[BugFix] Fix race in c128 prefill plan kernel on ragged extend (#32467)

Signed-off-by: EanWang211123 <wangyiheng@sangfor.com.cn>
This commit is contained in:
yiheng
2026-08-13 00:15:41 +08:00
committed by GitHub
parent d21eefc94f
commit 8549cce11b
@@ -154,11 +154,6 @@ __global__ __launch_bounds__(1024, 1) //
counter_c = 0;
counter_w = 0;
}
if (tx < kNumWarps) {
warp_max[tx] = 0;
warp_min[tx] = 0xFFFFFFFFu;
}
// === Stage B: min/max(extend_len) for MTP-uniform detection ===
// For min, treat threads outside `batch_size` as +inf so they don't pull the min down.
const uint32_t e_for_max = static_cast<uint32_t>(extend_len);