[jschmied]

What

With TF_QWEN_KV_PAGES on a stream (every stream under TF_QWEN_KV_PAGES_ALL or TF_QWEN_KV_SHARE), a Qwen3.5
prompt chunk currently goes through the scalar paged decode partials plus their merge. That costs about 2x TTFT
(the 27B measurement on #1357's original head a1804ab: cold 8k 7.3 s contiguous vs 15.7 s paged; warm turn 0.75 vs 2.2 s). This PR adds paged twins of the
contiguous prompt kernels and routes one-part paged prompt chunks to them.

• hpattn_paged_kernel / hpattn8_paged_kernel<256, 8, 2, 2> (prefill_attention.cu): hpattn_kernel /
  hpattn8_kernel with CAP = ptok plus a page table. A qwen3_5 page stores each key head's ptok rows
  consecutively (qwen35.cuh kv_row), which matches a contiguous cache of cap = ptok. Only the global address
  each 16-byte chunk is staged from changes: PT[row / CAP] * PS, with the page id looked up once per 32-key
  staging item and a per-row lookup only where an item straddles a page. The kernel folds the same 64-key tiles in
  the same order, so paged == contiguous, bit for bit.
• prefill_attention_rows.cuh: the head-major address takes the page term. For the existing kernels PT is the
  constexpr null, so the term folds away. Every existing function's SASS is unchanged on
  sm_80/86/89/120/121; 0 changed functions across all 97 fatbins (sm_121); 2 added.
• qwen3_5: cuda_prompt.run launches the paged twin when the stream's Seg has a table. cuda_forward takes the
  prompt branch for a one-part paged prompt chunk ahead of the scalar paged partials. Windows of several parts
  (verify) keep the paged partials.

No family changes its bytes. A paged Qwen stream now produces exactly the contiguous stream's prefill outputs,
where today it does not.

Kernel contract

zig/tests/cuda/contract_prompt_pages.zig:

• registers hpattn_kernel / hpattn8_kernel at the 27B's 24/4 heads against float64 (attention_ref bound,
  gates 1-4);
• registers the two paged twins, at pages of 256 and 48 positions shuffled in an arena with spare junk pages, each
  replaces its contiguous kernel bit for bit at 14 shapes, including 4096-row chunks and resumes at 2048/4096;
• plants a red control: a table whose first page id names a later page.

Launch records are in prompt_pages_records.zig (nvcc 13.0, plus 13.2 overrides).

• Filtered run: 9 kernels pass gates 1-4, and the control is red.
• Full run: 346 kernels pass gates 1-4, 0 fail, 95/95 controls red.
  • The 52 limit problems are all pre-existing qwen35 records, the same count as on main.
  • This PR's four launches pass live limits.

Gates (server, greedy, first-row logits hash of every prefill + reply text)

arm	vs	prefill logits	texts	c1/2/4/8 == solo
27B bf16 contiguous (this branch)	main	equal	equal	yes
27B bf16 PAGES_ALL (256)	main contiguous	equal (all)	1 of 17 differs, decode-only	yes
27B bf16 PAGE_TEST=48	main contiguous	equal (all)	same 1 differs	yes
27B bf16 PAGES_ALL, --no-drafts / no prefix cache	main contiguous	equal (all)	same 1 differs	yes
27B bf16 SHARE=8 (on the KV-share stack)	main contiguous	equal (all)	same 1 differs	yes
27B fp8 KV PAGES_ALL	main fp8	equal	equal	yes
35B-A3B PAGES_ALL	main contiguous	equal (all)	2 of 17 differ, decode-only	yes

• Coverage: turn-2 resumes at 4096/8192, a 20k needle, and a cold turn 2 without the cache (== cached).
• drafted == plain, resumed == fresh.
• The reply-text differences come from decode on pages, which still uses the scalar paged decode partials. The
  control shows this predates the PR: main with PAGES_ALL on the same short prompts gives the same divergent text.
  On main, paged prefill logits also differ for those prompts; this PR makes them equal.

Speed (TTFT s, fresh servers alternating, 2+2 for 27B and 3+3 for 35B; texts and logits identical across all arms and rounds)

model	arm	2k	8k	32k	warm turn (7.97k, 4096 cached)
27B	main contiguous	1.86 / 1.63	6.98 / 6.52	29.38 / 27.93	3.26 / 3.23
27B	PAGES_ALL (this PR)	1.60 / 1.60	6.46 / 6.36	27.79 / 27.49	3.17 / 3.16
27B	SHARE=8 (stack)	1.59 / 1.61	6.38 / 6.50	27.37 / 27.44	3.17 / 3.17
35B-A3B	main contiguous	0.389 / 0.393 / 0.385	1.45 / 1.47 / 1.42	6.42 / 6.62 / 6.80	0.704 / 0.727 / 0.698
35B-A3B	PAGES_ALL (this PR)	0.395 / 0.388 / 0.388	1.44 / 1.43 / 1.44	6.61 / 6.56 / 6.62	0.714 / 0.714 / 0.717
35B-A3B	SHARE=8 (stack)	0.390 / 0.388 / 0.390	1.42 / 1.41 / 1.42	6.45 / 6.41 / 6.46	0.713 / 0.709 / 0.713

Paged is within noise of contiguous at every length (the 27B main round 0 is the slower cold first start). For
scale: on the scalar paged partials, a1804ab's own 27B measurement (its prompt set) had cold 8k at 15.7 s paged
vs 7.3 s contiguous, and a warm turn at 2.2 vs 0.75 s.

Dependencies / notes

• Based on main 8c36359. It touches the same lines of qwen3_5/cuda_forward.zig as cuda qwen3_5: a paged prompt chunk's attention goes through most_verify rows at a time #1355 (which slices paged chunks over 128 rows into the partial buffers' size): whichever lands second takes one small conflict there. With this PR, a one-part paged prompt chunk no longer reaches cuda qwen3_5: a paged prompt chunk's attention goes through most_verify rows at a time #1355's path; cuda qwen3_5: a paged prompt chunk's attention goes through most_verify rows at a time #1355's slicing remains needed for multi-part paged windows over 128 rows, which do not occur today.
• This is the tensor-core path cuda qwen3_5: the prompt cache keeps KV pages on the device (TF_QWEN_KV_SHARE, #1156) #1357 (TF_QWEN_KV_SHARE) was held for; the SHARE rows above were measured with cuda qwen3_5: the prompt cache keeps KV pages on the device (TF_QWEN_KV_SHARE, #1156) #1357 stacked on this PR.
• The full contract run predates the last change (the page id looked up once per 32-key item instead of per row); the filtered run of the prompt-pages kernels covers it.
• Follow-up: a paged tensor-core decode tile would remove the remaining decode-only text differences on pages.

[jschmied]

Closing: 0843339 (#1373) gives paged prompt chunks the tensor-core tiles over the page table (and the decode tiles too), and contract_qwen_pages checks them against contiguous, which is what this PR did. The rest here (float64 gates for the twins, shuffled 48-position pages) is not worth a rebase on its own.