@novastera-oss/llamarn 0.3.1 → 0.4.1

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp

@@ -0,0 +1,65 @@
+#version 450
+
+#include "generic_binary_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+
+#define BLOCK_SIZE 128
+
+layout (constant_id = 1) const bool do_multiply = false;
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 3, std430) readonly buffer PartialsBuf {float partial_sums[];};
+
+shared FLOAT_TYPE sumsh[BLOCK_SIZE];
+
+void main() {
+    const uint ncols     = p.ne00;
+    const uint nrows     = gl_NumWorkGroups.x;
+    const uint nchannels = gl_NumWorkGroups.y;
+
+    const uint row       = 0;
+    const uint channel   = gl_WorkGroupID.y;
+    const uint samp      = gl_WorkGroupID.z;
+    // The work is split across multiple workgroups in the x dimension. Each invocation
+    // processes one element
+    const uint tid       = gl_GlobalInvocationID.x;
+
+    const uint stride_row       = p.nb01;
+    const uint stride_channel   = p.nb02;
+    const uint stride_sample    = p.nb03;
+
+    uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
+    uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
+    uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
+
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
+
+    uint32_t num_partials = p.param3;
+    for (uint32_t i = gl_SubgroupInvocationID; i < num_partials; i += gl_SubgroupSize) {
+        sum += partial_sums[i];
+    }
+    sum = subgroupAdd(sum);
+
+    uint col = tid;
+    if (col >= ncols) {
+        return;
+    }
+
+    const FLOAT_TYPE mean = sum / FLOAT_TYPE(ncols);
+    const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
+
+    if (do_multiply) {
+        if (ncols > p.ne10) {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+        } else {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+        }
+    } else {
+        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
+    }
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp

@@ -20,6 +20,7 @@ layout (push_constant) uniform parameter
     float m1;
     uint n_head_log2;
     uint nrows_x;
+    uint has_sinks;
 } p;
 
 #include "types.comp"
@@ -29,7 +30,8 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
-layout (binding = 2) buffer D {D_TYPE data_d[];};
+layout (binding = 2) readonly buffer Z {float data_c[];};
+layout (binding = 3) buffer D {D_TYPE data_d[];};
 
 shared FLOAT_TYPE vals[BLOCK_SIZE];
 
@@ -60,13 +62,13 @@ void soft_max(uint num_iters) {
         const uint h = (rowx / p.ne01) % p.ne02; // head index
 
         const float base = h < p.n_head_log2 ? p.m0 : p.m1;
-        const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
+        const uint   exp = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
 
         slope = pow(base, exp);
     }
 
     // Find max
-    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
+    FLOAT_TYPE max_val = p.has_sinks == 0 ? uintBitsToFloat(0xFF800000) : data_c[i02];
 
     // Cache values while we compute the max, so we don't need to read them
     // again when we're ready to compute exp(x-max).
@@ -148,6 +150,10 @@ void soft_max(uint num_iters) {
     }
     sum = vals[0];
 
+    if (p.has_sinks != 0) {
+        sum += FLOAT_TYPE(exp(FLOAT_TYPE(data_c[i02]) - max_val));
+    }
+
     FLOAT_TYPE rcpdivisor = 1.0/sum;
 
     [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp

@@ -0,0 +1,17 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sqrt(val));
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp

@@ -1,9 +1,9 @@
 #version 450
 
-#include "generic_head.comp"
 #include "types.comp"
 
 #extension GL_EXT_control_flow_attributes : enable
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -11,16 +11,49 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
+    const float weight = p.weight;
+
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
+
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
 
-    tmp[col] = FLOAT_TYPE(0.0f);
+    tmp[col] = FLOAT_TYPE(0.0);
 
-    for (uint i = col; i < p.KX; i += BLOCK_SIZE) {
-        tmp[col] += FLOAT_TYPE(data_a[row*p.KX + i]);
+    for (uint i = col; i < p.n_cols; i += BLOCK_SIZE) {
+        tmp[col] += FLOAT_TYPE(data_a[src_idx + i]);
     }
 
     barrier();
@@ -32,6 +65,6 @@ void main() {
     }
 
     if (col == 0) {
-        data_d[row] = D_TYPE(tmp[0]);
+        data_d[dst_idx] = D_TYPE(tmp[0] * weight);
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp

@@ -0,0 +1,14 @@
+#version 450
+
+#include "glu_head.comp"
+
+float op(float a, float b) {
+    float xi = min(a, p.limit);
+    float gi = max(min(b, p.limit), -p.limit);
+
+    float out_glu = xi / (1.0f + exp(-xi * p.alpha));
+    out_glu = out_glu * (1.0f + gi);
+    return out_glu;
+}
+
+#include "glu_main.comp"

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -1337,6 +1337,29 @@ struct block_iq4_nl_packed16
 #define A_TYPE_PACKED16 block_iq4_nl_packed16
 #endif
 
