte-identical rung bodies under two fps (4× program-region copies)

  target/spyre
  AI-native-Systems-Research/scratchy#957 · Filter by author starpit opened on Oct 28, 2026

• spyre-hw worker build carries dead-code warnings no CI gate compiles

  target/spyre
  AI-native-Systems-Research/scratchy#956 · Filter by author starpit opened on Oct 28, 2026

• Batch width 2: the emulator's two-row attention output diverges from the golden trace (pre-existing, reproduces on main)

  target/spyre
  AI-native-Systems-Research/scratchy#955 · Filter by author starpit opened on Oct 27, 2026

• fix(serving): honor thinking: disabled on /v1/messages by turning the model's thinking off

  AI-native-Systems-Research/scratchy#952 · Filter by author diegocastanibm opened on Oct 27, 2026

• feat(cuda): GLM-4.5-Air's MLX affine 4-bit routed experts, dequant-fused grouped expert GEMM and gathered GEMV on ThunderKittens over #924's 128-expert sort, bit-identical to the bf16 twins (4-bit step 4 of 4)

  target/cuda
  AI-native-Systems-Research/scratchy#950 · Filter by author diegocastanibm opened on Oct 27, 2026
  1

• feat(cuda): the first quantized kernels, MLX affine 4-bit dequant-fused GEMM (256- and 64-column), GEMV and embed gather on ThunderKittens, bit-identical to the host dequant and to the bf16 twins (4-bit step 3 of 4)

  target/cuda
  AI-native-Systems-Research/scratchy#949 · Filter by author diegocastanibm opened on Oct 27, 2026
  1

• feat(cuda): lower GLM-4.5-Air's MLX affine 4-bit tapes (affine GEMM, GEMV, embed and expert steps), load the packed weights group-major, refused NotLaunched (4-bit step 2 of 4, toolchain-free)

  target/cuda
  AI-native-Systems-Research/scratchy#948 · Filter by author diegocastanibm opened on Oct 27, 2026
  1

• feat(cuda): declare GLM-4.5-Air's MLX 4-bit g64 preset and hand its affine storage facts (dense router, quantized embed and lm_head) across the target-emitter seam (4-bit step 1 of 4, shared, toolchain-free)

  target/cuda
  AI-native-Systems-Research/scratchy#947 · Filter by author diegocastanibm opened on Oct 27, 2026
  1

• feat(cuda): lower and expand glm-4.5-air (glm4-moe) to static CUDA tapes: q/k/v bias add, sigmoid routing with the correction bias, per-expert bundle packing, head_dim 128 (14a, toolchain-free; bf16 does not fit one H100)

  target/cuda
  AI-native-Systems-Research/scratchy#938 · Filter by author diegocastanibm opened on Oct 27, 2026
  1