perf: cascaded N=3 MTP + share-warp Q8 — GB10 decode past 20 t/s (stacked on #14)

[TrevorS]

Stacked on #14 (gb10-mtp-combined-forward).

Summary

Two CUDA-only, weight-preserving decode-perf changes on top of combined-forward:

1. share-warp Q8 weight hoist — load each Q8_0 weight row's quants once per block and reuse across the N batched tokens, instead of re-reading them per token.
2. cascaded N=3 combined-forward — verify [first_token, drafts[0], drafts[1]] in one batched main-model pass instead of [first_token, drafts[0]], lifting accepted tokens/iter.

Together they push greedy --mtp decode past 20 t/s on the standard short prompt and to ~22 t/s at long context.

Speed — DGX Spark / GB10 (ds4flash, --mtp, temp 0, --nothink)

prompt	#14 (N=2)	this PR	Δ
knight, n=256	18.85	20.54	+1.69
chat prompt, n=200	19.39	21.66	+2.27
long-context ~30k	17.45	22.19	+4.74

knight is a 5-run mean; tokens/iter 1.79 → 2.24.

What's in the PR

1. share-warp Q8 weight hoist

matmul_q8_0_preq_batch_share_warp_kernel was calling the per-token dot, re-reading the row's Q8 weight quants N times — the one thing the share-warp kernel exists to avoid. Hoist the 8 int32 weight words into registers once per block and dp4a them against each token. Bit-identical: exact int32 dot, unchanged accumulation order.

2. cascaded N=3 combined-forward

• drafts[1] is chained off drafts[0] by threading the MTP block's own output HC (mtp_state_hc → mtp_next_hc), not the iter-start combined_prev_hc, so the verifier accepts it ~0.58 of the time given drafts[0] accepted — enough to pay for the extra row.
• Partial-accept rollback gains a prefix2 snapshot (compressor frontier at the row-1 boundary) for the commit==1 case (drafts[0] kept, drafts[1] rejected). prefix1/prefix2 share one restore helper.
• Fixes a latent bug the N=3 window is the first caller to hit: the verifier's top_rows ≥ 2 readback took the top-N of row 0 only and routed through the generic single-thread indexer_topk (~17 ms/call over the 128k vocab). Replaced with one fast tree-reduce argmax per row.
• DS4_MTP_NO_CASCADE=1 forces the prior two-token window (kill switch).

Correctness

The verifier defines the exact greedy stream, so the N=3 output is byte-identical to the two-token window on every prompt tested, including ~30k-context prompts where the compressor frontier is fully active and the prefix2 rewind fires for real. Strict mode (DS4_MTP_STRICT / --quality) is unchanged — it never enters combined-forward and stays byte-exact to plain decode.

Tested

• make clean && make cuda-spark — clean; make cpu — clean
• ./ds4_test --long-context, --tool-call-quality, --server — OK
• N=3 ≡ DS4_MTP_NO_CASCADE and strict ≡ plain, byte-identical at short / second-prompt / ~30k-context
• Two ds4_test checks fail identically on stock upstream/main — not introduced here, and outside the spec-decode path (both run plain/batch decode, not --mtp):
  • --logprob-vectors short_code_completion — markdown code-fence language-tag case near-tie (```c vs ```C); the generated code is byte-identical.
  • --metal-tensor-equivalence — MoE routed-expert down-projection float atomicAdd at n_tokens >= 128 (use_atomic_down) drifts at ulp scale across runs and occasionally flips a greedy argmax at long context; DS4_CUDA_MOE_NO_ATOMIC_DOWN=1 makes it bit-exact.

Hardware: NVIDIA DGX Spark (GB10 / sm_121), driver 580.142, CUDA 13.0
Model: DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf
MTP: DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf

AGENT.md compliance

• Preserve correctness before speed — output byte-identical to the prior decode path; strict mode preserves the byte-exact canonical path.
• No permanent semantic variants behind flags — DS4_MTP_NO_CASCADE is a diagnostic kill switch that reproduces the prior two-token release path, not a new permanent mode.