whispercpp 1.3.2 → 1.3.4

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp

@@ -0,0 +1,98 @@
+#version 450
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};   // src0 - kernel:    [K, Cout, Cin]
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};   // src1 - input:     [L, Cin]
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};     // dst - result      [KL, Cout]
+
+layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+} p;
+
+
+uint32_t Cout_idx = gl_WorkGroupID.x;
+const uint32_t bs = gl_WorkGroupSize.x;
+uint32_t tid = gl_LocalInvocationID.x;
+// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
+uint32_t tmp_len = bs*p.s0+p.K;
+shared D_TYPE tmp[4096];
+
+uint splitWork(uint workSize){
+    return (bs + workSize -1) / bs;
+}
+
+void main(){
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        if(idx < tmp_len){
+            tmp[idx] = 0.0;
+        }
+    }
+
+    uint32_t L_blocks = splitWork(p.L);
+    for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
+        if(L_block_id > 0){
+            barrier();
+            // Shift values in tmp to the current processing window
+            for(int i = 0; i < splitWork(tmp_len); i++){
+                uint32_t idx = i*bs+tid;
+                if(idx >= bs*p.s0 && idx < tmp_len){
+                    tmp[idx-bs*p.s0] = tmp[idx];
+                    tmp[idx] = 0.0;
+                }else if(idx >= p.K && idx < bs*p.s0){
+                    tmp[idx] = 0.0;
+                }
+            }
+        }
+        barrier();
+
+        // Save contributions of the block to tmp
+        uint32_t L_idx = L_block_id*bs + tid;
+        for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){
+            D_TYPE dp = 0.0;
+            for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){
+                A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02];
+                if(L_idx < p.L){
+                    B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11];
+                    dp = fma(elemKrn, elemInp, dp);
+                }
+            }
+            tmp[tid*p.s0 + K_idx] += dp;
+            barrier();
+        }
+
+        // Save the computed values except the last block that can have different size
+        uint32_t KLb_idx = L_block_id*bs*p.s0;
+        if(L_block_id < L_blocks-1){
+            for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){
+                uint32_t sh_idx = p.s0*tid+s0_idx;
+                uint32_t KL_idx = KLb_idx+sh_idx;
+                if(KL_idx < p.KL){
+                    data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx];
+                }
+            }
+        }
+    }
+
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx;
+        if(KL_idx < p.KL){
+            data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx];
+        }
+    }
+}

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

@@ -4,8 +4,8 @@
 #include "generic_unary_head.comp"
 #include "dequant_funcs.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
+#if defined(DATA_A_IQ4_NL) || defined(DATA_A_MXFP4)
+// 16 invocations needed for init_iq_shmem
 layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
 #else
 layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;

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

@@ -1,22 +1,33 @@
 #version 450
 
-#if RTE16
-#extension GL_EXT_spirv_intrinsics : enable
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif // RTE16
-
+#include "rte.comp"
 #include "types.comp"
-#include "generic_unary_head.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
-layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+#if defined(SET_ROWS) && QUANT_K == 1
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 512;
 #else
-layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 32;
 #endif
 
 layout (binding = 0) readonly buffer S {float data_s[];};
+
+#if defined(SET_ROWS)
+#include "generic_binary_head.comp"
+layout (binding = 1) readonly buffer C {B_TYPE data_i[];};
+layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
+
+#if B_SIZE == 64
+#define DATA_I_SWIZZLE .x
+#else
+#define DATA_I_SWIZZLE
+#endif
+
+#else
+#include "generic_unary_head.comp"
 layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
+#endif
 
 #if defined(DATA_A_Q4_0)
 void quantize(uint dst_idx, uint src_idx)
@@ -221,15 +232,56 @@ void quantize(uint dst_idx, uint src_idx)
 }
 #endif
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(data_s[src_idx]);
+}
+#endif
+
+#if defined(DATA_A_BF16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx]));
+}
+#endif
+
+#if defined(SET_ROWS)
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
-    if (gl_LocalInvocationIndex.x != 0) {
+#endif
+
+    const uint idx = ((gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x) * BLOCK_SIZE + gl_LocalInvocationID.x) * QUANT_K;
+
+    if (idx >= p.ne) {
         return;
     }
+
+    uint i00, i01, i02, i03;
+    get_indices(idx, i00, i01, i02, i03);
+
+    uint i12 = fastmod(i03, p.ne12);
+    uint i11 = fastmod(i02, p.ne11);
+    uint i10 = i01;
+
+    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()] DATA_I_SWIZZLE;
+
+    uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
+    uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();
+
+    quantize(dst_idx, src0_idx);
+}
+
+#else
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-    const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
+    const uint idx = (gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x) * QUANT_K;
 
