whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -55,7 +55,7 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
     return max(elem0, elem1);
 }
 
-#if defined(BLOCK_SIZE)
+#if BLOCK_SIZE > 1
 #define DECODEFUNC , DEQUANTFUNC
 #else
 #define DECODEFUNC
@@ -72,6 +72,28 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
+// Store O values for non-GQA split_k. Rows are tokens, not heads.
+D_TYPE perElemOpNonGqaSplitKStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t unused, const in uint32_t iq2, const in uint32_t N) {
+    uint32_t global_row = i * Br + r;
+    if (global_row < N && c < HSV) {
+        uint32_t o_off = HSV * p.ne1
+            * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+        data_o[o_off + iq2 * HSV + c] = D_TYPE(elem);
+    }
+    return elem;
+}
+
+// Store L/M values for non-GQA split_k.
+ACC_TYPE perElemOpNonGqaSplitKStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t lm_base, const in uint32_t iq2, const in uint32_t N) {
+    uint32_t global_row = i * Br + r;
+    if (global_row < N && c == 0) {
+        uint32_t lm_off = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3
+            + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+        data_o[lm_off + lm_base + iq2] = D_TYPE(elem);
+    }
+    return elem;
+}
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
@@ -85,7 +107,7 @@ void main() {
 
     tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0);
 
-#if defined(BLOCK_SIZE)
+#if BLOCK_SIZE > 1
     tensorLayoutK = setTensorLayoutBlockSizeNV(tensorLayoutK, 1, BLOCK_SIZE);
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
@@ -98,7 +120,7 @@ void main() {
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
     {
         q_stride &= ~7;
-#if !defined(BLOCK_SIZE)
+#if BLOCK_SIZE == 1
         k_stride &= ~7;
         v_stride &= ~7;
 #endif
@@ -111,13 +133,13 @@ void main() {
     coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
     coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
 
-    uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
+    uint32_t q_offset = gqa_iq1*p.nb01*4/*sizeof(float)*/ + iq2*p.nb02+iq3*p.nb03;
     coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
     Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -138,48 +160,67 @@ void main() {
         coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
     }
 
-    uint32_t m_offset = 0;
+    const uint32_t mo_stride = CEIL_DIV(KV, 16 * Bc);
+    // mo_offset will point to the tile starting at row i*Br and col 0
+    uint32_t mo_offset = mo_stride * i;
+
+    uint32_t m_offset = gqa_iq1*KV * 2 /*sizeof(float16_t)*/;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+        mo_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * CEIL_DIV(p.nem1, Br) * mo_stride;
     }
 
+    uint32_t mask_opt = 0;
+    uint32_t mask_opt_idx = ~0;
+
     [[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> 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));
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv = coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
+        if (MASK_ENABLE) {
 
-                // skip the block if the mask is entirely -inf
-                coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
-                if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
-                    continue;
+            if (USE_MASK_OPT && mask_opt_idx != j / 16) {
+                mask_opt_idx = j / 16;
+                mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
+            }
+            uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
+            if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
+                // skip this block
+                continue;
+            }
+            // Only load if the block is not all zeros
+            if (mask_opt_bits != MASK_OPT_ALL_ZERO) {
+                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> 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;
+                    }
                 }
             }
         }
@@ -192,14 +233,14 @@ void main() {
         coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK_pad), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
-        if (p.logit_softcap != 0.0f) {
+        if (LOGIT_SOFTCAP) {
             [[unroll]]
             for (int k = 0; k < S.length(); ++k) {
                 S[k] = ACC_TYPE(p.logit_softcap)*tanh(S[k]);
             }
         }
 
-        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+        if (MASK_ENABLE) {
             S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
         }
 
@@ -218,6 +259,8 @@ void main() {
 
         coopMatReduceNV(rowmax, S, gl_CooperativeMatrixReduceRowNV, maxReduce);
 
+        rowmax += coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(FATTN_KQ_MAX_OFFSET);
+
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> Mold = M;
 
