whispercpp 1.3.3 → 1.3.5

data/ext/sources/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;
 }
@@ -61,8 +65,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    if (r < N && c < D) {
-        uint32_t offset = (iq2 + r) * D + c;
+    if (r < N && c < HSV) {
+        uint32_t offset = (iq2 + r) * HSV + c;
         data_o[o_offset + offset] = D_TYPE(elem);
     }
     return elem;
@@ -86,9 +90,9 @@ void main() {
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
 
-    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
-    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
-    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
+    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, HSK);
+    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, HSK);
+    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, HSV);
 
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
@@ -104,16 +108,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
+    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;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -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);
 
@@ -130,15 +138,58 @@ void main() {
         coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
     }
 
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+    }
+
     [[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));
+
+                // 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, D, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC);
+        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) {
@@ -148,15 +199,7 @@ void main() {
             }
         }
 
-        if (p.mask != 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, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
-
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
             S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
         }
 
@@ -203,65 +246,97 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV_pad, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV_pad) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> eMdiag;
 
         // 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, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+        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, D, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+        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 = D * p.ne1 * split_k_index;
+        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 = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        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);
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> S;
+        coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
+
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Mr;
+
+        // resize M by using smear/reduce
+        coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
+
+        // O, Ldiag, Mr all have the same type so all element locations match
+        [[unroll]] for (uint32_t i = 0; i < Ldiag.length(); ++i) {
+            ACC_TYPE sink = S[i];
+
+            ACC_TYPE ms = ACC_TYPE(1.0f);
+            ACC_TYPE vs = ACC_TYPE(1.0f);
+
+            if (sink > Mr[i]) {
+                ms = exp(Mr[i] - sink);
+
+                O[i] *= ms;
+            } else {
+                vs = exp(sink - Mr[i]);
+            }
+
+            Ldiag[i] = Ldiag[i]*ms + vs;
+        }
+    }
+
     [[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;
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+#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, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+    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 {
         tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
-        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
+        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, HSV);
 
         // permute dimensions
         tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
 
-        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV_pad), tensorViewPermute);
     }
 }

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

@@ -2,58 +2,119 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
-#define BLOCK_SIZE 32
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {float data_a[];};
-layout (binding = 1) writeonly buffer D {float data_d[];};
+layout (binding = 1) readonly buffer B {float data_s[];};
+layout (binding = 2) writeonly buffer D {float data_d[];};
 
 layout (push_constant) uniform parameter {
     uint D;
     uint N;
+    uint ne3;
     uint k_num;
+    uint sinks;
 } p;
 
+shared float tmpsh[BLOCK_SIZE];
+
 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;
 
     uint D = p.D;
     uint N = p.N;
     uint k_num = p.k_num;
 
-    uint l_offset = D * N * k_num + n;
-    uint m_offset = D * N * k_num + N + n;
+    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;
 
     // Compute the max m value for the row
     float m_max = -1.0/0.0;
-    [[unroll]] for (uint k = 0; k < k_num; ++k) {
-        float m = data_a[m_offset + k * lm_stride];
+    for (uint k = 0; k + tid < k_num; k += BLOCK_SIZE) {
+        float m = data_a[m_offset + (k + tid) * lm_stride];
         m_max = max(m_max, m);
     }
 
+    // reduce across the workgroup
+    tmpsh[tid] = m_max;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            m_max = max(m_max, tmpsh[tid + s]);
+            tmpsh[tid] = m_max;
+        }
+        barrier();
+    }
+    m_max = tmpsh[0];
+
+    barrier();
+
     // Compute L based on m_max
     float L = 0;
-    [[unroll]] for (uint k = 0; k < k_num; ++k) {
-        float l = data_a[l_offset + k * lm_stride];
-        float m = data_a[m_offset + k * lm_stride];
+    for (uint k = 0; k + tid < k_num; k += BLOCK_SIZE) {
+        float l = data_a[l_offset + (k + tid) * lm_stride];
+        float m = data_a[m_offset + (k + tid) * lm_stride];
         L += exp(m - m_max) * l;
     }
 
-    L = 1.0 / L;
+    // reduce across the workgroup
+    tmpsh[tid] = L;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            L += tmpsh[tid + s];
+            tmpsh[tid] = L;
+        }
+        barrier();
+    }
+    L = tmpsh[0];
+
+    float sink;
+    if (p.sinks != 0) {
+        sink = data_s[n];
+
+        float ms = 1.0f;
+        float vs = 1.0f;
 
