[novkien]

Overview

Add an SM70-only sparse Flash Attention route for the Qwen QSA D256/GQA12/Q8_0 shape, using the 16x4 tile and the existing 2051-token mask bound. The route keeps full-Q=1024 and unsupported shapes on existing paths. Deterministic test fixtures cover disjoint and sliding-window masks.

Addresses #9

Validation

• Native V100 candidate CPU-reference checks: 2/2 passed. RPC RTX 5060 Ti positive checks: 2/2 passed. The 14-case fallback set passed 14/14, including 11 merged-base Volta regressions and three sparse guard controls.
• Five alternating operator pairs cover all eight KV-length, query-batch, and mask-pattern cells. Speedup is baseline time divided by candidate time; intervals use exact enumeration of 3,125 paired resamples.

KV length	Queries	Mask pattern	Speedup [95% CI]
131072	512	Disjoint regions	2.2763 [2.2716, 2.2805]
200704	512	Disjoint regions	3.4384 [3.4294, 3.4475]
131072	512	Sliding window	28.6091 [28.5778, 28.6405]
200704	512	Sliding window	36.9951 [36.8919, 37.0777]
131072	1024	Disjoint regions	1.0005 [0.9999, 1.0011]
200704	1024	Disjoint regions	0.9994 [0.9985, 1.0004]
131072	1024	Sliding window	0.9981 [0.9940, 1.0017]
200704	1024	Sliding window	1.0014 [0.9970, 1.0075]

A bounded trace of an actual long-context proxy prefill recorded 20 sparse Flash launches across all four V100s, including Q=1020 with KV=70144 and Q=326 with KV=69120. This confirms route selection for the traced shapes; the trace package was excluded from performance measurements. Cold 8k/32k K/V lengths remain below the sparse threshold, and full Q=1024 is explicitly excluded. These operator results do not claim an end-to-end model speedup.

The PR is ready for review. Final combined deployment, near-full-context, and performance acceptance remain with the root owner; issue #9 stays open until that work is recorded.

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: YES. Codex implemented and reviewed this change under the owner's approved plan. The owner authorized publication and is responsible for the submission; owner code review is not claimed.

[novkien]

CI triage and handoff (2026-10-03)

The two completed Windows failures repeat the same signatures seen on PR17:

• The current Windows Server job fails unit/test_completion.py::test_completion_unified[256-4-n_predict_vals3-expected_success3] with HTTP 500, Failed to restore prompt cache state; the same case, result, and 334-passed/4-skipped count appear in the PR17 Server job. PR21 changes CUDA attention and backend-op tests, not the server prompt-cache path.
• The current Windows x64 OpenBLAS job fails test-thread-safety using tinyllamas/stories15M-q4_0.gguf: it reports all threads finished without errors, then exits with 0xc0000374. The same test/model/outcome appears in the PR17 OpenBLAS job. PR21 does not change that test or its CPU code path.

These are recurring signatures, not a confirmed root cause or proof of an infrastructure failure. The owner authorized marking this reviewed head ready and squash-merging it after the exact-head, issue, and base checks. Root retains the combined build, deployment, performance, and context acceptance; issue #9 remains open.