[GLM-5.3 Flash] Serving optimizations: KDA/mHC fusion, DSA draft metadata graphs, prefill autotuning

[JustinTong0323]

Eleven validated serving optimizations for GLM-5.3-Flash, from a 4xGB300 benchmark campaign against vLLM (fixed workload: 80 random-token requests, chat 1000/1000 and summarization 8000/1000 at concurrency 1/16/80, natural NEXTN 5-1-6, TP4/EP1, FP8 KV).

Measured end state of the full campaign stack vs pinned vLLM:

Cell	Original gap	Final gap
chat C1	-8.6%	+22.3%
chat C16	-17.2%	+14.4%
chat C80	-52.9%	+15.2%
summarization C1	-10.8%	+20.8%
summarization C16	-20.9%	+6.2%
summarization C80	-45.2%	-0.1% (tie within 1%)

Commits, grouped:

• KDA kernels: dense-verification register-spill and state-tile tuning; CPU logical-token-count reuse for ordinary prefill
• mHC: small-token post parallelism, small-row gated RMSNorm tile, cross-layer post/pre fusion
• DSA draft metadata: pooled preparation capture, qualified graph reuse by default, padded-batch extension
• Prefill autotuning for GLM FP8 experts
• State fix: multi-CTAs KV counter lifetime for captured decode graphs (carried from [Fix] Don't free the multi-CTAs KV counter the decode graphs captured #39175, closed unmerged; its KDA-verify sibling already landed as [Fix] Don't write conv state from the fused KDA verify kernel #39219)

Validation so far: campaign-wide GSM8K runs at 90.8-92.7% with all-natural stops on every admitted step; pre-commit clean (GITHUB_BASE_REF=upstream/main); 63 focused tests on this series. A same-box attribution A/B confirms this series plus the deferred projection/mHC work recovers to within 0.5% of the campaign numbers.

Gate status: a full C80 re-gate of this exact series is pending reconciliation of the deferred items below; the PR stays draft until the gate passes.

Relationship to other PRs

Optimization area	Status here	Related PR
KDA projection fusion (qkvbfg)	NOT carried — deferred to the open PRs below; a validated unified implementation exists locally if needed	#39350 (open, attention-TP-aware design), #38254 (open, testless earlier design)
mHC attn→MLP boundary	NOT carried — deferred	#39200 (open)
CPU logical-token-count reuse	carried (36e3696799); overlapping intent, implementations differ — will coordinate/merge whichever lands second	#38431 (open, reuses host lengths to avoid KDA prefill sync)
KDA dense verify / state tiles	carried; functionally distinct from the dispatch rework	#37901 (open, hardware/shape-aware KDA dispatch)
KDA prefill OOM workspace guard	not carried, unrelated to this series	#37535 (open)
KPool metadata fusion	adopted as-is during the campaign, no local variant	#38845 (merged)
KDA verify conv-state fix	in base via upstream	#39219 (merged)
Cookbook recipe update (fixed MTP 5/1/6, EP1, TRTLLM MoE)	merged separately	#39213 (merged)

Co-authored-by trailers mark the commits incorporating @mmangkad's glm-5.3-flash-opt work; the carried state fix keeps his original authorship.

---

CI States

Latest PR Test (Base): ❌ Run #34851287867
Latest PR Test (Extra): ❌ Run #34851287364
Latest PR Test (AMD ROCm 10): ❌ Run #34851287532