feat(qwen35): support sampled concurrent speculation

[Graffioh]

Stack

This draft PR is built on top of #651 at f1c85e05. It therefore contains the exact current #642 history, #651's projection batching, and the separate replay-log rename.

#654 adds two commits:

1. 0445b161 — feat(sampler): add distribution-preserving spec verification
2. e0647848 — feat(qwen35): speculate sampled concurrent lanes

Both commits are patch-equivalent across the final restack.

Summary

• Keep temperature, top-p, top-k, repetition penalty, frequency penalty, and presence penalty lanes in DFlash2 chain rounds.
• Share one target forward across greedy speculative lanes, sampled speculative lanes, and autoregressive fallback lanes.
• Apply distribution-preserving speculative accept/reject sampling.
• Keep one transactional RNG state per slot and publish it only after prefix promotion succeeds.
• Roll back at the first rejected draft token.
• Commit only the accepted KV, target-feature, GDN, and convolution prefix.
• Batch sampled target-logit transfers by depth, stop transfers after a lane rejects, and reuse sampler buffers.
• Preserve the sparse ordinary AR path for lanes that cannot use the sampled protocol.

For a proposed token x, verification accepts with min(1, p(x) / q(x)). At the first rejection, it samples from normalized max(p - q, 0). If the chain fully accepts, it samples the bonus token from the target distribution on the final verifier row. Target and residual distributions use the same sampler policy as ordinary generation.

Three-lane reference benchmark

These measurements were taken on the previous stacked head 13590205, before the latest #625 refresh. They remain useful as a reference for the unchanged #654 patches, but they are not presented as final-head measurements. A clean rerun was deferred because another user continuously occupied the shared GPU host; contaminated numbers were excluded.

Setup:

• Radeon R9700 (gfx1201)
• Qwen3.8-27B Q4_K_M target
• Q8_0 DFlash2 draft with W8 and SWA 2048
• Concurrency 3, context 4,096, and a 12,288-token KV pool
• 96 output tokens per lane with DFLASH_MIN_TOKENS=96
• One warmup batch, followed by three measured batches per mode
• Aggregate throughput is 288 completion tokens divided by batch wall time

Sampled lanes use temperature 0.8, top-p 0.9, top-k 40, frequency penalty 0.4, and presence penalty 0.2. The mixed batch is greedy, sampled, greedy.

Lanes	DFlash2	One-token AR control	Speedup
Greedy / greedy / greedy	222.33 tok/s	46.93 tok/s	4.74x
Sampled / sampled / sampled	149.69 tok/s	44.33 tok/s	3.38x
Greedy / sampled / greedy	187.46 tok/s	45.43 tok/s	4.13x

Every measured batch produced exactly 288 tokens. The speculative server reported W8, SWA layers: 5/5 (window=2048), and one packed three-lane draft backbone. The AR control used the same binary without a drafter.

Verification

• git range-diff reports both feat(qwen35): support sampled concurrent speculation #654 commits patch-equivalent across the final restack.
• Confirmed that feat(qwen35): support sampled concurrent speculation #654's parent is exactly f1c85e05, the final perf(qwen35): batch draft projections across lanes for concurrency #651 head.
• Built the exact final gfx1201 server and server-unit binary at e0647848.
• Passed the qwen35: DSpark speculative decoding (Qwen3.8-27B drafters) #625 disk-cache version-2 assertion.
• Passed 92 sampler checks covering parity, acceptance probabilities, residual sampling, and EOS-mass redirection.
• Passed all 189 slot-manager checks and all 15 sequence-engine contract checks.
• Passed five host SWA policy cases.
• Passed batched-GDN CPU/GPU checks, all replay-log modes, and the real Qwen3.8 three-lane SWA qualification with max_abs=0 before the final test-only parent update.
• The reference benchmark completed 27 measured and 9 warmup speculative requests plus the same AR-control workload; every request produced exactly 96 tokens.
• Passed git diff --check.

[Graffioh]

Benchmark results after the commit-stack rewrite, at head 0e7da88.

Setup: Radeon gfx1201 32 GB; Qwen3.8-27B-Q4_K_M target; Q8_0 DFlash2 draft; concurrency 3; draft width 8; max context 4096. Each batch launched three concurrent requests and generated 96 tokens per lane with DFLASH_MIN_TOKENS=96. Results are aggregate completion tokens divided by batch wall time, after one warmup and across three measured batches. Every measured batch produced exactly 288 tokens.

Sampled lanes used temperature 0.8, top_p 0.9, top_k 40, frequency_penalty 0.4, and presence_penalty 0.2.

lanes	DFlash2 on #654	one-token AR control	speedup
greedy / greedy / greedy	219.92 tok/s	47.14 tok/s	4.67x
sampled / sampled / sampled	151.29 tok/s	44.62 tok/s	3.39x
greedy / sampled / greedy	188.55 tok/s	45.76 tok/s	4.12x

The control used the same #654 binary with no drafter loaded. It measures concurrent one-token AR and is not a source-level A/B against #651. Server logs confirmed packed DFlash2 batches with lanes=3 and q_len=8 for the speculative runs. The final gfx1201 build passed, as did the five focused CTests.