round 22 track b: router + glue #

delivered #

patch: /tmp/sol_r22_router_glue/sol_r22_router_glue.patch sha256: 87651f461c679bb12d1decbe0f954172d04cda6ca3730f50fcbae6cc59fd3ef4

only metal/kernels/glue.metal changes. tracked vendor sources were not edited. git -C vendor/strata apply --check exits 0. baseline vendor head: 296b59d2f73c8f04a805df29c2fa97655adace78; baseline actual kernel snapshots are in /tmp/sol_r22_router_glue/baseline/ (the checkout may contain prior changes, so snapshots, not head alone, define this experiment).

swiglu_prep moves scale tables to program-scope constants and replaces its 16-float threadgroup array/barrier with same-order shuffle sums. lane 0 still adds even indices and odd indices separately, then adds the two results. all threads execute the shuffle, outside the lane-0 output branch. production dispatch uses 16 threads, entirely within one simdgroup. pipeline-reported static threadgroup memory changes 64 -> 0 bytes; source has one -> zero threadgroup barriers. maximum threads remains 1024. no interface/layout or host dispatch changes.

no measured latency win established. the parent explicitly requested retaining this resource-only reduction after seeing initial results. it is not an end-to-end throughput claim.

rejected hypotheses #

• full route_body unrolling plus constant-index winner invalidation: byte-identical, but consistently slower. confirmed paired median regressions are below. reverted completely, including the entry barrier guard.
• router entry barrier guard alone: +0.3% to +3.8% paired median latency across both formats/windows, not a win. not retained.
• combine n_embd == 2560 constant-division fast path: approximately -1.2% to -0.3%, below the predeclared 3% significance threshold. not retained.
• combine simd-broadcast plan entries and shared sigmoid: +6.2%, +2.2%, +2.6%, and +0.7% for ntok 1/2/4/8, respectively. not retained. this candidate was only evaluated for width 2560, where groups do not cross token boundaries; it was not generalized to arbitrary widths.

no disassembly/register-count evidence was collected. increased register pressure / occupancy loss is a plausible explanation for the unroll regression, not a proven compiler diagnosis. source unrolling does not establish zero stack traffic or prove the original array spilled.

correctness #

• fuse_parity: exit 0, 5.664s, all pass. includes bf16/f32 router logits/ids/w/plan bitwise checks for ntok 1..8 and repeated counter reset; fused norm/combine regressions also pass.
• route_parity: exit 0, 2.180s; 200 k=512 and 200 k=256/moce windows, zero mismatches. this existing cpu-reference test uses a floating-weight tolerance, not memcmp; it is supplementary, not the byte-parity evidence.
• custom paired harness: full output-buffer memcmp, sentinel-filled padding, 8 trials per kernel/window; ntok 1/2/4/8. every comparison passed for logits, ids, w, plan, xp and bo. counter reset also checked. trial zero deliberately produces exact tied router logits; later trials vary activations. router trials alternate 256/512 experts, with/without moce and draft scaling, random-like fixed residency; rejected full unroll additionally passes rank-1 scale 2 specialization.
• delivered source rechecked for router_route_30/router_route_0/moe_combine_sh (delivered_*.log, unchanged kernels) and swiglu (production*.log, changed kernel). swiglu tests the actual width 640, a.n_tok=8 plane stride, and dispatched windows 1/2/4/8; unused rows remain sentinel-filled and match too. early exploratory swiglu runs used width 2560 and are not the production-shape evidence.

stock fuse compilation exceeded two original 4.9s deadlines. after explicit parent authorization for a 20s fuse timeout, its next run passed in 5.664s. all custom benchmark invocations stayed below 5s. no full model, no port 8080 interaction, no subagents.

paired gpu timings #

apple m5; runtime safe math; release appleclang 21.0.0.21000327. same process/library/inputs, separate outputs. final protocol: 0.10s alternating-kernel warmup, ten pairs of 256 dispatches, alternating which side runs first. table reports median baseline/candidate microseconds and median paired percentage (not ratio of medians). clocks/shared-gpu noise are visible in exploratory logs; unchanged combine control differs by about 1%, so changes near that scale are not credible wins.

production swiglu, three independent short invocations/window, 30 pairs pooled:

rejected full-unroll router, default k=512/no moce timed path (the parity trials still cover moce):

unchanged combine control, delivered source:

reproduction / artifacts #

all experimental sources, baseline snapshots, failed attempts and logs: /tmp/sol_r22_router_glue/; source hashes: source_sha256.json. plan.md records the predeclared threshold. ab.metal is the delivered baseline+candidate library; ab_full_unroll.metal, ab_barrier_only.metal, ab_combine_broadcast.metal retain rejected libraries. ab_initial.cpp, ab_v2.cpp, ab_full_confirmation.cpp preserve earlier evaluators; ab.cpp is the delivered evaluator.

cmake -S metal -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build --target fuse_parity route_parity -j 4
# commands above run in /tmp/sol_r22_router_glue/strata
c++ -std=c++20 -O2 -I/tmp/sol_r22_router_glue/strata/metal/src /tmp/sol_r22_router_glue/ab.cpp /tmp/sol_r22_router_glue/strata/build/libmtl_rt.a -framework Metal -framework Foundation -o /tmp/sol_r22_router_glue/ab
/tmp/sol_r22_router_glue/ab swiglu_prep 1
# repeat ntok 2/4/8; other kernels: router_route_30, router_route_0, moe_combine_sh
/tmp/sol_r22_router_glue/strata/build/fuse_parity
/tmp/sol_r22_router_glue/strata/build/route_parity
git -C vendor/strata apply --check /tmp/sol_r22_router_glue/sol_r22_router_glue.patch

current Window::comb_push is at 1083–1091, not the brief's stale 1015–1045 range; no host changes were needed. final image review completed for the entire diff and surrounding glue code; scale-table placement and shuffle idiom match neighboring gr.metal.