@novastera-oss/llamarn 0.4.1 → 0.4.3-beta4

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -7,9 +7,10 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #extension GL_KHR_shader_subgroup_shuffle : enable
+#extension GL_KHR_shader_subgroup_vote : enable
 
-#include "types.comp"
-#include "flash_attn_base.comp"
+#include "types.glsl"
+#include "flash_attn_base.glsl"
 
 const uint32_t HSK_per_thread = HSK / D_split;
 const uint32_t HSV_per_thread = HSV / D_split;
@@ -108,6 +109,38 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
+            float max_mask = NEG_FLT_MAX_OVER_2;
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) % Bc;
+                uint32_t r = (idx + tid) / Bc;
+                if (idx + tid < Bc * Br) {
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+                        masksh[c][r] = m;
+                        max_mask = max(max_mask, m);
+                    } else {
+                        masksh[c][r] = float(0);
+                    }
+                }
+            }
+            // skip the block if the mask is entirely -inf
+            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+            barrier();
+            if (gl_SubgroupInvocationID == 0) {
+                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+            }
+            barrier();
+            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                max_mask = max(max_mask, tmpsh[s]);
+            }
+            if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                continue;
+            }
+        }
+
         float Sf[Br][cols_per_thread];
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
@@ -117,6 +150,9 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                continue;
+            }
             [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
@@ -150,16 +186,6 @@ void main() {
         }
 
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
-
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
-                uint32_t c = (idx + tid) % Bc;
-                uint32_t r = (idx + tid) / Bc;
-                if (idx + tid < Bc * Br) {
-                    masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
-                }
-            }
-            barrier();
-
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
                 [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                     float mvf = masksh[c * cols_per_iter + col_tid][r];
@@ -172,8 +198,11 @@ void main() {
 
         float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br];
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            rowmaxf[r] = Sf[r][0];
+            rowmaxf[r] = NEG_FLT_MAX_OVER_2;
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                    continue;
+                }
                 rowmaxf[r] = max(rowmaxf[r], Sf[r][c]);
             }
             Moldf[r] = Mf[r];
@@ -190,6 +219,9 @@ void main() {
             // Compute sum across row of P
             rowsumf[r] = 0.0;
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                    continue;
+                }
                 rowsumf[r] += Pf[r][c];
             }
 
@@ -203,6 +235,9 @@ void main() {
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                continue;
+            }
             [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
@@ -328,12 +363,15 @@ void main() {
 
     float Lfrcp[Br];
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-        Lfrcp[r] = 1.0 / Lf[r];
+        Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
     }
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
+#if defined(ACC_TYPE_MAX)
+            Of[r][d] = clamp(Of[r][d], -vec4(ACC_TYPE_MAX), vec4(ACC_TYPE_MAX));
+#endif
         }
     }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/{flash_attn_base.comp → flash_attn_base.glsl}

@@ -13,6 +13,8 @@ layout (constant_id = 6) const uint32_t D_split = 16;
 const uint32_t HSK_pad = (HSK + 15) & ~15;
 const uint32_t HSV_pad = (HSV + 15) & ~15;
 
+const bool KV_bounds_check = Clamp != 0;
+
 layout (push_constant) uniform parameter {
     uint32_t N;
     uint32_t KV;
@@ -62,33 +64,69 @@ layout (binding = 4) readonly buffer S {float data_s[];};
 
 layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
 
-#if defined(A_TYPE_PACKED16)
 #define BINDING_IDX_K 0
 #define BINDING_IDX_V 1
-layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
+#if defined(DATA_A_F32)
+layout (binding = 1) readonly buffer K_PACKED {vec4 k_data_packed[];} k_packed;
+layout (binding = 2) readonly buffer V_PACKED {vec4 v_data_packed[];} v_packed;
+#elif defined(A_TYPE_PACKED16)
+layout (binding = 1) readonly buffer K_PACKED16 {A_TYPE_PACKED16 k_data_packed16[];} k_packed;
+layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16[];} v_packed;
+#endif
+
+#if defined(DATA_A_F32)
+#undef BLOCK_SIZE
+#define BLOCK_SIZE 4
+#define BLOCK_BYTE_SIZE 16
+
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    // iqs is currently always zero in the flash attention shaders
+    if (binding_idx == BINDING_IDX_K) {
+        return k_packed.k_data_packed[a_offset + ib];
+    } else {
+        return v_packed.v_data_packed[a_offset + ib];
+    }
+}
 #endif
 
