Warm 512-wide DeepSeek sparse-MLA kernels before serving

[Aaronontheweb]

Summary

• warm the observed 512-wide portable sparse-MLA Triton layout during startup
• retain the full 576-wide NoPE plus RoPE warmup
• cover both 32- and 64-head TP-local layouts and update locked source provenance

Evidence

A live 16K prefill compiled _accumulate_indexed_attention_chunk_multihead_kernel after startup with:

BLOCK_D=512
head_dim=512
num_heads=32
stride_q_t=16384
stride_q_h=512

The prior image only warmed 576-wide Q/K tensors. Triton treats head_dim and the contiguous strides as compile keys.

Validation

• ./scripts/validate-repository.sh
• Python syntax compilation for the changed warmup and test modules
• cumulative patch applies cleanly to pinned vLLM commit 568afb3a13806beb53bb2e6bd518269357b237c0
• pinned upstream source archive checksum matches dependency.lock.json