whispercpp 1.3.6 → 1.3.8

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

@@ -21,7 +21,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
 
     if (iqs == 0) {
-        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+        buf_a[buf_ib].dm = FLOAT_TYPEV2(data_a_packed32[ib].dm);
     }
 #endif
 }
@@ -72,7 +72,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
 
     if (iqs == 0) {
-        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+        buf_a[buf_ib].dm = FLOAT_TYPEV2(data_a_packed32[ib].dm);
         buf_a[buf_ib].qh = data_a_packed32[ib].qh;
     }
 #endif
@@ -203,7 +203,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     buf_a[buf_ib].qs[iqs] = vals0 | (vals1 << 2) | (vals2 << 4) | (vals3 << 6);
 
     if (iqs == 0) {
-        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+        buf_a[buf_ib].dm = FLOAT_TYPEV2(data_a_packed32[ib_k].dm);
         buf_a[buf_ib].scales = unpack8(uint32_t(data_a_packed16[ib_k].scales[iqs_k / 8])).xy; // vec4 used due to #12147
     }
 }
@@ -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_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
+        buf_a[buf_ib].d_scales = FLOAT_TYPEV2(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(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
+        buf_a[buf_ib].dm = FLOAT_TYPEV2(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_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
+        buf_a[buf_ib].d_scales = FLOAT_TYPEV2(float(data_a_packed16[ib_k].d) * vec2(scales));
     }
 }
 
@@ -426,7 +426,7 @@ void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bo
         const uint ib_inner = ib % 4;
 
         if (iqs == 0) {
-            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
+            buf_b[buf_ib].ds = FLOAT_TYPEV2(data_b[ib_outer].ds[ib_inner]);
         }
 
         const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
@@ -436,7 +436,7 @@ void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bo
         buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
     } else {
         if (iqs == 0) {
-            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
+            buf_b[buf_ib].ds = FLOAT_TYPEV2(0.0f);
         }
 
         buf_b[buf_ib].qs[iqs * 4    ] = 0;

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/norm.comp

@@ -1,26 +1,26 @@
 #version 450
 
-#include "generic_head.glsl"
 #include "types.glsl"
+#include "generic_unary_head.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
-
 shared vec2 sum[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
+    const uint a_base = get_aoffset() + src0_idx(row * p.ne00);
+    const uint d_base = get_doffset() + dst_idx(row * p.ne10);
+
     sum[tid] = vec2(0.0f, 0.0f);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        const float xi = float(data_a[row*p.KX + col]);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        const float xi = float(data_a[a_base + i0*p.nb00]);
         sum[tid].x += xi;
         sum[tid].y += xi * xi;
     }
@@ -34,11 +34,11 @@ void main() {
         barrier();
     }
 
-    const float mean = sum[0].x / p.KX;
-    const float var = sum[0].y / p.KX - mean * mean;
+    const float mean = sum[0].x / p.ne00;
+    const float var = sum[0].y / p.ne00 - mean * mean;
     const float inv_std = inversesqrt(var + p.param1);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        data_d[d_base + i0*p.nb10] = D_TYPE((float(data_a[a_base + i0*p.nb00]) - mean) * inv_std);
     }
 }

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

@@ -13,11 +13,11 @@ void main() {
     }
 
     // Destination multi-index (inlined dst_idx)
-    const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i13 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
     const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
-    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
     const uint i12_offset = i12*p.ne11*p.ne10;
-    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
     const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
     const uint d_idx = i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
 

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

@@ -20,11 +20,11 @@ void main() {
         return;
     }
 
-    const uint i3 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i3 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
     const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
-    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
     const uint i2_offset = i2*p.ne11*p.ne10;
-    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
     const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
 
     const uint p1 = floatBitsToUint(p.param1);

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

@@ -9,7 +9,7 @@ uint rope_a_coord(const uint i0, const uint i01, const uint i02, const uint i03,
     // Per-row offset in shared memory
     const uint ix = i0;
 #else
-    const uint ix = i03*p.nb03 + 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;
 }
@@ -48,6 +48,7 @@ void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_
         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]);
@@ -84,6 +85,7 @@ void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_
         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]);
@@ -121,6 +123,7 @@ void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope
         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]);
@@ -176,7 +179,7 @@ void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rop
         return;
     }
 