 #if defined(DATA_A_Q4_0)
 #define BLOCK_BYTE_SIZE 18
 
 vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
-    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
-    uint shift = (iqs & 0x10) >> 2;
-    vui_lo >>= shift;
-    vui_hi >>= shift;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+    if (binding_idx == BINDING_IDX_K) {
+        uint vui_lo = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+        uint vui_hi = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+        uint shift = (iqs & 0x10) >> 2;
+        vui_lo >>= shift;
+        vui_hi >>= shift;
+
+        return float(k_packed.k_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+    } else {
+        uint vui_lo = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+        uint vui_hi = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+        uint shift = (iqs & 0x10) >> 2;
+        vui_lo >>= shift;
+        vui_hi >>= shift;
+
+        return float(v_packed.v_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+    }
 }
 #endif
 
 #if defined(DATA_A_Q8_0)
 #define BLOCK_BYTE_SIZE 34
 vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
-    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+    if (binding_idx == BINDING_IDX_K) {
+        const i8vec2 v0 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+        const i8vec2 v1 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
 
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+        return float(k_packed.k_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+    } else {
+        const i8vec2 v0 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+        const i8vec2 v1 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+
+        return float(v_packed.v_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+    }
 }
 #endif
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -7,11 +7,12 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_vote : enable
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_cooperative_matrix : enable
 
-#include "types.comp"
-#include "flash_attn_base.comp"
+#include "types.glsl"
+#include "flash_attn_base.glsl"
 
 const uint32_t HSK_per_thread = HSK / D_split;
 const uint32_t HSV_per_thread = HSV / D_split;
@@ -148,18 +149,52 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
+        float mask_cache[Bc * Br / WorkGroupSize];
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
+            float max_mask = NEG_FLT_MAX_OVER_2;
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) % Bc;
+                uint32_t r = (idx + tid) / Bc;
+                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+                        mask_cache[idx / WorkGroupSize] = m;
+                        max_mask = max(max_mask, m);
+                    }
+                }
+            }
+            // skip the block if the mask is entirely -inf
+            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+            barrier();
+            if (gl_SubgroupInvocationID == 0) {
+                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+            }
+            barrier();
+            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                max_mask = max(max_mask, tmpsh[s]);
+            }
+            if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                continue;
+            }
+        }
+
         [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
             uint32_t d = (idx + tid) % (HSK / 4);
             uint32_t c = (idx + tid) / (HSK / 4);
             if (c < Bc && d < HSK / 4) {
+                f16vec4 K_Tf = f16vec4(0);
+                if (!KV_bounds_check || j * Bc + c < KV) {
 #if BLOCK_SIZE > 1
-                uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
-                uint ib = coord / BLOCK_SIZE;
-                uint iqs = (coord % BLOCK_SIZE);
-                f16vec4 K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                    uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
+                    uint ib = coord / BLOCK_SIZE;
+                    uint iqs = (coord % BLOCK_SIZE);
+                    K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
 #else
-                f16vec4 K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
+                    K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
 #endif
+                }
 
                 ksh[c * kshstride + d] = K_Tf;
             }
@@ -198,11 +233,16 @@ void main() {
         }
 
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]));
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                        float f = mask_cache[idx / WorkGroupSize];
+                        sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * f);
+                    }
                 }
             }
             barrier();
@@ -210,8 +250,11 @@ void main() {
 
         float eMf[rows_per_thread];
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            float rowmaxf = sfsh[tile_row(r) + (0 * cols_per_iter + col_tid) * sfshstride];
+            float rowmaxf = NEG_FLT_MAX_OVER_2;
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                    continue;
+                }
                 rowmaxf = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride]));
             }
             float Moldf = Mf[r];
@@ -233,6 +276,9 @@ void main() {
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                continue;
+            }
             float Pf[rows_per_thread];
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
@@ -345,8 +391,8 @@ void main() {
     }
 
     if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            float sink = perElemOpGetSink(tile_row(r), 0u, ACC_TYPE(0), iq2);
 
             float ms = 1.0f;
             float vs = 1.0f;
@@ -367,12 +413,15 @@ void main() {
 
     float Lfrcp[rows_per_thread];
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        Lfrcp[r] = 1.0 / Lf[r];
+        Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
     }
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] *= ACC_TYPE(Lfrcp[r]);
+#if defined(ACC_TYPE_MAX)
+            Of[r][d] = clamp(Of[r][d], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+#endif
         }
     }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -16,9 +16,9 @@
 #extension GL_KHR_shader_subgroup_vote : enable
 #extension GL_EXT_null_initializer : enable
 
-#include "types.comp"
-#include "dequant_funcs_cm2.comp"
-#include "flash_attn_base.comp"
+#include "types.glsl"
+#include "dequant_funcs_cm2.glsl"
+#include "flash_attn_base.glsl"
 
 layout (binding = 0) readonly buffer Q {uint8_t data_q[];};
 layout (binding = 1) readonly buffer K {uint8_t data_k[];};
@@ -29,6 +29,10 @@ ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
     return max(x, y);
 }
 
