Issues

• hrx: at -ngl 99 the weights are resident twice in GTT (one whole-model BO plus per-tensor BOs) - 5.9 GiB for a 2.38 GiB model

  bughrx
  1bit-MONSTER/engine#856 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  2

• npu: the OGA 0.5B pack loads and answers in the engine, but the first greedy token is wrong (flat head over a large-magnitude stream)

  bugnpu
  1bit-MONSTER/engine#855 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  3

• tests: hybrid_kvshare_route asserts a sub-millisecond share time, so it fails on a loaded box (not a code failure)

  bug
  1bit-MONSTER/engine#852 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• tests/hybrid_route.sh: the failing-provider arm flakes on CI (fixed ~10 s startup budget)

  bug
  1bit-MONSTER/engine#850 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  2

• npu-pack: the q4_1 zero-point nibbles are paired by halves, the codes by even/odd (main, since #829)

  bugnpu
  1bit-MONSTER/engine#849 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu: the forward is causal-only, so a noncausal_full_attention pack (d3-lite) cannot be reproduced — measured, not inferred

  bugnpu
  1bit-MONSTER/engine#847 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu: the forward derives GEMM K from hidden, but a Q4NX container pads K to the 256-column grid (hidden 896 -> 1024)

  bugenhancementnpu
  1bit-MONSTER/engine#845 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  3

• hrx: GGML_HRX_GRAPH_PROGRAM_CACHE_MB sees ~16 MiB per cached program while the cache grows 0.4-2.3 GiB per shape — the 64-program count is the only bound that bites

  bughrx
  1bit-MONSTER/engine#841 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• hrx: the AMDGPU HAL decides APU-ness from a bit this ROCr leaves clear (0x2), so host+device-local memory types are refused on Strix Halo

  bughrx
  1bit-MONSTER/engine#840 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu-window11: the 2 % ms/launch tolerance is smaller than this box's noise floor — the same binary against itself moves up to 7.7 %

  bug
  1bit-MONSTER/engine#834 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu: forward-serve accepts chat_template_kwargs.enable_thinking and silently ignores it, so logprobs.content is always empty for a thinking model

  bugnpu
  1bit-MONSTER/engine#832 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  2

• 1bit OS on Bosgame BeyondMax (AXB35-02, BIOS 1.09): only the front USB-A port works under Linux; Type-C and rear ports are electrically silent (Windows: all work)

  bugos
  1bit-MONSTER/engine#823 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• libamdf: XDNA import class advertises FOREIGN_API dma-buf execution, but execution on an imported range fails EFAULT on amdxdna 526c608 (PASID)

  bughrx
  1bit-MONSTER/engine#812 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  3

• npu plan P7: int8 KV with per-head Hadamard rotation, and NPU prefill in 1024/2048-token chunks with block-fp expansion

  enhancement
  1bit-MONSTER/engine#806 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  2

• npu plan P6: the shared drafter protocol - tau-gated chain, 4-row verify with leaf fill, deferred GDN commit; acceptance per target in the leaderboard

  enhancement
  1bit-MONSTER/engine#805 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu plan P5: Hexagon target - t2 gemv on HVX/HMX, QEMU v73 bit-exact, AI Hub X Elite timing, one exact layer, cost model

  enhancement
  1bit-MONSTER/engine#804 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  11

• npu plan P4: extract NpuTarget + NpuModel; xdna target wraps I8Ctx; host fallbacks for optional ops (lands after c20 ships)

  enhancement
  1bit-MONSTER/engine#803 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  4

• npu plan P3: t2 on-core expansion kernel for the sealed M=16 family (XDNA) - halve the weight stream

  enhancement
  1bit-MONSTER/engine#802 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu plan P2: the t2 chunk kind (2-bit ternary + bf16 scale per 128 K, 2.125 b/w) - spec, CPU dequant, converters, bit-exact oracle (additive, may land before c20)

  enhancement
  1bit-MONSTER/engine#801 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  9

• npu: the plan layer - one model description (plan.json), t2 ternary chunks, NpuTarget per NPU (XDNA, Hexagon)

  enhancement
  1bit-MONSTER/engine#799 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  1

• npu: a real d3 pack (vllm-sr/d3-lite, 0.8B-GDN) answering through /v1/systemone — converter transforms, name map, and the 7-design ELF set

  enhancement
  1bit-MONSTER/engine#796 · Filter by author bong-water-water-bong opened on Nov 2, 2026
  6

• hrx: Q4_K prefill at M>2048 is 3.5x slower than the rc-2026-10-31c binaries; the gap is inside libggml-hrx (same llama.cpp pin, same dispatch)

  hrxperf
  1bit-MONSTER/engine#782 · Filter by author bong-water-water-bong opened on Nov 1, 2026
  4

• Split prefill in the hybrid route: blocked on an NPU prefill GEMM ~4x faster (measured bar inside)

  enhancementperf
  1bit-MONSTER/engine#754 · Filter by author bong-water-water-bong opened on Nov 1, 2026
  6

• NPU: execute on an imported dma-buf (XRT cannot bind one today; iGPU->NPU sharing already works)

  enhancement
  1bit-MONSTER/engine#753 · Filter by author bong-water-water-bong opened on Nov 1, 2026
  8

• npu-pack: layer norms (and the final norm) are located only for hidden 896

  enhancement
  1bit-MONSTER/engine#734 · Filter by author bong-water-water-bong opened on Nov 1, 2026
  4