whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl

@@ -29,7 +29,10 @@ layout (push_constant) uniform parameter
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
+    uint expert_i1;
+    uint nbi1;
 #else
+    uint base_work_group_y;
     uint ne02;
     uint ne12;
     uint broadcast2;
@@ -43,9 +46,9 @@ uint expert_id;
 
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.y;
+    const uint expert_i0 = gl_WorkGroupID.y;
 #else
-    const uint batch_idx = gl_GlobalInvocationID.y;
+    const uint batch_idx = gl_WorkGroupID.y + p.base_work_group_y;
 #endif
 
 #ifndef MUL_MAT_ID
@@ -60,7 +63,7 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_i0 + p.expert_i1 * p.nbi1];
 #endif
 
     a_offset =
@@ -71,13 +74,13 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #endif
     b_offset =
 #ifdef MUL_MAT_ID
-            (expert_idx % p.ne11) * p.stride_b;
+            (expert_i0 % p.ne11) * p.stride_b + p.expert_i1 * p.batch_stride_b;
 #else
             batch_idx * p.batch_stride_b;
 #endif
     d_offset =
 #ifdef MUL_MAT_ID
-            expert_idx * p.stride_d;
+            expert_i0 * p.stride_d + p.expert_i1 * p.batch_stride_d;
 #else
             batch_idx * p.batch_stride_d;
 #endif
@@ -103,12 +106,12 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
                     temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
@@ -158,12 +161,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                     temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
@@ -203,12 +206,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                     tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {

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

@@ -90,6 +90,8 @@ layout (push_constant) uniform parameter
     uint nbi1;
     uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
     uint k_split;
     uint ne02;
     uint ne12;
@@ -139,7 +141,7 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
     if (ic * BN >= data_expert_count[expert_idx]) {
         return;
     }
@@ -149,7 +151,7 @@ void main() {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;
 
     const uint i13 = batch_idx / p.ne12;
     const uint i12 = batch_idx % p.ne12;
@@ -366,7 +368,7 @@ void main() {
     const uint dc = ic * BN + warp_c * WN;
 
 #ifndef MUL_MAT_ID
-    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif
 
 #ifdef COOPMAT
@@ -375,6 +377,7 @@ void main() {
         [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
             coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
 
+            barrier();
             [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                 const uint row_i = dc + cm_col * TN + col + store_c;
                 if (row_i >= _ne1) break;
@@ -385,6 +388,7 @@ void main() {
                     data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
                 }
             }
+            barrier();
         }
     }
 #else
@@ -402,18 +406,22 @@ void main() {
                 // Full coopMat is within bounds, but stride_d is not aligned
                 coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
 
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
                 [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                     data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
                 }
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
             } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
                 // Partial coopMat is within bounds
                 coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
 
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
                 [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                     if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
                         data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
                     }
                 }
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
             }
         }
     }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -53,6 +53,8 @@ layout (push_constant) uniform parameter
     uint nbi1;
     uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
     uint k_split;
     uint ne02;
     uint ne12;
@@ -165,7 +167,9 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
         uint id = ids[iter++];
         uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
 
-        ballots_sh[gl_SubgroupID] = ballot;
+        if (gl_SubgroupInvocationID == 0) {
+            ballots_sh[gl_SubgroupID] = ballot;
+        }
         barrier();
 
         uint subgroup_base = 0;
@@ -197,7 +201,7 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
     if (ic * BN >= data_expert_count[expert_idx]) {
         return;
     }
@@ -215,7 +219,7 @@ void main() {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;
 
     const uint i13 = batch_idx / p.ne12;
     const uint i12 = batch_idx % p.ne12;
@@ -255,7 +259,7 @@ void main() {
 #else
     uint pos_a = batch_idx_a * (p.batch_stride_a / QUANT_K);
     uint pos_b = batch_idx * p.batch_stride_b;
-    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif
 
     uint stride_a = p.stride_a / QUANT_K;

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl

@@ -43,7 +43,9 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
         uint id = ids[iter++];
         uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
 
-        ballots_sh[gl_SubgroupID] = ballot;
+        if (gl_SubgroupInvocationID == 0) {
+            ballots_sh[gl_SubgroupID] = ballot;
+        }
         barrier();
 
         uint subgroup_base = 0;

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

@@ -57,6 +57,8 @@ layout (push_constant) uniform parameter
     uint nbi1;
     uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
     uint k_split;
     uint ne02;
     uint ne12;
@@ -108,7 +110,7 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
     if (ic * BN >= data_expert_count[expert_idx]) {
         return;
     }
@@ -118,7 +120,7 @@ void main() {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;
 
     const uint i13 = batch_idx / p.ne12;
     const uint i12 = batch_idx % p.ne12;
@@ -276,7 +278,7 @@ void main() {
     const uint dc = ic * BN + warp_c * WN;
 