+float16_t maxReduceFp16(const in float16_t x, const in float16_t y) {
+    return max(x, y);
+}
+
 ACC_TYPE smearReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
     return x;
 }
@@ -121,7 +125,11 @@ void main() {
     const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
 
     L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
+#if defined(ACC_TYPE_MAX)
+    M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(-ACC_TYPE_MAX / ACC_TYPE(2));
+#else
     M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(NEG_FLT_MAX_OVER_2);
+#endif
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0);
 
@@ -138,6 +146,44 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
+            if (nem1_bounds_check) {
+                tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+                tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
+                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
+                tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
+
+                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv, mvmax;
+
+                coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+
+                // skip the block if the mask is entirely -inf
+                coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
+                if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
+            } else {
+                tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
+                // Don't clamp against nem1 when GQA is enabled
+                uint32_t m_height = p.gqa_ratio > 1 ? ~0 : p.nem1;
+                tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, m_height, KV);
+                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
+
+                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
+
+                coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+
+                // skip the block if the mask is entirely -inf
+                coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
+                if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
+            }
+        }
+
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
         coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
@@ -154,14 +200,6 @@ void main() {
         }
 
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
-            tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
-            tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
-            tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
-
-            coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
-
-            coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
-
             S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
         }
 
@@ -278,11 +316,15 @@ void main() {
 
     [[unroll]]
     for (int k = 0; k < Ldiag.length(); ++k) {
-        Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];
+        Ldiag[k] = (Ldiag[k] == 0.0) ? ACC_TYPE(0.0) : (ACC_TYPE(1.0) / Ldiag[k]);
     }
 
     O = Ldiag*O;
 
+#if defined(ACC_TYPE_MAX)
+    [[unroll]] for (uint i = 0; i < O.length(); ++i) { O[i] = clamp(O[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
+#endif
+
     uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp

@@ -91,7 +91,7 @@ void main() {
         L = L*ms + vs;
     }
 
-    L = 1.0 / L;
+    L = (L == 0.0) ? 0.0 : 1.0 / L;
 
     // D dimension is split across workgroups in the y dimension
     uint d = tid + gl_WorkGroupID.y * BLOCK_SIZE;
@@ -111,6 +111,10 @@ void main() {
             }
         }
         O *= L;
+
+        const float FLT_MAX = uintBitsToFloat(0x7F7FFFFF);
+        O = clamp(O, -FLT_MAX, FLT_MAX);
+
         data_d[iq3 * D * N + D * n + d] = O;
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(floor(x));
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "glu_head.comp"
+#include "glu_head.glsl"
 
 const float GELU_COEF_A    = 0.044715f;
 const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
@@ -10,4 +10,4 @@ float op(float a, float b) {
     return 0.5f*a*(2.0f - 2.0f / (exp(2 * val) + 1)) * b;
 }
 
-#include "glu_main.comp"
+#include "glu_main.glsl"

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "glu_head.comp"
+#include "glu_head.glsl"
 
 // based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
 // ref: https://www.johndcook.com/blog/python_erf/
@@ -24,4 +24,4 @@ float op(float a, float b) {
     return 0.5f * a * (1.0f + erf_approx) * b;
 }
 
-#include "glu_main.comp"
+#include "glu_main.glsl"

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "glu_head.comp"
+#include "glu_head.glsl"
 
 const float GELU_QUICK_COEF = -1.702f;
 
@@ -8,4 +8,4 @@ float op(float a, float b) {
     return a * (1.0f / (1.0f + exp(GELU_QUICK_COEF * a))) * b;
 }
 
-#include "glu_main.comp"
+#include "glu_main.glsl"

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/{generic_binary_head.comp → generic_binary_head.glsl}

@@ -1,8 +1,11 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_control_flow_attributes : require
 
-#include "rte.comp"
-#include "utils.comp"
+#include "rte.glsl"
+#include "utils.glsl"
+#if RMS_NORM_ROPE_FUSION
+#include "rope_params.glsl"
+#endif
 
 layout (push_constant) uniform parameter
 {
@@ -12,11 +15,16 @@ layout (push_constant) uniform parameter
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint misalign_offsets;
     float param1; float param2; int param3;
+#if RMS_NORM_ROPE_FUSION
+    rope_params rope;
+#endif
 } p;
 
+#if !RMS_NORM_ROPE_FUSION
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#endif
 
 // true if src0/src1 are the same shape and the indices can be reused without additional modulus
 layout(constant_id = 0) const bool norepeat = false;