[merceod]

Stacked on #1102 (pr-36-t6-fused-draft).

Why

Decode linears on ROCm were routed by three configs (decode_gemv, decode_gemv_silu_mul, decode_linear) and a chain of passes, each deciding for its own kernel. Nothing put AITER's skinny GEMM, the Triton GEMV and tuned hipBLASLt side by side for one size, and moving a size to another kernel meant editing a different file per kernel.

What changed

Stacked on #1102, nothing picked first: the table replaces the decode GEMV configs #1102 added rows 3, 6, 12 and the fc to, and the last commit puts those rows into the table.

• KernelChoice in mstar/kernels/registry.py reads one per-device table, configs/gfx942/mi325x/decode_gemm.json, which replaces the three configs. Each entry keeps every candidate's time and parameters beside the choice, so moving a size is a one-word edit. A fused gate/up has its own gated1- entry.
• mstar/platforms/rocm_gemm.py: one pass route for AITER, Triton, tuned hipBLASLt and plain hipBLASLt at rows 1-128 in decode buckets. DecodeLinearPass and the pass chain are gone. The route keys on a bucket's rows, so Speculative decoding on Qwen3.5: a fused draft step and an fp8 draft head #1102's verify buckets (8 rows, 24 tokens) keep their passes. Tuned solutions apply only on the stack they were tuned on.
• python -m benchmark.kernels.promote decode_gemm --store DIR writes the table from the harness store. A size leaves what the tree serves only for a 2 % gain and one graph node (a fused gate/up for the 2 % alone): without the node floor a first table moved 2B's rows-1 out_proj to Triton for 0.17 us a call and broke an Inductor fold (247 to 265 kernels a step).
• 185 entries: 159 from the search (144 sizes over 9B, 2B and 0.8B, 15 gate/up), Speculative decoding on Qwen3.5: a fused draft step and an fp8 draft head #1102's 16 rows with the kernel and knobs it shipped, and 10 naming the class default. The search store alone gives back the 159 unchanged.
• Off ROCm skinny_gemm_passes() returns None before any table read, and products without a table keep today's defaults. Docs in docs/kernels.rst.

Verification

• Window I (job 551), 8x MI325X, exclusive, text-chat-256, three timed runs per lane on two lanes in opposite orders, median tok/s, f7122c3 against the same tree with the 159-entry table (7806537, no Speculative decoding on Qwen3.5: a fused draft step and an fp8 draft head #1102 rows):

model	c=1	c=4	c=8	c=16	c=32	c=48	c=64
9B	224 / 224	766 / 767	1473 / 1467	2586 / 2657	4313 / 4636	5793 / 6060	7139 / 7273
2B	608 / 619	2035 / 2026	3824 / 3765	6760 / 6816	10585 / 11193	14585 / 15006	17408 / 18453
0.8B	880 / 839	2601 / 2639	4881 / 5020	8819 / 8923	13974 / 14243	19167 / 19254	23107 / 22923

• 9B +2.7 % at c=16, +7.5 % at c=32, +4.6 % at c=48, +1.9 % at c=64, ITL p50 7.10 to 6.56 ms at c=32. 2B +5.7 % at c=32 and +6.0 % at c=64 (ITL 3.27 to 3.11 ms), -1.5 % at c=8. 0.8B +1 to +3 % at c=4-32, flat at c=48-64, and -2 to -5 % at c=1 on the median, where both arms carry stalled runs (best run -1.9 %).
• Stalled runs are as common on the tip as here: 41 of 144 against 40 of 138. Two losses a shared node showed (2B c=1 -20 %, 9B c=4 -34 %) did not come back on the exclusive node.
• On the week-3 tip with Speculative decoding on Qwen3.5: a fused draft step and an fp8 draft head #1102's rows in, one job at a time on an empty queue, direction only: 9B 225.9 against 223.8 tok/s at c=1 and 7287 against 7147 at c=64 (+0.6 %, +1.8 %). 9B mtp2 reads 409 against 410 at 2.45 tokens a chunk, so Speculative decoding on Qwen3.5: a fused draft step and an fp8 draft head #1102's path keeps its numbers.
• Correctness: 25 greedy passes on 9B and 2B with 0 unexplained divergences, and greedy PASS on every window I arm. 1416 GPU cases check every kernel the table holds against the parity bar, and each served one from a graph replay.
• Tests, from the branch and integration trees, not this cut branch: CPU Core failing set equal to the base's (34) on the branch and on the week-3 tip, test/rocm 1965 passed and 9 skipped on the tip with the merged table, ruff clean.

Not in this PR

• The 0.8B c=1 -2 % is unexplained. Its rows-1 entries come from one search pass, and a rerun of those 7 sizes on a quiet node changed no pick. The two passes disagree on 9B rows-16 gate/up, which rests on the pooled time.
• Rows 96 and 128 keep hipBLASLt's pick, and the Triton kernels weren't tried past 16 rows. With Speculative decoding on Qwen3.5: a fused draft step and an fp8 draft head #1102's rows in, only 9B was served (c=1, c=64, mtp2 c=1): no 2B, 0.8B, 9B c=16 or c=32 yet.
• Only MI325X has a table. Norm has nothing to pick (Triton wins at every decode row), and attention needs a bench that times a candidate inside a step first. NVIDIA not run, and nothing there reads the table.

Rebased onto main

Rebased onto main (ff7819d, 2026-11-05); the same 9 commits. The MI325X numbers here were measured on the AMD line before its rebase; the rebased line read level on the Qwen3-TTS, Qwen3-Omni, Qwen3.5 mtp2 and FLUX.2 klein smoke rows, and -8.7 % on the Qwen3.5-0.8B text c=64 row, which waits on the twin-slot re-port (phase 2).