 #ifndef MUL_MAT_ID
-    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif
 
     [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl

@@ -264,7 +264,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
         const i8vec2 scales = i8vec2(unpack8(uint32_t(((data_a_packed16[ib_k].scales[(is % 8      ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
                                                      (((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4))).xy); // vec4 used due to #12147
 
-        buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
+        buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
     }
 }
 
@@ -334,7 +334,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
                               (data_a[ib_k].scales[is+4] >>  4) | ((data_a[ib_k].scales[is  ] & 0xC0) >> 2));
         }
 
-        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
     }
 }
 
@@ -385,7 +385,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
         const uint is = iqs_k / 4;
         const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
 
-        buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
+        buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
     }
 }
 

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

@@ -112,12 +112,11 @@ void rms_norm(uint num_iters) {
 #if RMS_NORM_ROPE_FUSION
     barrier();
     rope_params rp = p.rope;
-    uint rope_row = (samp*nchannels + channel)*nrows + row;
     for (uint t = 2*tid; t < ncols; t += 2*BLOCK_SIZE) {
         if (rp.rope_mode == GGML_ROPE_TYPE_NEOX) {
-            rope_neox(t, rope_row, rp);
+            rope_neox(t, row, channel, samp, rp);
         } else if (rp.rope_mode == GGML_ROPE_TYPE_NORMAL) {
-            rope_norm(t, rope_row, rp);
+            rope_norm(t, row, channel, samp, rp);
         }
     }
 #endif

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl

@@ -4,12 +4,12 @@ float rope_yarn_ramp(const float low, const float high, const uint i0) {
     return 1.0f - min(1.0f, max(0.0f, y));
 }
 
-uint rope_a_coord(const uint i0, const uint i01, const uint i02, rope_params p) {
+uint rope_a_coord(const uint i0, const uint i01, const uint i02, const uint i03, rope_params p) {
 #if RMS_NORM_ROPE_FUSION
     // Per-row offset in shared memory
     const uint ix = i0;
 #else
-    const uint ix = i02*p.nb02 + i01*p.nb01 + i0;
+    const uint ix = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
 #endif
     return ix;
 }
@@ -34,26 +34,19 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out
     sin_theta = sin(theta) * mscale;
 }
 
-void rope_norm(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
+void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0;
-    const uint ix = rope_a_coord(i0, i01, i02, p);
+    uint idst = i0 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -63,7 +56,7 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
         return;
     }
 
-    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
 
@@ -77,25 +70,19 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
     rope_data_d[idst + 1] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }
 
-void rope_neox(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
+void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0/2;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0/2;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -105,7 +92,7 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
         return;
     }
 
-    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
 
@@ -120,26 +107,19 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
 }
 
 
-void rope_multi(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
+void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0/2;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0/2;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -156,26 +136,26 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     float theta_base = 0.0;
     if (p.is_imrope != 0) {
         if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
-            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
         } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
         } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
-            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
         } else {
-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
         }
     } else {
         if (sector < p.sections[0]) {
-            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= p.sections[0] && sector < sec_w) {
-            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
         }
     }
 
@@ -191,20 +171,13 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }
 
-void rope_vision(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
+void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    const uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    const uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     const int sect_dims = p.sections[0] + p.sections[1];
     const int sec_w = p.sections[1] + p.sections[0];
@@ -213,11 +186,11 @@ void rope_vision(const uint i0, const uint i1, rope_params p) {
     float theta_base = 0.0;
     if (sector < p.sections[0]) {
         const uint p0 = sector;
-        theta_base = rope_data_pos[i02]*pow(p.theta_scale, p0);
+        theta_base = rope_data_pos[i2]*pow(p.theta_scale, p0);
     }
     else if (sector >= p.sections[0] && sector < sec_w) {
         const uint p0 = sector - p.sections[0];
-        theta_base = rope_data_pos[i02 + ne2]*pow(p.theta_scale, p0);
+        theta_base = rope_data_pos[i2 + p.ne02]*pow(p.theta_scale, p0);
     }
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_multi(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_multi(i0, i1, i2, i3, pc);
 }

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_neox(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_neox(i0, i1, i2, i3, pc);
 }

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_norm(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_norm(i0, i1, i2, i3, pc);
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl

@@ -5,24 +5,29 @@
 
 struct rope_params {
     uint rope_mode;
-    uint ncols;
     uint nrows;
     uint n_dims;
     float freq_scale;
-    uint p_delta_rows;
     float freq_base;
     float ext_factor;
     float attn_factor;
     float corr_dims[2];
     float theta_scale;
     uint has_ff;
-    uint ne02;
-    uint nb01;
-    uint nb02;
     int sections[4];
     uint is_imrope;
     uint is_back;
     uint set_rows_stride;
+
+    uint ne00;
+    uint ne01;
+    uint ne02;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
 };
 
 #endif // !defined(GGML_ROPE_PARAMS)

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_vision(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_vision(i0, i1, i2, i3, pc);
 }

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

@@ -0,0 +1,21 @@
+#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;
+    }
+
+    data_d[i] = D_TYPE(sign(float(data_a[i])));
+}