record(SPEC-DFLASH2): #1456's conclusion is retracted at the artifact, and the 0.8017 denominator inherits it (#1796)

[localai-bot]

#1456 measured a real source build at CUDA_ARCHS=12.0 emitting sm_80 for
_vllm_fa2_C and sm_75 for _vllm_fa3_C, and concluded that vLLM's vendored
flash-attention cannot target sm_12x, so the GB10 oracle has no FLASH_ATTN
denominator. The measurement stands. The conclusion does not, and showing that
needed no lease. Both staged oracle wheels were opened with zipfile and
walked as fatbinaries on the CPU dev box, and every one of the 76 FA2 and 192
FA3 fatbinaries in each wheel carries a PTX image beside its SASS image. The
first FA2 payload decompresses to .version 9.0 / .target sm_80, which is the
+PTX half of FA2_ARCHS "8.0+PTX" and is the mechanism by which the module CAN
reach sm_121. The artifact establishes a necessary condition, that
forward-JITtable code is shipped; that the JIT then ran is an inference from that
plus a run selecting FLASH_ATTN and generating, which is enough to retract
"cannot target sm_12x" and is not a trace.
cudaErrorUnsupportedPtxVersion is the opposite failure, PTX ISA newer than the
driver. vLLM says the same in its own source:
FlashAttentionBackend.supports_compute_capability returns
capability >= DeviceCapability(8, 0).

The downstream is why this is a change rather than a note. #1456's body records
that the DFlash2 speed gate's denominator would be vLLM pinned to TRITON_ATTN,
by developer decision on 2026-08-20, explicitly because of that conclusion, and
the first DFlash2 speed ratio 0.8016987337853048 was taken against it.
AGENTS.md requires vLLM's production configuration as the denominator, and on
this box vLLM's own auto-selection picks FLASH_ATTN; the gate run's log carries
both branches in one process, cuda.py:426 forcing TRITON_ATTN for the 27B
target and cuda.py:486 auto-selecting FLASH_ATTN for the draft. So if
TRITON_ATTN is the slower backend, which vLLM's priority ordering implies
rather than states and which nothing here measures, the denominator is too low
and the ratio is too high. An
error in our own favour is the one nobody chases, so it is recorded with its
sign.

Nothing is re-measured and no denominator is substituted, because a wave must
not replace a denominator the developer declared. What remains is one lease
running vLLM against itself on this workload with attention_backend=FLASH_ATTN
against TRITON_ATTN, each read back off the built engine, and no GPU lease is
authorised, so it is PENDING.

The same read also settles #1685's three open readings.
vllm/v1/worker/gpu/spec_decode/dflash/utils.py:31-46 sets the draft's
backend=speculative_config.attention_backend unconditionally, so a harness
that sets only the engine backend leaves the draft at None and the draft
re-runs auto-selection; the sibling dspark/utils.py:24-28 and
gemma4/speculator.py:66-89 both carry the target's backend through, and
dspark's comment names the hazard. The DFlash path is the un-defended case, not
an upstream intent.

A fresh review re-derived all twelve cited anchors and confirmed them, with one
enumeration wrong and four notes, and a delta review then caught the repair
shipping a second wrong count. Both are fixed. grep -rln FA_USABLE . --exclude-dir=.git returns three files: .agents/benchmark-record.md, the
DFlash2 spec, and .agents/issue-index.md, where the #1456, #1658 and #1685
rows each quote FA_USABLE=0 as a live constraint. Two of the three are
append-only — the index by rule and gate, benchmark-record.md by its own
opening line, though only as a convention, which is #1373's subject — so the
spec is the only one of the three that can carry the retraction, and ## Owed
O33 says so and names the un-annotated site. What both reviews confirmed
unchanged is that .agents/oracles/ and .agents/upstream-sync.md contain none
of FA_USABLE, FLASH_ATTN or TRITON_ATTN, so no oracle file needs
retracting.

Two departures ride in this branch and are argued rather than assumed. The
appended .agents/issue-index.md row was EDITED by the two repair commits
instead of appended again; scripts/check-issue-index-append-only.py:45-52
diffs merge-base(origin/main, HEAD)..HEAD so it reads the net effect, the row
has never existed on main, and this squash lands one clean append — but a
reader of git log -p on the branch will see the row change shape, which is why
it is said here. And the same implementer who wrote the findings repaired them:
AGENTS.md bars the COORDINATOR from repairing and this is a helper session, but
the same step asks for a FRESH implementer, and these were prose-accuracy
repairs in a records-only change. The freshness that protects correctness held
on the review side — the reviewer re-derived the FA_USABLE enumeration
independently before being told, and re-verified every repaired claim on the new
head.

What is NOT claimed: that an FA2 kernel was traced entering the SM, or that the
forward JIT is free. Construction is proven and coherent output is measured on
both arms; the cost is unmeasured.

Records only, no code. .agents/issue-index.md gains one appended row keyed on
#1796; .agents/specs/dflash2-spec-decode.md gains ## Owed O33 and a fifth
caveat beside the ratio itself. #1456, #1679 and #1685 each received a comment,
and a collapse onto #1685 is recommended there. All three stay open; ending
them is not this branch's call.

Preflight is green apart from test_cpu_x86_llamacpp_floor, which failed with
exit=4 at loadavg 72 on a contended box. That is #618 verbatim and is
unrelated to a records-only change.

Issue: #1796, which stays OPEN because the lease it owes is PENDING. It is
listed under ## Owed O33 of the DFlash2 spec.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]