     if (idx >= p.ne) {
         return;
@@ -240,3 +292,5 @@ void main() {
 
     quantize(dst_idx, src_idx);
 }
+
+#endif

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

@@ -434,6 +434,18 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    vec2 v0 = dequantize(ib, iqs, a_offset);
+    vec2 v1 = dequantize(ib, iqs + 1, a_offset);
+    return vec4(v0.x, v0.y, v1.x, v1.y);
+}
+#endif
+
 #if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(0, 0);
@@ -455,8 +467,150 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(e8m0_to_fp32(data_a[a_offset + ib].e), 0);
+}
+#endif
+
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
 }
 #endif
+
+#if defined(DATA_A_Q2_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+    const uint scalesi = iqs / 8;                      // 0..15
+    const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
+
+    const uvec2 qs = uvec2(data_a[a_offset + ib].qs[qsi], data_a[a_offset + ib].qs[qsi + 1]);
+    const uint scales = data_a[a_offset + ib].scales[scalesi];
+    const vec2 d = vec2(data_a[a_offset + ib].d);
+
+    return d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q3_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 64;                     // 0,1
+    const uint qsi = n * 32 + (iqs % 16) * 2;    // 0,2,4..62
+    const uint hmi =          (iqs % 16) * 2;    // 0,2,4..30
+    const uint j = (iqs % 64) / 4;               // 0..3
+    const uint is = iqs / 8;                     // 0..15
+    const uint halfsplit = ((iqs % 64) / 16);    // 0,1,2,3
+    const uint qsshift = halfsplit * 2;          // 0,2,4,6
+    const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
+
+    const int8_t us = int8_t(((data_a[a_offset + ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
+                          | (((data_a[a_offset + ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
+    const float dl = float(data_a[a_offset + ib].d) * float(us - 32);
+
+    return vec2(dl * float(int8_t((data_a[a_offset + ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[a_offset + ib].hmask[hmi    ] & m) != 0) ? 0 : 4)),
+                dl * float(int8_t((data_a[a_offset + ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[a_offset + ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 32;                   // 0,1,2,3
+    const uint b = (iqs % 32) / 16;            // 0,1
+    const uint is = 2 * n + b;                 // 0..7
+    const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+
+    const vec2 loadd = vec2(data_a[a_offset + ib].d);
+
+    const uint scidx0 = (is < 4) ? is : (is + 4);
+    const uint scidx1 = (is < 4) ? is : (is - 4);
+    const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint scidxshift1 = (is < 4) ? 0 : 2;
+    const uint mbidx0 = is + 4;
+    const uint mbidx1 = (is < 4) ? is + 4 : is;
+    const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+    const uint mbidxshift0 = (is < 4) ? 0 : 4;
+    const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+    const uint8_t sc = uint8_t((data_a[a_offset + ib].scales[scidx0] & 0xF) | ((data_a[a_offset + ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+    const uint8_t mbyte = uint8_t((data_a[a_offset + ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[a_offset + ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+    const float d = loadd.x * sc;
+    const float m = -loadd.y * mbyte;
+
+    return vec2(fma(d, float((data_a[a_offset + ib].qs[qsi    ] >> (b * 4)) & 0xF), m),
+                fma(d, float((data_a[a_offset + ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q5_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 32;                   // 0,1,2,3
+    const uint b = (iqs % 32) / 16;            // 0,1
+    const uint is = 2 * n + b;                 // 0..7
+    const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+    const uint qhi = (iqs % 16) * 2;           // 0,2,4..30
+
+    const uint8_t hm = uint8_t(1 << (iqs / 16));
+
+    const vec2 loadd = vec2(data_a[a_offset + ib].d);
+
+    const uint scidx0 = (is < 4) ? is : (is + 4);
+    const uint scidx1 = (is < 4) ? is : (is - 4);
+    const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint scidxshift1 = (is < 4) ? 0 : 2;
+    const uint mbidx0 = is + 4;
+    const uint mbidx1 = (is < 4) ? is + 4 : is;
+    const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+    const uint mbidxshift0 = (is < 4) ? 0 : 4;
+    const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+    const uint8_t sc    = uint8_t((data_a[a_offset + ib].scales[scidx0] & 0xF)                         | ((data_a[a_offset + ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+    const uint8_t mbyte = uint8_t(((data_a[a_offset + ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[a_offset + ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+    const float d = loadd.x * sc;
+    const float m = -loadd.y * mbyte;
+
+    return vec2(fma(d, float((data_a[a_offset + ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[a_offset + ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m),
+                fma(d, float((data_a[a_offset + ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[a_offset + ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q6_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 64;                    // 0,1
+    const uint b = (iqs % 64) / 32;             // 0,1
+    const uint is_b = (iqs % 16) / 8;           // 0,1
+    const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+    const uint is = 8 * n + qhshift + is_b;     // 0..15
+    const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+    const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+    const float dscale = float(data_a[a_offset + ib].d) * float(data_a[a_offset + ib].scales[is]);
+
+    return vec2(dscale * float(int8_t(((data_a[a_offset + ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[a_offset + ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32),
+                dscale * float(int8_t(((data_a[a_offset + ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[a_offset + ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -654,6 +654,25 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufMXFP4 {
+   block_mxfp4 block;
+};
+
+float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float d = e8m0_to_fp32(bl.block.e);
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+    float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
+    return ret;
+}
+#endif
+
 #if defined(DATA_A_Q4_0)
 #define dequantFuncA dequantFuncQ4_0
 #elif defined(DATA_A_Q4_1)
@@ -696,4 +715,6 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 #define dequantFuncA dequantFuncIQ4_XS
 #elif defined(DATA_A_IQ4_NL)
 #define dequantFuncA dequantFuncIQ4_NL
+#elif defined(DATA_A_MXFP4)
+#define dequantFuncA dequantFuncMXFP4
 #endif