+        if (sink > m_max) {
+            ms = exp(m_max - sink);
+        } else {
+            vs = exp(sink - m_max);
+        }
+
+        L = L*ms + vs;
+    }
+
+    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;
     // Scale and sum the O contributions based on m_max and store the result to memory
-    for (uint d = tid; d < D; d += BLOCK_SIZE) {
+    if (d < D) {
         float O = 0.0;
         [[unroll]] for (uint k = 0; k < k_num; ++k) {
-            uint o_offset = D * N * k + D * n + d;
+            uint o_offset = D * N * (k + iq3 * k_num) + D * n + d;
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
+        if (p.sinks != 0) {
+            if (sink > m_max) {
+                float ms = 1.0f;
+                ms = exp(m_max - sink);
+                O *= ms;
+            }
+        }
         O *= L;
-        data_d[D * n + d] = O;
+
+        const float FLT_MAX = uintBitsToFloat(0x7F7FFFFF);
+        O = clamp(O, -FLT_MAX, FLT_MAX);
+
+        data_d[iq3 * D * N + D * n + d] = O;
     }
 }

data/ext/sources/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));
+}

data/ext/sources/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"

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

@@ -0,0 +1,27 @@
+#version 450
+
+#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/
+const float p_erf  = 0.3275911f;
+const float a1_erf = 0.254829592f;
+const float a2_erf = -0.284496736f;
+const float a3_erf = 1.421413741f;
+const float a4_erf = -1.453152027f;
+const float a5_erf = 1.061405429f;
+
+const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+
+float op(float a, float b) {
+    const float a_div_sqr2 = a * SQRT_2_INV;
+    const float sign_x = sign(a_div_sqr2);
+    const float x = abs(a_div_sqr2);
+    const float t = 1.0f / (1.0f + p_erf * x);
+    const float y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    const float erf_approx = sign_x * y;
+
+    return 0.5f * a * (1.0f + erf_approx) * b;
+}
+
+#include "glu_main.glsl"

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

@@ -0,0 +1,11 @@
+#version 450
+
+#include "glu_head.glsl"
+
+const float GELU_QUICK_COEF = -1.702f;
+
+float op(float a, float b) {
+    return a * (1.0f / (1.0f + exp(GELU_QUICK_COEF * a))) * b;
+}
+
+#include "glu_main.glsl"

data/ext/sources/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
 

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

@@ -0,0 +1,39 @@
+#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() {
+    // based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+    // ref: https://www.johndcook.com/blog/python_erf/
+    const float p_erf  = 0.3275911f;
+    const float a1_erf = 0.254829592f;
+    const float a2_erf = -0.284496736f;
+    const float a3_erf = 1.421413741f;
+    const float a4_erf = -1.453152027f;
+    const float a5_erf = 1.061405429f;
+
+    const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float a = float(data_a[i]);
+    const float a_div_sqr2 = a * SQRT_2_INV;
+    const float sign_x = sign(a_div_sqr2);
+    const float x = abs(a_div_sqr2);
+    const float t = 1.0f / (1.0f + p_erf * x);
+    const float y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    const float erf_approx = sign_x * y;
+
+    data_d[i] = D_TYPE(0.5f * a * (1.0f + erf_approx));
+}

data/ext/sources/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
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{generic_binary_head.comp → generic_binary_head.glsl}

@@ -1,6 +1,12 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_control_flow_attributes : require
 
+#include "rte.glsl"
+#include "utils.glsl"
+#if RMS_NORM_ROPE_FUSION
+#include "rope_params.glsl"
+#endif
+
 layout (push_constant) uniform parameter
 {
     uint ne;
@@ -9,11 +15,23 @@ 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[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 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;
@@ -26,25 +44,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
 uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
 uint get_doffset() { return p.misalign_offsets & 0xFF; }
 
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
 
 void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
 }
 
 uint src0_idx(uint i00, uint i01, uint i02, uint i03) {

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{generic_head.comp → generic_head.glsl}

@@ -6,4 +6,6 @@ layout (push_constant) uniform parameter
     uint KY;
     float param1;
     float param2;
+    float param3;
+    float param4;
 } p;

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{generic_unary_head.comp → generic_unary_head.glsl}

@@ -18,6 +18,13 @@ layout (push_constant) uniform parameter
 } p;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 uint get_idx() {