-    const uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    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];

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_params.glsl

@@ -1,8 +1,6 @@
 #if !defined(GGML_ROPE_PARAMS)
 #define GGML_ROPE_PARAMS
 
-#include "rte.glsl"
-
 struct rope_params {
     uint rope_mode;
     uint nrows;
@@ -28,6 +26,9 @@ struct rope_params {
     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/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

@@ -6,12 +6,15 @@
 
 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_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;
@@ -45,6 +48,13 @@ void main() {
         }
     }
 
+    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"
 
@@ -18,11 +17,11 @@ void main() {
         return;
     }
 
-    const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L);
+    const uint i03 = fastdiv(idx, p.ne0_012mp, fastdiv_L(p.ne0_Ls, 0));
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L);
+    const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, fastdiv_L(p.ne0_Ls, 1));
     const uint i02_offset = i02*p.ne01*p.ne00;
-    const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L);
+    const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, fastdiv_L(p.ne0_Ls, 2));
     const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
 
     int param = floatBitsToInt(p.param1);

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

@@ -31,6 +31,7 @@
 #else
 #define A_TYPE float16_t
 #endif
+#define A_TYPE_PACKED32 f16vec2
 #endif
 
 #if defined(DATA_A_BF16)
@@ -44,6 +45,7 @@
 #else
 #define A_TYPE uint16_t
 #endif
+#define A_TYPE_PACKED32 uint32_t
 #endif
 
 #define QUANT_K_Q4_0 32
@@ -188,6 +190,22 @@ struct block_q8_0_packed16
 #define DATA_A_QUANT_LEGACY
 #endif
 
+#define QUANT_K_Q1_0 128
+#define QUANT_R_Q1_0 1
+
+struct block_q1_0
+{
+    float16_t d;
+    uint8_t qs[QUANT_K_Q1_0 / 8];
+};
+
+#if defined(DATA_A_Q1_0)
+#define QUANT_K QUANT_K_Q1_0
+#define QUANT_R QUANT_R_Q1_0
+#define QUANT_AUXF 1
+#define A_TYPE block_q1_0
+#endif
+
 #define QUANT_K_Q8_1 32
 #define QUANT_R_Q8_1 1
 
@@ -580,9 +598,10 @@ const uint[1024] iq1s_grid_const = {
     0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
 };
 
+#if defined(NEEDS_IQ1S_GRID_GPU)
 // Same content as iq1s_grid_const except each 2-bit value is expanded to 4-bit
 // and has 1 added to it (allows packed values to be extracted with & 0x0F0F0F0F
-// and 0xF0F0F0F0).
+// and 0xF0F0F0F0). This is only used by the q8_1/int-dot vector path.
 const uint32_t[2048] iq1s_grid_gpu_const = {
     0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
     0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
@@ -841,9 +860,12 @@ const uint32_t[2048] iq1s_grid_gpu_const = {
     0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
     0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
 };
+#endif
 
 shared uint16_t iq1s_grid[2048];
+#if defined(NEEDS_IQ1S_GRID_GPU)
 shared uint32_t iq1s_grid_gpu[2048];
+#endif
 
 #define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
@@ -857,12 +879,14 @@ void init_iq_shmem(uvec3 wgsize)
             iq1s_grid[2*idx+1] = g.y;
         }
     }
+#if defined(NEEDS_IQ1S_GRID_GPU)
     [[unroll]] for (uint i = 0; i < iq1s_grid_gpu_const.length(); i += wgsize.x) {
         uint idx = i + gl_LocalInvocationIndex.x;
         if (iq1s_grid_gpu_const.length() % wgsize.x == 0 || idx < iq1s_grid_gpu_const.length()) {
             iq1s_grid_gpu[idx] = iq1s_grid_gpu_const[idx];
         }
     }
+#endif
     barrier();
 }
 #endif
@@ -1676,6 +1700,7 @@ struct block_iq4_nl_packed16
 #if defined(DATA_A_IQ4_NL)
 #define QUANT_K QUANT_K_IQ4_NL
 #define QUANT_R QUANT_R_IQ4_NL
+#define QUANT_AUXF 1
 #define A_TYPE block_iq4_nl
 #define A_TYPE_PACKED16 block_iq4_nl_packed16
 #endif
@@ -1696,6 +1721,29 @@ struct block_mxfp4
 #define A_TYPE block_mxfp4
 #endif
 
