fix: harden attention kernel boundaries

- fix scalar prefill head_dim=32 out-of-bounds via G=4 dispatch - fix MMA decode 4D mask head indexing and invalid-row mask access - add q_head/kv_head divisibility and head-dim contiguity checks - validate split-KV scratch and decode out_buf layout in bindings - set max dynamic shared memory for scalar decode D=256 - cover scalar prefill D=32 in pure C test
2026-08-09 14:53:24 +08:00
parent 596c35fd71
commit d565d44c43
8 changed files with 50 additions and 19 deletions
@@ -323,6 +323,7 @@ int main() {
    // ---- PREFILL ----
    {
        const int configs[][7] = {
+            {1,2,1,64,128,32,0},     // scalar fallback D=32
            {1,2,1,64,128,64,0},     // tiny: B,Hq,Hk,q,kv,D,causal
            {1,32,4,512,512,128,0},  // standard
            {1,32,4,128,256,128,0},  // medium