whispercpp 1.3.5 → 1.3.7

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

@@ -1,4 +1,13 @@
-#if defined(DATA_A_Q4_0)
+#if defined(FA_MMQ_MIXED)
+// Mixed-K flash attention MMQ: superset cache that fits Q4_0/Q4_1/Q5_0/Q5_1/Q8_0.
+// Q4_*/Q5_* only use qs[0..3] and (for Q5_*) qh. Q8_0 uses qs[0..7]. Single-scale
+// types (Q4_0/Q5_0/Q8_0) leave dm.y unused.
+struct block_a_cache {
+    int32_t qs[8];
+    uint32_t qh;
+    FLOAT_TYPEV2 dm;
+};
+#elif defined(DATA_A_Q4_0)
 #define QUANT_R_MMQ 2
 struct block_a_cache {
     uint32_t qs[16/4];
@@ -8,7 +17,7 @@ struct block_a_cache {
 #define QUANT_R_MMQ 2
 struct block_a_cache {
     uint32_t qs[16/4];
-    FLOAT_TYPE_VEC2 dm;
+    FLOAT_TYPEV2 dm;
 };
 #elif defined(DATA_A_Q5_0)
 #define QUANT_R_MMQ 2
@@ -22,7 +31,7 @@ struct block_a_cache {
 struct block_a_cache {
     uint32_t qs[16/4];
     uint32_t qh;
-    FLOAT_TYPE_VEC2 dm;
+    FLOAT_TYPEV2 dm;
 };
 #elif defined(DATA_A_Q8_0)
 #define QUANT_R_MMQ 1
@@ -32,6 +41,12 @@ struct block_a_cache {
     int32_t qs[32/4];
     FLOAT_TYPE dm;
 };
+#elif defined(DATA_A_IQ4_NL)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE dm;
+};
 #elif defined(DATA_A_MXFP4)
 #define QUANT_R_MMQ 2
 struct block_a_cache {
@@ -43,36 +58,36 @@ struct block_a_cache {
 struct block_a_cache {
     uint32_t qs[2];
     u8vec2 scales;
-    FLOAT_TYPE_VEC2 dm;
+    FLOAT_TYPEV2 dm;
 };
 #elif defined(DATA_A_Q3_K)
 #define QUANT_R_MMQ 2
 struct block_a_cache {
     uint32_t qs[4];
-    FLOAT_TYPE_VEC2 d_scales;
+    FLOAT_TYPEV2 d_scales;
 };
 #elif defined(DATA_A_Q4_K)
 #define QUANT_R_MMQ 2
 struct block_a_cache {
     uint32_t qs[4];
-    FLOAT_TYPE_VEC2 dm;
+    FLOAT_TYPEV2 dm;
 };
 #elif defined(DATA_A_Q5_K)
 #define QUANT_R_MMQ 1
 struct block_a_cache {
     int32_t qs[8];
-    FLOAT_TYPE_VEC2 dm;
+    FLOAT_TYPEV2 dm;
 };
 #elif defined(DATA_A_Q6_K)
 #define QUANT_R_MMQ 1
 struct block_a_cache {
     int32_t qs[8];
-    FLOAT_TYPE_VEC2 d_scales;
+    FLOAT_TYPEV2 d_scales;
 };
 #endif
 
 struct block_b_cache
 {
     int32_t qs[8];
-    FLOAT_TYPE_VEC2 ds;
+    FLOAT_TYPEV2 ds;
 };

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

@@ -8,7 +8,6 @@
 #extension GL_KHR_shader_subgroup_basic : enable
 #endif
 
-#include "rte.glsl"
 #include "types.glsl"
 #include "utils.glsl"
 

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 = p.a_offset + i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
 #endif
     return ix;
 }
@@ -34,27 +34,21 @@ 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;
     }
+    idst += p.d_offset;
 
     if (i0 >= p.n_dims) {
         rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
@@ -63,7 +57,7 @@ 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;
 
@@ -77,26 +71,21 @@ 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;
     }
+    idst += p.d_offset;
 
     if (i0 >= p.n_dims) {
         rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
@@ -105,7 +94,7 @@ 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;
 
@@ -120,27 +109,21 @@ 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;
     }
+    idst += p.d_offset;
 
     if (i0 >= p.n_dims) {
         rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
@@ -156,26 +139,26 @@ 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 +174,13 @@ 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 = p.d_offset + 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 +189,11 @@ 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_head.glsl

@@ -2,7 +2,6 @@
 
 #extension GL_EXT_shader_16bit_storage : require
 
-#include "rte.glsl"
 #include "rope_params.glsl"
 
 layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;

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

@@ -1,28 +1,34 @@
 #if !defined(GGML_ROPE_PARAMS)
 #define GGML_ROPE_PARAMS
 
-#include "rte.glsl"
-
 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;
+
+    uint a_offset;
+    uint d_offset;
 };
 
