[jschmied]

With TF_QWEN_KV_PAGES=n and TF_QWEN_KV_PAGES_ALL=1, main runs every stream on pool pages, but it still reserves the engine's own window, a full stream cache at --context, and leaves it idle. The drafters also keep one ring of their own. With this change the engine's own stream becomes a pool lane too (engine_paged), as on GLM, Gemma 4 (#1619) and Kolibri-1 (#1620). The plan sets one window's worth of pages aside for it, no window is reserved, and DFlash2 and MTP keep their rings in the lane runs. Without _ALL, nothing changes. This follows #1156 and #1357 (the lane count fix, d60cc24, is already on main).

• attach makes streams + 1 lanes when the engine's stream is paged; Loaded.engine_paged tells the host's plan.
• reserve makes no window then: shapes warm on the first pool lane.
• DFlash2's window timing gets null and makes its own window_rows cache, as cost_adapter.resolve already does for MTP.
• RingOf.all: neither drafter keeps a ring of its own.

Startup line at --parallel 8 --context 16384:

	main _ALL	this _ALL
Qwen3.8-27B	all 8 on 959 pages (14.98 GiB), the engine's own window (1.90 GiB) idle	the engine's own and 7 more on 1039 pages (16.23 GiB), no window
Qwen3.6-35B-A3B	all 8 on 994 pages (4.85 GiB), the engine's own window (0.59 GiB) idle	the engine's own and 7 more on 1074 pages (5.24 GiB), no window

No kernel changed. The 27B and 35B attend over pages through tf_q35_attn_partial_paged (qwen35_attn.cuh), not through window_rows, so the paged merge that slowed #1653 (tf_wrows_merge_256_pg, fixed there in 7c0a80b) is not on this path.

Gates (GB10, sm_121)

These are the #1357 gate suite: 37 greedy outputs per arm (solo, concurrent 1/2/4/8, resumes at 8192/8292/8448/4796), --parallel 8 --context 16384, a fresh server per arm. Models: Qwen3.8-27B MLX-4bit with DFlash2, and Qwen3.6-35B-A3B MLX-4bit with its MTP layer, each with bf16 and fp8 KV. Main is the merge base 5d19272.

gate	27B bf16	27B fp8	35B bf16	35B fp8
unset: this == main	37/37	37/37	37/37	37/37
_ALL: this == main _ALL (solo, resumes)	22/22	22/22	22/22	22/22
_ALL == contiguous (solo, resumes)	22/22	22/22	22/22	22/22
_ALL: concurrent == solo	15/15	15/15	15/15	15/15
_ALL: resumed == main _ALL	12/12	12/12	12/12	12/12
_ALL: drafted == plain	22/22	22/22	22/22	22/22
PAGES alone: concurrent == window or page solo	15/15	15/15	15/15	15/15

Speed against main (GB10, sm_121)

Setup: same models and drafters, bf16 KV, --parallel 8 --context 16384 --prompt-cache-gib 0, a fresh server per arm. The four arms (main, this, main _ALL, this _ALL) alternate in ABBA order: 3 rounds for the 27B, 4 for the 35B.

• Prefill: token-id prompts of 2,048, 9,216 and 15,000 tokens, 3 each per server; a server's value is the median.
• Decode: 8 fixed chat prompts × 256 tokens (ignore_eos), 1, 4 or 8 in flight; aggregate tok/s = completion tokens / wall.
• Deltas: medians over rounds, with the per-round paired deltas in brackets.
• Tokens: greedy token_sha equal across arms in every cell.