         // M = max(rowmax, Mold)
@@ -260,11 +303,8 @@ void main() {
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
-        // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
-        PV = coopMatMulAdd(P_A, V, PV);
-
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
+        O *= coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(eMdiag);
+        O = coopMatMulAdd(P_A, V, O);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
@@ -272,12 +312,19 @@ void main() {
     if (p.k_num > 1) {
         coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
-        coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
-
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
-        coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
-        coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
+
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
+            coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
+        } else {
+            coopMatPerElementNV(O_D, O_D, perElemOpNonGqaSplitKStore, 0u, iq2, N);
+            coopMatPerElementNV(L, L, perElemOpNonGqaSplitKStoreCol0, 0u, iq2, N);
+            coopMatPerElementNV(M, M, perElemOpNonGqaSplitKStoreCol0, p.ne1, iq2, N);
+        }
         return;
     }
 
@@ -305,7 +352,7 @@ void main() {
             if (sink > Mr[i]) {
                 ms = exp(Mr[i] - sink);
 
-                O[i] *= ms;
+                O[i] *= float16_t(ms);
             } else {
                 vs = exp(sink - Mr[i]);
             }
@@ -319,15 +366,16 @@ void main() {
         Ldiag[k] = (Ldiag[k] == 0.0) ? ACC_TYPE(0.0) : (ACC_TYPE(1.0) / Ldiag[k]);
     }
 
-    O = Ldiag*O;
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
+
+    O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(Ldiag)*O_D;
 
 #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); }
+    [[unroll]] for (uint i = 0; i < O_D.length(); ++i) { O_D[i] = clamp(O_D[i], D_TYPE(-ACC_TYPE_MAX), D_TYPE(ACC_TYPE_MAX)); }
 #endif
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + 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);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp

