perf(glm52): batch TP4 prefill execution

[xiaguan]

Summary

Speed up the native GLM-5.2 FP8 TP4 prefill-only request path by batching routed-expert work and increasing the outer prefill tile from 32 to 512 rows.

The previous implementation launched the FP8 projection chain separately for every active expert. On a 1K prompt this made prefill roughly 24 seconds and dominated the entire request. This change:

• packs routed rows once and executes the routed experts through two batched FlashInfer/CUTLASS groupwise FP8 GEMMs per MoE layer;
• keeps the indexer and sparse-attention subtiles at their validated 32-row shape while dense and MoE projections use a 512-row outer tile;
• uses NCCL BF16 all-reduce for the large-M prefill path and avoids allocating decode-only LL/VMM buffers in prefill-only mode;
• reduces the packed routed/shared outputs with one deterministic weighted-reduction kernel;
• adds a 512-row startup preflight and accounts for the larger workspace in the VRAM budget.

This PR deliberately does not change startup loading. The remaining large performance gap is the padded 512-row MoE execution itself; closing it requires a deployable fused-MoE backend that can consume most or all of the 16K prefill chunk directly.

Performance

Single request, greedy one-token boundary output, concurrency 1, prefix-cache misses, 16K chunking, one 4×GB300 host. Values are mean TTFT.

requested input	merged implementation	this PR	vLLM TP4+EP FP8-KV
1K	24,451 ms	1,843 ms	319 ms
4K	—	5,310 ms	310 ms
16K	—	20,138 ms	901 ms

The 1K path is 13.3× faster than the merged implementation. Moving from the validated 256-row point to 512 rows improved TTFT by another 4.9% at 1K, 5.8% at 4K, and 3.5% at 16K. A 1024-row experiment faulted and is not used.

The final 512-row startup preflight passes on all four ranks. Post-build free VRAM is 84.0 GiB per rank with --max-model-len 131136 and --glm52-prefill-chunk-size 16384.

Nsight Systems attributes the remaining request GPU time primarily to FP8 CUTLASS GEMM (65.4%); NCCL is 9.0% and sparse attention is 1.0%. The new weighted reduction is about 0.3% of the request trace. A temporary 256×9-route microbenchmark measured 12.1 µs per launch; Nsight Compute measured 85.3% achieved occupancy. The temporary benchmark was not committed.

Validation

• cargo build --release -p openinfer-server --features glm52
• cargo test -p openinfer-glm52 --lib — 78 passed, 17 ignored
• cargo fmt --all -- --check
• cargo clippy -p openinfer-kernels --features glm52 --lib -- -D warnings
• live TP4 startup with the 512-row kernel preflight
• five 1K requests and three requests each at 4K and 16K
• git diff --check

Affected-crate strict Clippy still reports existing lints outside this diff in the GLM-5.2 crate (remote match arms, scheduler boolean parameters, and an existing topology match).