[Bug] Default value of ChunkQ in deepgemm could lead to divided_by_0 error

[qli88]

GLM 5.1 FP8 fails on MI300/355 because its index_n_heads=32 and ChunkQ = 64, which will lead to divided by zero error when calculation triton grid.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:sglang	SGLang integration tests
ci:atom	ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm	vLLM benchmark
ci:all	All of the above

Add labels via the sidebar or gh pr edit 2891 --add-label <label>

[Copilot]

Pull request overview

This PR fixes a divide-by-zero failure in the DeepGEMM FP8 paged MQA logits Triton wrapper that can occur when the default ChunkQ is larger than heads, causing TileQCount to become 0 and breaking the Triton grid/SplitKV computation.

Changes:

• Clamp ChunkQ to heads in deepgemm_fp8_paged_mqa_logits_stage1 to prevent TileQCount == 0.
• Prevent downstream TotalCuCount // TileQCount from raising a ZeroDivisionError in the reported configuration (e.g., heads=32, ChunkQ=64).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.