Skip to content

Actions: penta2himajin/qwisp

Actions

benchtest-ingest

Actions

Loading...
Loading

Show workflow options

Create status badge

Loading
benchtest-ingest.yml
50 workflow runs
50 workflow runs

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
[benchtest] Apple M2 (10-core GPU) · 24GB · streaming benchtest-ingest #50: Issue #170 labeled by jeno007
10s
10s
kv: 8-bit KV cache as a bolt-tier long-context decode lever (Step 0 — measure the ceiling before any kernel) benchtest-ingest #49: Issue #168 labeled by penta2himajin
1s
1s
chat: piped stdin is truncated to the first line (cat file | qwisp chat silently answers on line 1) benchtest-ingest #48: Issue #166 labeled by penta2himajin
1s
1s
strict streaming collapses ~10x on real 16GB Macs (systematic across 5 machines) — investigation + call for a probe run benchtest-ingest #47: Issue #69 labeled by penta2himajin
10s
10s
instrument: ~7.3GB of a prefill's IOAccelerator growth is invisible to both MLX and currentAllocatedSize benchtest-ingest #46: Issue #163 labeled by penta2himajin
9s
9s
serialize: MLX free-buffer pool is unbounded (lane-only cacheLimit) — measured +7.7GB on a long prefill benchtest-ingest #45: Issue #162 labeled by penta2himajin
1s
1s
serialize: expert arenas are never released — bolt's outlives its request, and the segmented path rebuilds one per segment with no clearCache benchtest-ingest #44: Issue #165 labeled by penta2himajin
1s
1s
serialize: expert arenas are never released — bolt's outlives its request, and the segmented path rebuilds one per segment with no clearCache benchtest-ingest #43: Issue #165 labeled by penta2himajin
1s
1s
sizing: budget from actual free memory at startup, not total device RAM (field-driven; expected contentious) benchtest-ingest #42: Issue #164 labeled by penta2himajin
1s
1s
instrument: ~7.3GB of a prefill's IOAccelerator growth is invisible to both MLX and currentAllocatedSize benchtest-ingest #41: Issue #163 labeled by penta2himajin
1s
1s
serialize: MLX free-buffer pool is unbounded (lane-only cacheLimit) — measured +7.7GB on a long prefill benchtest-ingest #40: Issue #162 labeled by penta2himajin
1s
1s
streaming/bolt: concurrent requests to amortise expert I/O (Step 0 — not a port of resident lane mode) benchtest-ingest #39: Issue #160 labeled by penta2himajin
1s
1s
[benchtest] Apple M1 Pro (14-core GPU) · 32GB · resident benchtest-ingest #38: Issue #156 labeled by jonathanlourette
13s
13s
lanes: WS-B step 3 — shared-prefix KV across lanes: value + mechanism feasibility (Step 0, expected NO-GO) benchtest-ingest #37: Issue #154 labeled by penta2himajin
1s
1s
serialize: does Tell.prefill retain the same autoreleased Metal wrappers as the lane path did? (measure before fixing) benchtest-ingest #36: Issue #150 labeled by penta2himajin
7s
7s
lanes: footprint ratchets to OOM under sustained load (predates Stage B; blocks v0.3.10) benchtest-ingest #35: Issue #148 labeled by penta2himajin
7s
7s
perf(arena): expert-major weight relayout — 9 preads per expert miss → 1 (streaming tier I/O count, not bytes) benchtest-ingest #34: Issue #147 labeled by penta2himajin
1s
1s
perf(seedless): concurrent MTLComputeCommandEncoder with an explicit per-layer dependency table — blocked on #143 benchtest-ingest #33: Issue #146 labeled by penta2himajin
6s
6s
perf(seedless): fuse gate qmm8 + route_top8 into one dispatch (−40/token; removes a 1-threadgroup serialization point per layer) benchtest-ingest #32: Issue #145 labeled by penta2himajin
1s
1s
perf(seedless): fold the cross-layer resid_add + input rmsnorm into the existing gdn_resid_postnorm_rows kernel (−39 dispatch/token) benchtest-ingest #31: Issue #144 labeled by penta2himajin
10s
10s
perf: Step 0 — decode dispatch inventory + serial-vs-concurrent encoder bubble probe benchtest-ingest #30: Issue #143 labeled by penta2himajin
1s
1s
[handoff] WS-B Stage B: ctx-adaptive lane admission (lift QWISP_LANE_CTX 16K cap) benchtest-ingest #29: Issue #141 labeled by penta2himajin
Skipped
Skipped
[handoff] WS-B Stage A: token-budget admission scheduler benchtest-ingest #28: Issue #139 labeled by penta2himajin
1s
1s
[handoff] WS-A: matrix-unit attention-prefill kernel (strict L1, re-canonicalization on GO) benchtest-ingest #27: Issue #137 labeled by penta2himajin
1s
1s
prefix-cache-aware admission: parallel sub-agent fan-out should share the prefill, not re-pay it (measured: batch path loses 2.6x to serialize for lack of this) benchtest-ingest #26: Issue #121 labeled by penta2himajin
10s
10s