[AzeezIsh]

Summary

Two pieces from the ATLAS_QSA_ATTN_L8 gating work:

• spark-model: refuse ATLAS_QSA_ATTN_L8 on atlas_scale — the qsa_prefill_attn_l8 kernel (shared into strix-hip's qsa_indexer.cu by symlink from the GB10 target) destroys the GPU context mid-prefill on gfx1151, not merely diverges numerically. New pure qsa_attn_l8_selectable mirrors qsa_attn_tc2_decision, plus a one-line warn when the env opt-in is refused. NVIDIA selection is unchanged.
• scripts/ppl.py — the gate the _l8/tc/tc2 kernel comments point at: legacy /v1/completions echo+logprobs+max_tokens=0 loglikelihood scorer (stdlib-only), plus --compare for per-token Δlp and ppl ratio between two captures.

Evidence

• winbox plan1006p9, same build both arms (atlas-p5 @ 35b4db9bc): REF (default QSA) scored all 4 × ~18k-token windows — ppl=4.5917 over 73,591 tokens, 0 faults. L8ON serve died mid first window; teardown swept 229,520 allocations and cuMemFree_v2 → status 719. Reported on [job] Qwen3.8-Flash-Next: long-context cold prefill to SGLang parity (stage 1 = #68) #74.
• Unit tests verified on strix-hip (cargo test -p spark-model qsa_attn): qsa_attn_l8_selectable_table + qsa_attn_tc2_decision_table green.
• ppl.py verified live on strix-27B (ppl=5.82 docs window) and winbox Flash-Next (4.5917).

Related: #74 (QSA campaign), #163 comment for the SSM/prefix battery attribution.

🤖 Generated with Devin

[AzeezIsh]

gb10 counterpart check: ATLAS_QSA_ATTN_L8=1 boots and serves Flash-Next
cleanly on GB10 (the kernel's intended hardware — no faults on either arm).
A windowed-perplexity A/B (scripts/ppl.py, corpus ~225k tokens) is queued on
reiner to measure the L8 quality delta vs the default path.

Net: the refusal here is gfx1151-scoped only — on GB10 the opt-in still works
and will have ppl evidence.

(Devin)

[AzeezIsh]

_l8 ppl gate on GB10 — FAILS on numerical drift (its home turf)

Flash-Next on reiner (gb10, postmerge 1d3324989), 8 windows × ~20k tokens
(scripts/ppl.py, corpus ~/code/corpus.txt):

• ppl: 2.4125 → 2.4183 (ratio 1.0024 — under the 2% bar)
• per-token |Δlogprob|: max 6.005 nats (worst: '\n' −1.31→−7.32), mean 0.093
• wall: 395.1s → 417.3s (~5.7% slower — L8 is not even faster here)
• 0 faults on either arm

VERDICT: FAIL — a token moved >0.5 nats. The _l8 path perturbs outputs
measurably even on the hardware it's written for, at no speed gain. The
gfx1151 refusal in this PR stands; on gb10 the opt-in still exists but this
is now evidence it shouldn't be promoted past env-gated without a quality
review.

(Devin — job 744-gb10-l8-ppl-r2)