@@ -0,0 +1,162 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 128;
+layout (constant_id = 1) const uint NUM_SUBGROUPS = 4;
+layout (constant_id = 2) const uint Br = 32;
+layout (constant_id = 3) const uint Bc = 32;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {float16_t data_a[];};
+layout (binding = 0) readonly buffer Av4 {f16vec4 data_av4[];};
+layout (binding = 1) writeonly buffer D {uint data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint nem0;
+    uint nem1;
+    uint nem2;
+    uint nbm1;
+    uint nbm2;
+    uint nbm3;
+    uint nbd1;
+    uint nbd2;
+    uint nbd3;
+};
+
+#define MASK_OPT_ALL_NEG_INF 1
+#define MASK_OPT_ALL_ZERO 2
+
+shared float minsh[NUM_SUBGROUPS];
+shared float maxsh[NUM_SUBGROUPS];
+
+float FLT_MAX_OVER_2 = uintBitsToFloat(0x7EFFFFFF);
+
+void loadvec4(inout uint result, const uint i0, const uint i1, const uint i2, const uint i3, const bool need_bounds_check) {
+    const uint tid = gl_LocalInvocationIndex;
+
+    [[unroll]] for (uint block_x = 0; block_x < 16; ++block_x) {
+        float min_v = FLT_MAX_OVER_2;
+        float max_v = -FLT_MAX_OVER_2;
+        [[unroll]] for (uint i = 0; i < Br * Bc / 4; i += BLOCK_SIZE) {
+            uint j0 = (i + tid) % (Bc / 4);
+            uint j1 = (i + tid) / (Bc / 4);
+
+            j0 *= 4;
+            j0 += (i0 * 16 + block_x) * Bc;
+            j1 += i1 * Br;
+
+            if (!need_bounds_check || j0 + 3 < nem0) {
+                vec4 f = vec4(data_av4[(j0 + j1 * nbm1 + i2 * nbm2 + i3 * nbm3) / 4]);
+                [[unroll]] for (int c = 0; c < 4; ++c) {
+                    min_v = min(min_v, f[c]);
+                    max_v = max(max_v, f[c]);
+                }
+            } else {
+                [[unroll]] for (int c = 0; c < 4; ++c) {
+                    if (j0 + c < nem0) {
+                        float f = float(data_a[j0 + j1 * nbm1 + i2 * nbm2 + i3 * nbm3]);
+                        min_v = min(min_v, f);
+                        max_v = max(max_v, f);
+                    }
+                }
+            }
+        }
+        min_v = subgroupMin(min_v);
+        max_v = subgroupMax(max_v);
+        if (gl_SubgroupInvocationID == 0) {
+            minsh[gl_SubgroupID] = min_v;
+            maxsh[gl_SubgroupID] = max_v;
+        }
+        barrier();
+        if (tid == 0) {
+            [[unroll]] for (uint i = 0; i < NUM_SUBGROUPS; ++i) {
+                min_v = min(min_v, minsh[i]);
+                max_v = max(max_v, maxsh[i]);
+            }
+            if (max_v <= -FLT_MAX_OVER_2) {
+                result |= 1 << (2*block_x);
+            }
+            if (min_v == 0.0f && max_v == 0.0f) {
+                result |= 2 << (2*block_x);
+            }
+        }
+        barrier();
+    }
+}
+
+// For each Br x Bc block of the mask (input) buffer, read all values and check
+// if it's all -inf or all zero. Write out a two-bit code indicating which it is
+// (or zero for neither). Each workgroup processes 16 tiles and writes out a
+// 32-bit result mask.
+//
+// TODO: This is a lot of work per workgroup, might make sense to split this into
+// more workgroups in the future.
+void main() {
+    // Each workgroup handles a row
+    const uint tid = gl_LocalInvocationIndex;
+    const uint i0 = gl_WorkGroupID.x;
+    const uint i1 = gl_WorkGroupID.y;
+    const uint i2 = gl_WorkGroupID.z % nem2;
+    const uint i3 = gl_WorkGroupID.z / nem2;
+
+    uint result = 0;
+
+    // Fast path for fully in-bounds blocks where we can do f16vec4 loads
+    if ((nem0 % Bc) == 0 && (nem1 % Br) == 0 &&
+        ((Br * Bc) % (BLOCK_SIZE * 4)) == 0) {
+        if ((i0 + 1) * 16 * Bc <= nem0) {
+            loadvec4(result, i0, i1, i2, i3, false);
+        } else {
+            loadvec4(result, i0, i1, i2, i3, true);
+        }
+    } else {
+        [[unroll]] for (uint block_x = 0; block_x < 16; ++block_x) {
+            float min_v = FLT_MAX_OVER_2;
+            float max_v = -FLT_MAX_OVER_2;
+            [[unroll]] for (uint i = 0; i < Br * Bc; i += BLOCK_SIZE) {
+                if ((Br * Bc % BLOCK_SIZE) != 0 && i + tid >= Br * Bc) {
+                    continue;
+                }
+                uint j0 = (i + tid) % Bc;
+                uint j1 = (i + tid) / Bc;
+
+                j0 += (i0 * 16 + block_x) * Bc;
+                j1 += i1 * Br;
+
+                if (j0 < nem0 && j1 < nem1) {
+                    float f = float(data_a[j0 + j1 * nbm1 + i2 * nbm2 + i3 * nbm3]);
+                    min_v = min(min_v, f);
+                    max_v = max(max_v, f);
+                }
+            }
+            min_v = subgroupMin(min_v);
+            max_v = subgroupMax(max_v);
+            if (gl_SubgroupInvocationID == 0) {
+                minsh[gl_SubgroupID] = min_v;
+                maxsh[gl_SubgroupID] = max_v;
+            }
+            barrier();
+            if (tid == 0) {
+                [[unroll]] for (uint i = 0; i < NUM_SUBGROUPS; ++i) {
+                    min_v = min(min_v, minsh[i]);
+                    max_v = max(max_v, maxsh[i]);
+                }
+                if (max_v <= -FLT_MAX_OVER_2) {
+                    result |= 1 << (2*block_x);
+                }
+                if (min_v == 0.0f && max_v == 0.0f) {
+                    result |= 2 << (2*block_x);
+                }
+            }
+            barrier();
+        }
+    }
+
+    if (tid == 0) {
+        data_d[i0 + i1 * nbd1 + i2 * nbd2 + i3 * nbd3] = result;
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp

@@ -12,7 +12,8 @@ layout (binding = 2) writeonly buffer D {float data_d[];};
 
 layout (push_constant) uniform parameter {
     uint D;
-    uint N;
+    uint ne1;
+    uint ne2;
     uint ne3;
     uint k_num;
     uint sinks;
@@ -24,15 +25,15 @@ void main() {
     // Each workgroup handles a row
     const uint n = gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
-    const uint iq3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.z % p.ne2;
+    const uint i3 = gl_WorkGroupID.z / p.ne2;
 
     uint D = p.D;
-    uint N = p.N;
     uint k_num = p.k_num;
 
