Issues

tracking: zero-copy on the serving paths (hybrid KV handoff, NPU activations in BOs, ONNX EP IO binding, audit)

1bit-MONSTER/engine#698 · opened

on Oct 31, 2026

npu: the pb16i/sealed-ELF path emits garbage — 48 launches/token missing, 16 int8 M=128 QKV launches on full-attention layers (blocks #692)

1bit-MONSTER/engine#697 · opened

on Oct 31, 2026

hrx: hrx_synchronous_h2d/d2h are not ordered against the compute stream

1bit-MONSTER/engine#696 · opened

on Oct 31, 2026

hw-gate: every GPU step takes flock ~/npu.lock, so the gate cannot run while an NPU serve is up (c16 gate cut 3x by its own 20-min hold)

1bit-MONSTER/engine#682 · opened

on Oct 31, 2026

hrx: a separate MTP draft head (-md) is rejected for qwen4exp on our route, while a cut-vocab head gives +55% decode on the fork

1bit-MONSTER/engine#681 · opened

on Oct 31, 2026

hrx: gfxhub page fault + MODE2 GPU reset when GTT nears the 120 GiB ceiling (Flash-Next UD-IQ3_XXS, 18.8k ctx, 106.5 GiB GTT)

1bit-MONSTER/engine#680 · opened

on Oct 31, 2026

hrx: long-context decode 9.70 tok/s at 18.8k ctx (prefill 74-93) vs 31.15 / 459 on a gfx1151 fork — same GGUF, same hour, short-prompt parity 28.6 vs 28.9

1bit-MONSTER/engine#679 · opened

on Oct 31, 2026

2

hrx: any graph whose weight ops all run on the CPU computes garbage (1 s reproducer)

1bit-MONSTER/engine#674 · opened

on Oct 31, 2026

1

npu: load Ryzen AI OGA NPU model packs on Linux through the engine (pack import -> Q4NX -> our kernels)

1bit-MONSTER/engine#671 · opened

on Oct 30, 2026

1

HRX route pins ~107 GiB outside process RSS on a 217 GiB mmap (DeepSeek-V4.1 UD-IQ1_S), independent of -ngl/ubatch/prefetch

1bit-MONSTER/engine#669 · opened

on Oct 30, 2026

4

npu: readback can observe stale data after the ERT completion signal; needs a fence or verify-and-reissue

1bit-MONSTER/engine#666 · opened

on Oct 30, 2026

4

test-backend-ops (fork pin 40c4ccb): a run that executes zero cases exits 0 and prints OK

1bit-MONSTER/engine#657 · opened

on Oct 30, 2026

1

npu: two resident engines need disjoint-array-half kernel sets (per-engine column partition)

1bit-MONSTER/engine#656 · opened

on Oct 30, 2026

2

npu: /v1/systemone decision endpoint (Decision 3.0 / d3 packs) — the gap that keeps d3 out of the catalog

1bit-MONSTER/engine#653 · opened

on Oct 30, 2026

1

npu: constrained decoding past GBNF — tool_choice, parallel_tool_calls, json_schema, and one answer parser for every family (FLM v1.0.10)

1bit-MONSTER/engine#652 · opened

on Oct 30, 2026

1

npu: logprobs / top_logprobs on /v1/chat/completions (FLM v1.0.11)

1bit-MONSTER/engine#651 · opened

on Oct 30, 2026

1

npu: no multi-prompt prefix cache — every request prefills from scratch (FLM v1.0.11 has 8 slots)

1bit-MONSTER/engine#650 · opened

on Oct 30, 2026

1

census: new HF model class uncovered by the registry

1bit-MONSTER/engine#619 · opened

on Oct 27, 2026

7

NPU: Qwen3.8-27B-NPU2 (FastFlowLM Q4NX, 17.1 GB) on the NPU route and as the hybrid prefill side

1bit-MONSTER/engine#587 · opened

on Oct 26, 2026

7

HRX: Qwen3.8-Flash-Next prefill is 287 tok/s on UD-IQ4_XS vs 1,679 for gufo on the same box (and 500-2,400 across published Strix Halo engines)

1bit-MONSTER/engine#572 · opened

on Oct 25, 2026

8

Navigation Menu

Repository navigation

Issues

tracking: zero-copy on the serving paths (hybrid KV handoff, NPU activations in BOs, ONNX EP IO binding, audit)

npu: the pb16i/sealed-ELF path emits garbage — 48 launches/token missing, 16 int8 M=128 QKV launches on full-attention layers (blocks #692)

hrx: hrx_synchronous_h2d/d2h are not ordered against the compute stream

hw-gate: every GPU step takes flock ~/npu.lock, so the gate cannot run while an NPU serve is up (c16 gate cut 3x by its own 20-min hold)

hrx: a separate MTP draft head (-md) is rejected for qwen4exp on our route, while a cut-vocab head gives +55% decode on the fork

hrx: gfxhub page fault + MODE2 GPU reset when GTT nears the 120 GiB ceiling (Flash-Next UD-IQ3_XXS, 18.8k ctx, 106.5 GiB GTT)

hrx: long-context decode 9.70 tok/s at 18.8k ctx (prefill 74-93) vs 31.15 / 459 on a gfx1151 fork — same GGUF, same hour, short-prompt parity 28.6 vs 28.9

hrx: any graph whose weight ops all run on the CPU computes garbage (1 s reproducer)

npu: load Ryzen AI OGA NPU model packs on Linux through the engine (pack import -> Q4NX -> our kernels)

HRX route pins ~107 GiB outside process RSS on a 217 GiB mmap (DeepSeek-V4.1 UD-IQ1_S), independent of -ngl/ubatch/prefetch

npu: readback can observe stale data after the ERT completion signal; needs a fence or verify-and-reissue

test-backend-ops (fork pin 40c4ccb): a run that executes zero cases exits 0 and prints OK

npu: two resident engines need disjoint-array-half kernel sets (per-engine column partition)

npu: /v1/systemone decision endpoint (Decision 3.0 / d3 packs) — the gap that keeps d3 out of the catalog

npu: constrained decoding past GBNF — tool_choice, parallel_tool_calls, json_schema, and one answer parser for every family (FLM v1.0.10)

npu: logprobs / top_logprobs on /v1/chat/completions (FLM v1.0.11)

npu: no multi-prompt prefix cache — every request prefills from scratch (FLM v1.0.11 has 8 slots)

census: new HF model class uncovered by the registry

NPU: Qwen3.8-27B-NPU2 (FastFlowLM Q4NX, 17.1 GB) on the NPU route and as the hybrid prefill side

HRX: Qwen3.8-Flash-Next prefill is 287 tok/s on UD-IQ4_XS vs 1,679 for gufo on the same box (and 500-2,400 across published Strix Halo engines)