+#define QUANT_K_NVFP4 64
+#define QUANT_R_NVFP4 1
+
+struct block_nvfp4
+{
+    uint8_t d[QUANT_K_NVFP4 / 16];
+    uint8_t qs[QUANT_K_NVFP4 / 2];
+};
+
+struct block_nvfp4_packed32
+{
+    uint32_t d[QUANT_K_NVFP4 / 16 / 4];
+    uint32_t qs[QUANT_K_NVFP4 / 2 / 4];
+};
+
+#if defined(DATA_A_NVFP4)
+#define QUANT_K QUANT_K_NVFP4
+#define QUANT_R QUANT_R_NVFP4
+#define QUANT_AUXF 1
+#define A_TYPE block_nvfp4
+#define A_TYPE_PACKED32 block_nvfp4_packed32
+#endif
+
 #if defined(DATA_A_IQ4_NL) || defined(DATA_A_IQ4_XS)
 const int8_t kvalues_iq4nl_const[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
@@ -1715,7 +1763,7 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
-#if defined(DATA_A_MXFP4)
+#if defined(DATA_A_MXFP4) || defined(DATA_A_NVFP4)
 const int8_t kvalues_mxfp4_const[16] = {
     int8_t(0), int8_t(1), int8_t(2), int8_t(3), int8_t(4), int8_t(6), int8_t(8), int8_t(12),
     int8_t(0), int8_t(-1), int8_t(-2), int8_t(-3), int8_t(-4), int8_t(-6), int8_t(-8), int8_t(-12),
@@ -1723,6 +1771,24 @@ const int8_t kvalues_mxfp4_const[16] = {
 
 shared int8_t kvalues_mxfp4[16];
 
+#if defined(DATA_A_NVFP4)
+// UE4M3 scale in NVFP4 blocks use only 7 bits; sign (bit 7) is always zero.
+shared float ue4m3_fp32_lut[128];
+
+float ue4m3_to_fp32_build(uint u) {
+    if (u == 0u || u == 127u) {
+        return 0.0;
+    }
+    const uint exp = (u >> 3) & 15u;
+    const uint man = u & 7u;
+    if (exp == 0u) {
+        return float(man) * (1.0 / 512.0);
+    }
+    const uint bits = (exp + 120u) << 23 | (man << 20);
+    return uintBitsToFloat(bits);
+}
+#endif
+
 #define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
@@ -1730,6 +1796,11 @@ void init_iq_shmem(uvec3 wgsize)
     for (uint i = gl_LocalInvocationIndex.x; i < kvalues_mxfp4.length(); i += wgsize.x) {
         kvalues_mxfp4[i] = kvalues_mxfp4_const[i];
     }
+#if defined(DATA_A_NVFP4)
+    for (uint i = gl_LocalInvocationIndex.x; i < 128u; i += wgsize.x) {
+        ue4m3_fp32_lut[i] = ue4m3_to_fp32_build(i);
+    }
+#endif
     barrier();
 }
 #endif
@@ -1766,6 +1837,12 @@ float e8m0_to_fp32(uint8_t x) {
     return uintBitsToFloat(bits);
 }
 
+#if defined(DATA_A_NVFP4)
+float ue4m3_to_fp32(uint8_t x) {
+    return ue4m3_fp32_lut[uint(x)];
+}
+#endif
+
 #if BDA
 
 #extension GL_EXT_buffer_reference : enable