[namgyu-yun]

Overview

Adds a measured B200 (SM100) table for the GLM-5.2/5.3 low-latency BF16 decode GEMM; SM100 used cuBLAS. Related: #59822.

Claims

• 1.14–2.35x over cuBLAS per projection at M ≤ 8.

Validation

pytest tests/kernels/test_bf16_skinny_gemm.py -k "glm or dsv3"   → 84 passed, 180 skipped (B200)
run_glm52_plan vs F.linear, every routed cell (CUPTI, CUDA graph) → 40 cells, all ≥ 1.05x cold and hot L2

projection	(N, K)	routed M	M=1 cuBLAS → table (µs, cold L2)	speedup over routed M
q_a + kv_a	(2624, 6144)	1–4, 6, 8	12.00 → 8.48	1.21–1.42x
q_b TP8	(2048, 2048)	1–4, 6, 8	6.02 → 4.06	1.46–2.00x
q_b TP4 / indexer wq_b	(4096, 2048)	1–4	6.94 → 5.76	1.15–1.29x
o_proj TP8	(6144, 2048)	1–2	8.03 → 6.91	1.14–1.16x
indexer wk + weights	(160, 6144)	1–4, 6, 8	6.08 → 2.62	1.48–2.35x
shared gate_up TP8	(512, 6144)	1–4, 6, 8	8.00 → 4.26	1.37–1.88x
shared gate_up TP4	(1024, 6144)	1–4, 6, 8	9.09 → 5.22	1.16–1.74x
MTP eh_proj	(6144, 12288)	1–4	37.63 → 26.24	1.31–1.43x

Details

Cells kept only if ≥ 5 % faster on cold and hot L2; others stay on cuBLAS. No e2e (needs 4–8 B200).

---

Pull Request Checklist

• Used /pr-checklist.

• AI-assisted.

• Design Fit:

• Testing and Validation:

• Code Quality and Style:

• Pull Request Contents:

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[namgyu-yun]

@gau-nernst could you check this PR as well? it's same SM100 change as #64482, for GLM model

[github-actions]

✅ @namgyu-youn, CI is now available for this PR.

• /ci run starts upstream CI; /amd-ci run starts AMD CI only.
• Run commands test your branch with every commit currently on its upstream target branch. If it is behind, the bot merges the target branch into it first (pull before pushing again). Append --allow-stale to a run command to test an outdated branch as is, at your own risk.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
• /ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

[namgyu-yun]

/ci run

[github-actions]

✅ Triggered Buildkite CI #98250 for commit 6364150e39c9.

ℹ️ This PR is 2 commits behind upstream main. Merged upstream main into this branch as 6364150e39c9 before starting CI. Pull the branch before pushing more changes.