[bong-water-water-bong]

Unsloth's sub-4-bit UD files mix IQ1/IQ2/IQ3 types inside FFN gate/up pairs. Until now, a pair whose two formats need different codebooks was declined by the fused K-quant SwiGLU decode kernel, because the kernel staged only one codebook in workgroup memory. Those pairs fell back to the generic path, and that tracked the slow sub-3-bit decode on the leaderboard.

Mixed pairs per file and decode speed (FFN gate/up pairs needing two different codebooks, out of 64, from each file's tensor types; tg128 on dd74f6b):

File	Mixed pairs	tg128 (tok/s)
UD-Q4_K_XL	0	9.71
UD-IQ1_S	6	5.54
UD-IQ1_M	12	3.52
UD-IQ2_XXS	14	2.09
UD-IQ2_S	24	1.74
UD-Q2_K_XL	27	2.02
UD-IQ3_XXS	30	2.56

Change:

• Both SwiGLU decode kernels (1 token and 2–8 tokens) allocate a 4 KiB codebook buffer each for gate and up, and fill each one for its own format. Same-format pairs fill both buffers.
• The matcher no longer declines mixed pairs, and the kquant_grid helper is removed.

Measured on strixhalo (HRX0, balanced power mode, interleaved rounds):

	dd74f6b	this PR
Qwen3.8-27B UD-IQ2_S tg128	3.12 / 3.12	8.32 / 8.31
27B UD-Q4_K_XL tg128 (no mixed pairs)	11.88 / 11.88	11.81 / 11.87
ZAYA1-8B Q4_K_M tg128	93.0 / 93.2	92.4–95.4* / 92.5

* The first round's first sample was 37.6 (JIT warm-up).

• Logits on UD-IQ2_S against the previous build: KLD 0.00013, same top token 99.1%. That run used 3 × 512 tokens with 8-token micro-batches, so decode goes through the K-quant kernels.
• test-backend-ops -b HRX0: 996/996.

🤖 Generated with Claude Code

[bong-water-water-bong]

Review (PR-Agent duty). Correct.

• Gate and up now each read their own 4 KiB codebook (%grid / %grid_u), each filled only when its format needs one. That makes kquant_grid() and the mixed-pair decline in dispatch-kquant-decode.cpp unnecessary, and removing them is right.
• The evidence covers it: UD-IQ2_S tg128 goes from 3.12 to 8.32 tok/s, with KLD 0.00013 against the previous pin. Uniform K-quant files are unchanged (27B 11.88 vs 11.84, ZAYA within noise), and test-backend-ops passes 996/996.

One measurement before merge: a uniform IQ file whose gate and up use the same grid (for example a pure IQ2_XXS or IQ3_XXS model, as in #51). Every SwiGLU workgroup now fills the same codebook twice. If that costs anything measurable there, fill %grid_u only when up_format's grid differs from gate_format's, and otherwise pass %grid for both (same idea as the old kquant_grid id, now just a select).

The 8 KiB of LDS in the two SwiGLU kernels is fine given the uniform-file numbers.

[bong-water-water-bong]

Pure-format check (review): filling the same codebook twice cost about 1% on a pure IQ3_XXS Qwen3-4B (tg128 32.6 → 32.2). The fix is pushed: when gate and up share a grid (same format, or IQ1_S with IQ1_M), grid_u isn't filled and the up lanes read the gate grid.

Re-measured after the change, HRX0, balanced power mode, three interleaved rounds:

	dd74f6b	this PR
Qwen3-4B pure IQ3_XXS tg128	32.99 / 32.07 / 32.09	31.99 / 31.87 / 32.27
Qwen3.8-27B UD-IQ2_S tg128	3.06 / 3.07 / 3.08	8.38 / 8.30 / 8.34

The first round of each build includes warm-up. Rounds 2 and 3 average 32.08 vs 32.07, so the cost is gone.

• UD-IQ2_S logits against dd74f6b: KLD 0.00013, same top token 99.1%.
• test-backend-ops -b HRX0: 996/996.

🤖 Generated with Claude Code