-    uint l_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + n;
-    uint m_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + N + n;
-    uint lm_stride = N * 2;
+    uint l_offset = D * p.ne1 * p.ne2 * p.ne3 * k_num + p.ne1 * 2 * (0/*split_k_index*/ + p.k_num * (i2 + p.ne2 * i3)) + n;
+    uint m_offset = D * p.ne1 * p.ne2 * p.ne3 * k_num + p.ne1 * 2 * (0/*split_k_index*/ + p.k_num * (i2 + p.ne2 * i3)) + p.ne1 + n;
+    uint lm_stride = p.ne1 * 2;
 
     // Compute the max m value for the row
     float m_max = -1.0/0.0;
@@ -99,7 +100,7 @@ void main() {
     if (d < D) {
         float O = 0.0;
         [[unroll]] for (uint k = 0; k < k_num; ++k) {
-            uint o_offset = D * N * (k + iq3 * k_num) + D * n + d;
+            uint o_offset = D * p.ne1 * (k + p.k_num * (i2 + p.ne2 * i3)) + D * n + d;
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
@@ -115,6 +116,6 @@ void main() {
         const float FLT_MAX = uintBitsToFloat(0x7F7FFFFF);
         O = clamp(O, -FLT_MAX, FLT_MAX);
 
-        data_d[iq3 * D * N + D * n + d] = O;
+        data_d[(i3 * p.ne2 + i2) * p.ne1 * D + D * n + d] = O;
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp

