[festeraeb]

Fixes #839.

The stage-1 decode kernel keeps the Q tile and one K and one V tile in shared memory. The launch code budgeted only K and V and never went below 32 KV rows, so head_dim 512 overflows a 64 KiB part. This counts the Q tile and halves BLOCK_N, down to tl.dot's 16-row minimum, until the three tiles fit (16 + 16 + 16 KiB for head_dim 512). GPUs whose budget already fits (all sm_80+) take the same tiles as before.

Tested on an RTX 2060 (sm_75), driver 580.178, CUDA 13.3, main @ a2ce636 built from source with sglang-kernel uninstalled, checkpoint and command as in #839:

build	result
main	first decode: OutOfResources ... Required: 81920, Hardware limit: 65536
this branch	decode passes; the next failure is the extend tile (Required: 98304), which #24 fixes
this branch + #24's _select_extend_tile hunk	serves: 53.5 / 65.8 tok/s decode on 44- / 400-token answers, TTFT 1.7 s warm

🤖 Generated with Claude Code