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp

@@ -29,7 +29,7 @@ void main() {
         uint qs = data_a[ib].qs[4 * ib32 + l];
         const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
         qs |= (qh << (8 - 2 * l)) & 0x300;
-        const uvec2 grid = iq2s_grid[qs & 511];
+        const uvec2 grid = iq2s_grid[qs];
         const u8vec4 grid0 = unpack8(grid.x);
         const u8vec4 grid1 = unpack8(grid.y);
         data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp

@@ -33,7 +33,8 @@ void main() {
     [[unroll]] for (uint l = 0; l < 4; ++l) {
         const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
         const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit
-        const uvec2 grid = iq2xxs_grid[data_a[ib].qs[8 * is + l]];
+        const uint qs = data_a[ib].qs[8 * is + l];
+        const uvec2 grid = iq2xxs_grid[qs];
         const u8vec4 grid0 = unpack8(grid.x);
         const u8vec4 grid1 = unpack8(grid.y);
         data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp

@@ -22,15 +22,16 @@ void main() {
     const uint b_idx = 256 * ib + 32 * is;
 
     const float d = float(data_a[ib].d);
-    const float db = d * (1 + 2 * ((data_a[ib].scales[is] >> (4 * (is % 2))) & 0xf));
+    const float db = d * (1 + 2 * ((data_a[ib].scales[is / 2] >> (4 * (is % 2))) & 0xf));
 
     // We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes.
     uint qh = data_a[ib].qh[is];
     [[unroll]] for (uint l = 0; l < 8; ++l) {
-        uint qs = data_a[ib].qs[8 * is + l];
-        uint gidx = qs | ((qh << (8 - l)) & 256);
-        uint8_t signs = data_a[ib].signs[8 * is + l / 2] >> (4 * (l & 1));
-        u8vec4 grid = unpack8(iq3s_grid[gidx]);
+        const uint iqs = 8 * is + l;
+        const uint qs = data_a[ib].qs[iqs];
+        const uint gidx = qs | ((qh << (8 - l)) & 256);
+        const uint8_t signs = data_a[ib].signs[iqs / 2] >> (4 * (l & 1));
+        const u8vec4 grid = unpack8(iq3s_grid[gidx]);
         data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0));

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp

@@ -35,8 +35,10 @@ void main() {
         const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
         // Restore parity bit.
         const uint sign8 = sign7 | (bitCount(sign7) << 7);
-        const u8vec4 grid0 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l]]);
-        const u8vec4 grid1 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l + 1]]);
+        const uint qs0 = data_a[ib].qs[8 * is + 2 * l];
+        const uint qs1 = data_a[ib].qs[8 * is + 2 * l + 1];
+        const u8vec4 grid0 = unpack8(iq3xxs_grid[qs0]);
+        const u8vec4 grid1 = unpack8(iq3xxs_grid[qs1]);
         data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp

@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_mxfp4 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
+
+    const float d = e8m0_to_fp32(data_a[ib].e);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]);
+        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]);
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
         const uint tid = gl_LocalInvocationID.x;

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

@@ -0,0 +1,21 @@
+#version 450
+
+#include "rte.comp"
+#include "generic_head.comp"
+#include "types.comp"
+
+#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(exp(float(data_a[i])));
+}