@@ -0,0 +1,128 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+layout(constant_id = 0) const uint S_V = 128;
+layout(constant_id = 1) const uint KDA = 0;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform Parameters {
+    uint H;
+    uint n_tokens;
+    uint n_seqs;
+    uint s_off;
+    uint sq1, sq2, sq3;
+    uint sv1, sv2, sv3;
+    uint sb1, sb2, sb3;
+    uint neq1, rq3;
+    float scale;
+};
+
+layout(binding = 0) readonly  buffer QBuf     { FLOAT_TYPE data_q[];     };
+layout(binding = 1) readonly  buffer KBuf     { FLOAT_TYPE data_k[];     };
+layout(binding = 2) readonly  buffer VBuf     { FLOAT_TYPE data_v[];     };
+layout(binding = 3) readonly  buffer GBuf     { FLOAT_TYPE data_g[];     };
+layout(binding = 4) readonly  buffer BetaBuf  { FLOAT_TYPE data_beta[];  };
+layout(binding = 5) readonly  buffer StateBuf { FLOAT_TYPE data_state[]; };
+layout(binding = 6)           buffer DstBuf   { FLOAT_TYPE data_dst[];   };
+
+shared FLOAT_TYPE s_k[S_V];
+shared FLOAT_TYPE s_q[S_V];
+shared FLOAT_TYPE s_g[S_V]; // KDA only: cached exp(g[i])
+
+void main() {
+    const uint head_id = gl_WorkGroupID.x;
+    const uint seq_id  = gl_WorkGroupID.y;
+    const uint col     = gl_LocalInvocationID.x;
+
+    const uint iq1 = head_id % neq1;
+    const uint iq3 = seq_id / rq3;
+
+    const uint state_size = S_V * S_V;
+    const uint state_base = (seq_id * H + head_id) * state_size;
+
+    FLOAT_TYPE state[S_V];
+    [[unroll]] for (uint i = 0; i < S_V; i++) {
+        state[i] = FLOAT_TYPE(data_state[state_base + col * S_V + i]);
+    }
+
+    uint attn_off = (seq_id * n_tokens * H + head_id) * S_V;
+
+    for (uint t = 0; t < n_tokens; t++) {
+        const uint q_off = iq3 * sq3 + t * sq2 + iq1 * sq1;
+        const uint k_off = q_off;
+        const uint v_off = seq_id * sv3 + t * sv2 + head_id * sv1;
+
+        s_q[col] = FLOAT_TYPE(data_q[q_off + col]);
+        s_k[col] = FLOAT_TYPE(data_k[k_off + col]);
+
+        const uint gb_off = seq_id * sb3 + t * sb2 + head_id * sb1;
+
+        if (KDA != 0) {
+            const uint g_base = gb_off * S_V;
+            s_g[col] = exp(FLOAT_TYPE(data_g[g_base + col]));
+        }
+
+        barrier();
+
+        const FLOAT_TYPE v_val = FLOAT_TYPE(data_v[v_off + col]);
+        const FLOAT_TYPE beta_val = FLOAT_TYPE(data_beta[gb_off]);
+
+        if (KDA == 0) {
+            const FLOAT_TYPE g_val = exp(FLOAT_TYPE(data_g[gb_off]));
+
+            FLOAT_TYPE kv_col = 0.0;
+            [[unroll]] for (uint i = 0; i < S_V; i += 4) {
+                kv_col += dot(
+                    vec4(state[i], state[i+1], state[i+2], state[i+3]),
+                    vec4(s_k[i], s_k[i+1], s_k[i+2], s_k[i+3])
+                );
+            }
+
+            FLOAT_TYPE delta_col = (v_val - g_val * kv_col) * beta_val;
+
+            FLOAT_TYPE attn_col = 0.0;
+            [[unroll]] for (uint i = 0; i < S_V; i += 4) {
+                vec4 sv = vec4(state[i], state[i+1], state[i+2], state[i+3]);
+                vec4 kv = vec4(s_k[i], s_k[i+1], s_k[i+2], s_k[i+3]);
+                sv = g_val * sv + kv * delta_col;
+                state[i] = sv.x; state[i+1] = sv.y; state[i+2] = sv.z; state[i+3] = sv.w;
+
+                attn_col += dot(sv, vec4(s_q[i], s_q[i+1], s_q[i+2], s_q[i+3]));
+            }
+
+            data_dst[attn_off + col] = attn_col * scale;
+        } else {
+            FLOAT_TYPE kv_col = 0.0;
+            [[unroll]] for (uint i = 0; i < S_V; i += 4) {
+                vec4 gv = vec4(s_g[i], s_g[i+1], s_g[i+2], s_g[i+3]);
+                vec4 sv = vec4(state[i], state[i+1], state[i+2], state[i+3]);
+                vec4 kv = vec4(s_k[i], s_k[i+1], s_k[i+2], s_k[i+3]);
+                kv_col += dot(gv * sv, kv);
+            }
+
+            FLOAT_TYPE delta_col = (v_val - kv_col) * beta_val;
+
+            FLOAT_TYPE attn_col = 0.0;
+            [[unroll]] for (uint i = 0; i < S_V; i += 4) {
+                vec4 gv = vec4(s_g[i], s_g[i+1], s_g[i+2], s_g[i+3]);
+                vec4 sv = vec4(state[i], state[i+1], state[i+2], state[i+3]);
+                vec4 kv = vec4(s_k[i], s_k[i+1], s_k[i+2], s_k[i+3]);
+                sv = gv * sv + kv * delta_col;
+                state[i] = sv.x; state[i+1] = sv.y; state[i+2] = sv.z; state[i+3] = sv.w;
+
+                attn_col += dot(sv, vec4(s_q[i], s_q[i+1], s_q[i+2], s_q[i+3]));
+            }
+
+            data_dst[attn_off + col] = attn_col * scale;
+        }
+
+        attn_off += S_V * H;
+        barrier();
+    }
+
+    [[unroll]] for (uint i = 0; i < S_V; i++) {
+        data_dst[s_off + state_base + col * S_V + i] = state[i];
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "generic_head.glsl"
+#include "generic_unary_head.glsl"
 #include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
@@ -8,19 +8,22 @@
 
 layout(local_size_x = BLOCK_SIZE, 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[];};
-
 shared FLOAT_TYPE sum[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
+    const uint i3 = row / (p.ne11 * p.ne12);
+    const uint i3_offset = i3 * p.ne12 * p.ne11;
+    const uint i2 = (row - i3_offset) / p.ne11;
+    const uint i2_offset = i2 * p.ne11;
+    const uint i1 = row - i3_offset - i2_offset;
+
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]);
         sum[tid] += xi * xi;
     }
 
@@ -33,9 +36,9 @@ void main() {
         barrier();
     }
 
-    const FLOAT_TYPE scale = inversesqrt(max(sum[0], FLOAT_TYPE(p.param1)));
+    const FLOAT_TYPE scale = 1.0f / max(sqrt(sum[0]), FLOAT_TYPE(p.param1));
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));
     }
 }