 #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])));
+}

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

@@ -0,0 +1,49 @@
+#version 450
+
+#include "types.glsl"
+
+// Fused snake activation: y = x + sin(b * x)^2 * c
+//   data_a [ne0, ne1]   per element activation x    (A_TYPE)
+//   data_b [1,   ne1]   per channel multiplier      (float)
+//   data_c [1,   ne1]   per channel inverse scale   (float, precomputed as 1 / freq)
+//   data_d [ne0, ne1]   output y                    (D_TYPE)
+layout (binding = 0) readonly  buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly  buffer B {float  data_b[];};
+layout (binding = 2) readonly  buffer C {float  data_c[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint32_t ne0;
+    uint32_t ne1;
+} p;
+
+// Load A_TYPE to float
+float load_val(uint32_t idx) {
+#if defined(DATA_A_BF16)
+    return bf16_to_fp32(uint32_t(data_a[idx]));
+#else
+    return float(data_a[idx]);
+#endif
+}
+
+// Store float as D_TYPE
+void store_val(uint32_t idx, float v) {
+#if defined(DATA_D_BF16)
+    data_d[idx] = D_TYPE(fp32_to_bf16(v));
+#else
+    data_d[idx] = D_TYPE(v);
+#endif
+}
+
+void main() {
+    const uint32_t i0 = gl_GlobalInvocationID.x;
+    const uint32_t i1 = gl_GlobalInvocationID.y;
+    if (i0 >= p.ne0 || i1 >= p.ne1) return;
+
+    const uint32_t idx = i0 + i1 * p.ne0;
+    const float xi = load_val(idx);
+    const float s  = sin(data_b[i1] * xi);
+    store_val(idx, xi + s * s * data_c[i1]);
+}

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

@@ -5,12 +5,16 @@
 #include "types.glsl"
 
 layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(constant_id = 1) const uint TOKENS_PER_WG = 16;
+layout(constant_id = 2) const bool APPLY_BIAS = false;
+layout(constant_id = 3) const bool APPLY_SILU = false;
 
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z = 1) in;
 
 layout(binding = 0) readonly buffer Src0 { float src0[]; };
 layout(binding = 1) readonly buffer Src1 { float src1[]; };
-layout(binding = 2) buffer Dst { float dst[]; };
+layout(binding = 2) readonly buffer Bias { float bias[]; };
+layout(binding = 3) buffer Dst { float dst[]; };
 
 layout(push_constant) uniform PushConstants {
     uint nb01; uint nb02;
@@ -20,25 +24,37 @@ layout(push_constant) uniform PushConstants {
 };
 
 void main() {
-    const uint global_thread_id = gl_GlobalInvocationID.x;
-    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_GlobalInvocationID.x;
+    const uint i2 = gl_WorkGroupID.y * TOKENS_PER_WG + gl_LocalInvocationID.y;
     const uint i3 = gl_WorkGroupID.z;
 
-    if (global_thread_id >= nr || i2 >= n_t || i3 >= n_s) {
+    if (i1 >= nr || i2 >= n_t || i3 >= n_s) {
         return;
     }
 
-    const uint i1 = global_thread_id;
     const uint src0_base = i3 * (nb02 / 4) + i2 + i1 * (nb01 / 4);
     const uint src1_base = i1 * (nb11 / 4);
-    const uint dst_idx = i3 * (dst_nb2 / 4) + i2 * (dst_nb1 / 4) + i1;
 
     float sum = 0.0;
-    [[unroll]] for (uint i0 = 0; i0 < nc; i0++) {
-        const uint src0_idx = src0_base + i0;
-        const uint src1_idx = src1_base + i0;
-        sum += src0[src0_idx] * src1[src1_idx];
+
+    if (nc == 4) {
+        sum = dot(
+            vec4(src0[src0_base], src0[src0_base + 1], src0[src0_base + 2], src0[src0_base + 3]),
+            vec4(src1[src1_base], src1[src1_base + 1], src1[src1_base + 2], src1[src1_base + 3])
+        );
+    } else {
+        [[unroll]] for (uint i0 = 0; i0 < nc; i0++) {
+            sum += src0[src0_base + i0] * src1[src1_base + i0];
+        }
+    }
+
+    if (APPLY_BIAS) {
+        sum += bias[i1];
+    }
+    if (APPLY_SILU) {
+        sum = sum / (1.0f + exp(-sum));
     }
 
+    const uint dst_idx = i3 * (dst_nb2 / 4) + i2 * (dst_nb1 / 4) + i1;
     dst[dst_idx] = sum;
 }

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

@@ -1,6 +1,5 @@
 #version 450
 
-#include "rte.glsl"
 #include "types.glsl"
 #include "generic_unary_head.glsl"