[Performance] GLM-OCR MTP: ~1.5x speedup on eager path but ~0.62x on default CUDA Graph path (OCRBench)

[airlsyn]

For zai-org/GLM-OCR with native MTP speculative decoding on vLLM, we consistently observe:

tested on OCRBench (1000 samples) with 80GB A100

mtp=1 + eager ≈ 1.5× faster than no_mtp + eager — matches the GLM-OCR tech report's ~50% throughput claim.
mtp=1 + cuda graph ≈ 0.62× slower than no_mtp + graph on the default CUDA Graph path — despite high draft acceptance

Could you help with how to get MTP speedup on the CUDA Graph path?

---

And we see a issue on vllm #37551 test, with same result

Case	no_mtp_eager	mtp_eager	eager speedup	no_mtp_graph	mtp_graph	graph ratio	MTP acceptance
receipt	60.08 tok/s	91.79 tok/s	1.53×	465.31 tok/s	260.37 tok/s	0.56×	~69%, mean accept len ~1.69
ocr_demo	59.48 tok/s	90.69 tok/s	1.53×	397.39 tok/s	277.29 tok/s	0.70×	~81%, mean accept len ~1.81
mean	—	—	1.526×	—	—	0.623×	—