perf(qwen3.6): Q4_K on-read LM head — requantize the Q6_K output projection at load

[minion1227]

What

Requantize the Qwen3.6 output projection (LM head) from Q6_K to Q4_K once at model load, then decode it on-read via the existing int8 dp4a Q4_K MMVQ. The LM head (vocab 248320 × hidden 2048) is the single largest weight read per decoded token; Q6_K→Q4_K trims that read ~31% (6.5625 → 4.5 bpw).

Why

The output projection runs once per token and is the largest single weight read on the decode path (~11–12% of decode bandwidth). Every other big projection already runs int8 dp4a / split-K; the LM head was still full Q6_K on-read.

How

• New kernels/csrc/cuda/quant/lmhead_requant.cu — a per-32-value affine k-quant fit (weighted least-squares over a short offset ladder) that emits standard Q4_K super-blocks. Runs once at load; no per-token cost.
• Load path (qwen35.cpp) — dequant the Q6_K output → requant to Q4_K → swap in, set lm_head_type = 12. Scoped to the Qwen3.6 output (vocab 248320), so the Qwen3-30B guard model stays on its original path.
• Forward — a Q4_K branch reuses launch_mmvq_q4k_f32 against the pre-quantized Q8_1 activation.
• SPARKINFER_LMHEAD_Q4K=0 restores the Q6_K path.

Results — RTX 5090 (sm_120), same-box A/B (SPARKINFER_LMHEAD_Q4K=0 vs =1)

context	Q6_K tok/s	Q4_K tok/s	Δ
128	280.08	289.47	+3.35%
512	277.42	286.59	+3.31%
4k	265.78	273.91	+3.06%

No regression at any context (uniform +3%, since it only shrinks a per-token read).

Accuracy vs llama.cpp (pinned reference 6f4f53f, teacher-forced, 100 positions):

• top-1 token agreement: 0.970 (bar ≥ 0.90)
• mean KL(llama‖spark): 0.0197 nats (bar ≤ 0.20)
• PPL: 2.83 → 2.86

Correctness stays well inside the gate because the reduction is confined to the final projection and the k-quant fit is weighted toward the high-magnitude logits.

[skyrocket2026]

⏸ Merge conflict — rebase before eval

This branch conflicts with main, so the RTX 5090 eval is skipped until you rebase.

Please rebase onto main and push — the bot re-evaluates on the next poll once the PR is cleanly mergeable.

[skyrocket2026]

The round's merge-first PR was just merged. Please rebase this branch onto main — once you push the rebase the bot re-evaluates it against the new frontier (crediting your marginal gain on top of what merged).