[zijiecode]

Motivation

MiniMax-M3 fails to start with EAGLE3 on ROCm because target verification expects missing prefill metadata. This PR adds the missing verify path.

Modifications

• Reuse sparse decode for linear draft chains (--speculative-eagle-topk 1), with the correct causal KV length for each token.
• Keep dense-layer metadata unchanged and size the shared top-k buffer for all verification tokens.
• Add regression tests for verification metadata and CUDA Graph replay.

Accuracy Tests

amd/MiniMax-M3-MXFP4, full 5-shot GSM8K, chat-completions, 16384-token output budget, temperature 0, and real EAGLE3 verification. Results are as expected.

Configuration	Strict-match	Flexible-extract
Baseline	96.59%	96.51%
This PR	96.74%	96.66%

Benchmarking and Profiling

AgentX on MI355X, TP4/EP1, C1 and C32, no KV offload, 3600 seconds per run, the baseline uses standard decoding and the fixed run enables EAGLE3.

C1:

Metric	Baseline	This PR	Change
Standard p90 interactivity (tok/s/user)	124.47	249.88	+100.76%
Total throughput (tok/s/GPU)	2916.94	4366.28	+49.69%
Output throughput (tok/s/GPU)	23.55	32.75	+39.05%
TPOT mean (ms/token)	7.215	3.694	-48.81%
TPOT p50 (ms/token)	7.185	3.744	-47.89%

C32:

Metric	Baseline	This PR	Change
Standard p90 interactivity (tok/s/user)	25.04	28.99	+15.78%
Total throughput (tok/s/GPU)	25174.33	29012.44	+15.25%
Output throughput (tok/s/GPU)	175.65	199.80	+13.75%
TPOT mean (ms/token)	27.180	22.478	-17.30%
TPOT p50 (ms/token)	24.007	16.759	-30.19%

Checklist

• Format your code according to the Format code with pre-commit guide.
• Add unit tests.
• Update documentation.
• Provide accuracy and speed benchmark results.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ✅ Run #36686661208
Latest PR Test (Extra): ❌ Run #36686660506
Latest PR Test (AMD ROCm 10): ⏳ Run #36686661067

[zijiecode]

/rerun-failed-ci