Default (no env):

cell	27B main	27B this	delta (per round)	35B main	35B this	delta (per round)
prefill 2,048 (tok/s)	1,407.9	1,414.1	+0.44 % (+0.5, +0.1, -0.3)	5,091.2	5,154.2	+1.24 % (-0.2, -0.1, +2.0, +3.2)
prefill 9,216 (tok/s)	1,365.3	1,372.1	+0.50 % (+0.7, +0.2, -0.7)	5,459.8	5,509.9	+0.92 % (+0.2, -0.2, +1.1, +3.4)
prefill 15,000 (tok/s)	1,348.0	1,342.8	-0.39 % (-0.4, -0.2, -0.6)	5,390.6	5,428.5	+0.70 % (+0.2, -0.0, +1.0, +1.9)
1 stream, code, T 1.0	66.3	66.9	+0.82 % (+3.4, +0.2, +0.7)	131.9	132.0	+0.06 % (+0.3, -0.3, -0.7, +1.3)
1 stream, code, greedy	73.3	73.8	+0.64 % (+3.5, +0.2, +0.6)	134.1	132.9	-0.91 % (-0.8, -1.7, -1.5, +0.4)
1 stream, chat, T 1.0	42.4	42.8	+0.74 % (+3.5, +0.3, +0.7)	113.7	112.4	-1.15 % (-2.6, -0.9, -2.2, +1.0)
1 stream, chat, greedy	45.8	46.9	+2.35 % (+3.5, +0.5, +2.3)	115.6	115.0	-0.56 % (-1.1, -0.2, -2.0, +0.0)
4 streams, code, T 1.0	177.7	180.2	+1.38 % (+1.8, -0.3, +1.4)	295.3	293.8	-0.52 % (-0.5, -0.3, -1.5, +0.4)
4 streams, code, greedy	181.9	185.6	+2.04 % (+3.0, +0.0, +1.3)	293.0	291.4	-0.56 % (-0.6, -0.3, -2.4, +0.3)
4 streams, chat, T 1.0	114.3	115.5	+1.11 % (+1.7, +0.7, +1.1)	256.4	255.0	-0.54 % (-0.9, -0.1, -1.1, +0.4)
4 streams, chat, greedy	130.6	129.4	-0.92 % (+1.9, -1.7, -1.0)	261.4	260.3	-0.42 % (-0.6, -0.3, -0.3, +0.4)
8 streams, code, T 1.0	242.4	242.9	+0.23 % (+1.4, +0.2, -0.8)	397.8	396.5	-0.32 % (-0.4, -0.2, -0.4, +0.2)
8 streams, code, greedy	255.0	256.3	+0.49 % (+1.5, +0.1, +0.5)	399.1	397.3	-0.46 % (-0.5, -0.5, -0.2, +0.2)
8 streams, chat, T 1.0	170.0	170.9	+0.53 % (+1.6, +0.0, +0.5)	347.2	345.9	-0.37 % (-0.2, -0.4, -0.7, +0.5)
8 streams, chat, greedy	181.6	182.5	+0.47 % (+1.8, +0.7, +0.4)	360.6	359.3	-0.36 % (-0.2, -0.7, -0.1, +0.2)

Without the env, the decode path is the same code (the new fields are read only at startup), and the 103 fatbins embedded in the two binaries are byte-identical (md5). Two cells still had one sign in every round, so I checked both:

• 35B decode was below main in rounds 1-3. Round 4, run in a separate block with an A/A arm (main's binary on a second server), came out above main in every cell, so no cell keeps its sign. The A/A arm itself differed from main by -1.6 to +3.4 % per cell. A graph-level nsys trace (--cuda-graph-trace=graph, 1 and 8 streams of code greedy, main/this/this/main) shows the same GPU work per decode round: busy time per tf_q35_argmax launch was 4.83 / 4.78 / 4.80 / 4.83 ms. The rest is idle time between steps, and it varies with the server, not with the binary.
• 27B prefill at 15,000 was below main in all three rounds (-0.4 %). A prefill-only rerun (3 more rounds, 5 prompts each, plus the A/A arm) gave -0.52 / -0.05 / -0.03 % against main, and the A/A arm itself gave -0.62 / +0.23 / +0.10 %. Under nsys, three 15,000-token prompts kept the GPU busy for 99.9 % of their span in every arm. The span was 37,321 and 36,635 ms on main against 36,842 and 36,807 ms here, with the same kernels.

