vulkan: Use unclamped loads for flash attention mask

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1787,6 +1787,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
             // can't use 256 for D==80.
             uint32_t wg_size = (small_rows && (D % 32) == 0) ? 256 : 128;
             auto rows_cols = fa_rows_cols(D, clamp, type, small_rows);
+            // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
+            GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0);
             return {wg_size, rows_cols[0], rows_cols[1], (D), clamp};
         };
 
@@ -5464,6 +5466,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                    // the "aligned" shader variant will forcibly align strides, for performance
                    (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0;
 
+    // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
+    GGML_ASSERT((nem1 % GGML_KQ_MASK_PAD) == 0);
+
     vk_pipeline pipeline = pipelines[aligned];
     assert(pipeline);
 

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -256,7 +256,7 @@ void main() {
         }
 
         if (p.mask != 0) {
-            tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+            tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
             tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
             // When using grouped query attention, all rows use the same mask.
             if (p.gqa_ratio > 1) {