perf(draft): batch append projections across lanes

[Graffioh]

Depends on #651. While #651 is open, this draft includes its commits. The diff collapses to this follow-up when #651 lands.

Profiling campaign and rerunnable evidence: Luce-Org/luce_box#69.

Summary

Batches the fixed-width draft append projections across active lanes. The graph packs each lane's append feature columns before the shared fc, wk, and wv matrix multiplications. It slices the projected columns before lane-local normalization, RoPE, destination rows, and KV-cache writes.

• replaces the singular append builder with build_draft_kv_appends()
• sends one lane through the normal and exact-bulk graphs, preserving the C1 graph shape
• sends C1-C6 lanes through the packed batch graph
• keeps a_step fixed, so padded rows still target only each lane's trash slot
• preserves the existing graph-cache key over the backend and ordered lane-state pointers
• leaves build_draft_kv_steps(), dynamic-convolution history and apply, attention, masks, and selector ownership unchanged

Isolated draft result on R9700

Qwen3.8-27B-UD-IQ4_XS, DFlash2 Q8 width 8, Q8 K/V cache, ten fixed-width HumanEval code-only waves, 256 forced tokens, temperature 0:

C	Matched baseline draft p50	Packed draft p50	Improvement	Append dispatches/round
1	5.1955 ms	5.1784 ms	0.33%	22 → 22
2	7.1411 ms	6.5494 ms	8.28%	44 → 22
3	9.4684 ms	8.2753 ms	12.60%	66 → 22
4	11.3953 ms	9.7563 ms	14.38%	88 → 22

The dispatch reduction is exact. Eleven append projections each avoid one Q8 quantize and one MMQ launch per additional lane. The remaining block-16 Q8 launches belong to lane-local dynamic convolution and remain out of scope.

End-to-end code-prompt result

This reruns the PR #651 HumanEval-style protocol on the current #651 head and this candidate. Each cell uses a fresh server, a same-C eight-token warmup, temperature 0, natural EOS after DFLASH_MIN_TOKENS=8, a 256-token cap, five server slots, K8 prefill admission, and two measured repetitions. C1-C4 use disjoint slices of the same ten checked-in raw prompts. llama.cpp was not rerun.

The published #651 column is context. The causal comparison uses a fresh build of current #651 at 22b65ba0 because the published table came from the earlier 8ba84d3d restack.

Radeon AI PRO R9700

C	Published #651 tok/s	Current #651 tok/s	Candidate tok/s (range)	vs current	Decode vs current	TTFT vs current	Output parity
1	110.02	110.97	110.57 (110.31-110.83)	-0.36%	-0.73%	-1.21%	yes
2	153.67	154.06	154.00 (153.48-154.53)	-0.03%	+0.04%	-0.85%	yes
3	200.33	201.77	203.68 (203.28-204.08)	+0.94%	+0.77%	-1.68%	yes
4	188.85	190.63	191.80 (191.75-191.85)	+0.61%	+0.65%	-1.73%	yes

Ryzen AI Max+ 395 / Strix Halo

C	Published #651 tok/s	Current #651 tok/s	Candidate tok/s (range)	vs current	Decode vs current	TTFT vs current	Output parity
1	40.63	47.97	47.75 (47.66-47.84)	-0.45%	-0.42%	+0.60%	yes
2	58.69	65.80	65.92 (65.70-66.14)	+0.18%	+0.81%	+0.36%	yes
3	63.52	70.05	70.71 (70.66-70.76)	+0.95%	+1.65%	+0.77%	yes
4	63.20	68.56	69.30 (69.30-69.31)	+1.08%	+1.53%	-0.78%	yes

The end-to-end goodput gain is marginal. This change remains worthwhile because the production implementation is net +29 lines, deletes the singular runtime path, removes 66 append-projection dispatches at C4, and cuts isolated C4 draft compute by 14.38%. TTFT is not a general justification because the Strix C1-C3 result regressed slightly.

All published, current-baseline, candidate, and repeat output sets have identical completion-token counts and output digests at C1-C4.

Correctness

The multilane qualification compares three independent single-lane reference graphs with one packed graph at append counts 0, 9, and 18 for a_step=18. It checks all five K/V cache layers and explicitly verifies padded trash-row writes.

Verified locally:

• HIP build of dflash_server and test_draft_swa_multilane for gfx1151;gfx1201
• test_draft_swa_multilane ... dflash2-q8_0.gguf 0 --append-only
• test_qwen35_roctx
• test_seq_engine_contract
• test_inference_profile
• test_observability
• two-GPU C1-C4 HumanEval output parity and repeat audit
• product and campaign git diff --check
• independent final audit with no P0-P3 findings

Dynamic-convolution coefficient packing and the device-resident selector handoff remain separate measured PRs.