TF_QWEN_KV_PAGES=256 TF_QWEN_KV_PAGES_ALL=1 (opt-in):

cell	27B main	27B this	delta (per round)	35B main	35B this	delta (per round)
prefill 2,048 (tok/s)	1,384.3	1,391.3	+0.51 % (+1.3, -2.3, +0.6)	5,105.3	5,081.0	-0.48 % (-0.1, -0.1, -1.1)
prefill 9,216 (tok/s)	1,335.8	1,339.6	+0.29 % (+1.1, -1.6, +0.3)	5,406.0	5,377.7	-0.52 % (-0.2, +0.1, -0.7)
prefill 15,000 (tok/s)	1,305.9	1,301.5	-0.34 % (+1.5, -1.8, -0.3)	5,275.8	5,240.2	-0.68 % (-0.3, -0.1, -0.8)
1 stream, code, T 1.0	67.0	66.4	-0.93 % (-1.2, -1.1, -0.1)	132.8	132.0	-0.60 % (-0.1, -0.5, -0.6)
1 stream, code, greedy	74.0	73.3	-0.98 % (-1.4, -0.5, -0.0)	133.8	132.2	-1.20 % (-1.7, -1.2, +0.5)
1 stream, chat, T 1.0	42.9	42.5	-0.91 % (-1.3, -0.4, +0.1)	113.9	113.5	-0.31 % (-0.9, -0.4, +2.0)
1 stream, chat, greedy	47.0	46.7	-0.80 % (-1.3, -0.5, +0.1)	114.3	114.4	+0.12 % (-0.3, -1.4, +0.6)
4 streams, code, T 1.0	178.9	176.6	-1.27 % (-1.4, +0.2, -0.5)	293.2	295.5	+0.76 % (-0.2, +0.2, +1.3)
4 streams, code, greedy	185.3	184.0	-0.65 % (-2.1, -0.2, +0.2)	290.7	293.1	+0.83 % (-0.1, +0.3, +1.3)
4 streams, chat, T 1.0	114.7	114.1	-0.48 % (-1.4, -0.5, -0.4)	254.7	257.0	+0.92 % (-0.0, +0.2, +1.4)
4 streams, chat, greedy	130.3	128.3	-1.57 % (-2.5, -0.1, -1.0)	259.3	261.3	+0.77 % (-0.1, +0.1, +1.3)
8 streams, code, T 1.0	241.6	242.2	+0.26 % (-0.0, +0.6, +0.7)	395.0	397.6	+0.66 % (-0.3, +0.2, +1.5)
8 streams, code, greedy	255.1	253.7	-0.57 % (-1.1, -0.5, +0.2)	394.6	398.2	+0.91 % (-0.4, +0.1, +1.3)
8 streams, chat, T 1.0	170.1	169.2	-0.51 % (-1.2, -0.3, +0.2)	345.7	346.7	+0.28 % (-0.6, +0.0, +1.1)
8 streams, chat, greedy	181.7	180.3	-0.74 % (-1.1, -0.1, -0.2)	358.1	360.1	+0.54 % (-0.2, +0.3, +1.0)

Under _ALL, a lone stream is now the engine's own stream on pages, where main ran it on the reserved window. The 27B's 1-stream cells came out 0.8-1.0 % lower in their medians, but no cell held its sign in all three rounds. I did not trace that case. The 35B's third _ALL main arm was rerun later, because the first run was lost when its loop was stopped.

Checks

• zig build test test-kernel-contract test-golden test-server-security test-pull-security fuzz-cortheon and git diff --check: pass on this head. They also pass on this head merged into main 94b90c9. The branch merges cleanly there and into e5b50c8 too, so it was not rebased.
• Lean: 0 new problems.
• No kernel file changed.

Whether pages should be the default for qwen3_5, as #1619 and #1620 propose for their families, is left open here.