+#define QUANT_K_MXFP4 32
+#define QUANT_R_MXFP4 2
+
+struct block_mxfp4
+{
+    uint8_t e;
+    uint8_t qs[QUANT_K_MXFP4/2];
+};
+
+//struct block_mxfp4_packed16
+//{
+//    uint8_t e;
+//    uint16_t qs[QUANT_K_MXFP4/2/2];
+//};
+
+#if defined(DATA_A_MXFP4)
+#define QUANT_K QUANT_K_MXFP4
+#define QUANT_R QUANT_R_MXFP4
+#define QUANT_AUXF 1
+#define A_TYPE block_mxfp4
+//#define A_TYPE_PACKED16 block_mxfp4_packed16
+#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),
@@ -1356,6 +1379,25 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+const FLOAT_TYPE kvalues_mxfp4_const[16] = {
+    FLOAT_TYPE(0.0f), FLOAT_TYPE(0.5f), FLOAT_TYPE(1.0f), FLOAT_TYPE(1.5f), FLOAT_TYPE(2.0f), FLOAT_TYPE(3.0f), FLOAT_TYPE(4.0f), FLOAT_TYPE(6.0f),
+    FLOAT_TYPE(-0.0f), FLOAT_TYPE(-0.5f), FLOAT_TYPE(-1.0f), FLOAT_TYPE(-1.5f), FLOAT_TYPE(-2.0f), FLOAT_TYPE(-3.0f), FLOAT_TYPE(-4.0f), FLOAT_TYPE(-6.0f)
+};
+
+shared FLOAT_TYPE kvalues_mxfp4[16];
+
+#define NEEDS_INIT_IQ_SHMEM
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < kvalues_mxfp4.length(); i += wgsize.x) {
+        kvalues_mxfp4[i] = kvalues_mxfp4_const[i];
+    }
+    barrier();
+}
+#endif
+
 // returns the bfloat value in the low 16b.
 // See ggml_compute_fp32_to_bf16
 uint32_t fp32_to_bf16(float f)
@@ -1370,4 +1412,17 @@ float bf16_to_fp32(uint32_t u)
     return uintBitsToFloat(u << 16);
 }
 
+float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = x;
+        bits = bits << 23;
+    }
+
+    return uintBitsToFloat(bits);
+}
+
 #endif // !defined(GGML_TYPES_COMP)

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp

@@ -0,0 +1,25 @@
+#ifndef UTILS_COMP
+#define UTILS_COMP
+
+// 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, uint ne00, uint ne01, uint ne02, uint ne03) {
+    i03 = fastdiv(idx, (ne02*ne01*ne00));
+    const uint i03_offset = i03 * ne02*ne01*ne00;
+    i02 = fastdiv((idx - i03_offset), (ne01*ne00));
+    const uint i02_offset = i02*ne01*ne00;
+    i01 = (idx - i03_offset - i02_offset) / ne00;
+    i00 = idx - i03_offset - i02_offset - i01*ne00;
+}
+
+#endif // UTILS_COMP

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -64,9 +64,16 @@ const std::vector<std::string> type_names = {
     "iq3_s",
     "iq4_xs",
     "iq4_nl",
+    "mxfp4",
     "bf16",
 };
 
+enum MatMulIdType {
+    NONE,
+    DEFAULT,
+    SUBGROUP,
+};
+
 namespace {
 void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
 #ifdef _WIN32
@@ -118,7 +125,7 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
     CloseHandle(pi.hProcess);
     CloseHandle(pi.hThread);
 #else
-int stdout_pipe[2];
+    int stdout_pipe[2];
     int stderr_pipe[2];
 
     if (pipe(stdout_pipe) != 0 || pipe(stderr_pipe) != 0) {
@@ -222,7 +229,8 @@ void string_to_spv_func(const std::string& _name, const std::string& in_fname, c
     std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
 
     // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
-    std::string opt_level = coopmat ? "" : "-O";
+    // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
+    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos) ? "" : "-O";
 
     #ifdef _WIN32
         std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
@@ -291,7 +299,7 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
     compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat, coopmat2, f16acc));
 }
 
-void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool f16acc) {
+void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc) {
     std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
     std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
     std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
@@ -301,9 +309,13 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
     };
     std::string shader_name = "matmul";
 
-    if (matmul_id) {
+    if (matmul_id_type == MatMulIdType::DEFAULT) {
         base_dict["MUL_MAT_ID"] = "1";
         shader_name = "matmul_id";
+    } else if (matmul_id_type == MatMulIdType::SUBGROUP) {
+        base_dict["MUL_MAT_ID"] = "1";
+        base_dict["MUL_MAT_ID_USE_SUBGROUPS"] = "1";
+        shader_name = "matmul_id_subgroup";
     }
 
     if (fp16) {
@@ -362,7 +374,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         std::string load_vec_quant = "2";
         if ((tname == "q4_0") || (tname == "q4_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
             load_vec_quant = "8";
-        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl"))
+        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
             load_vec_quant = "4";
 
         if (tname == "bf16") {
@@ -387,7 +399,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
+        if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
             string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
@@ -399,26 +411,28 @@ void process_shaders() {
     std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
 
     // matmul
-    for (const auto& matmul_id : {false, true}) {
+    for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
         // No coopmats
         // fp32
-        matmul_shaders(false, matmul_id, false, false, false);
+        matmul_shaders(false, matmul_id_type, false, false, false);
 
         // fp16, fp32acc and fp16acc
-        matmul_shaders(true, matmul_id, false, false, false);
-        matmul_shaders(true, matmul_id, false, false, true);
+        matmul_shaders(true, matmul_id_type, false, false, false);
+        matmul_shaders(true, matmul_id_type, false, false, true);
 
+        if (matmul_id_type != MatMulIdType::DEFAULT) {
 #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
-        // Coopmat, fp32acc and fp16acc
-        matmul_shaders(true, matmul_id, true, false, false);
-        matmul_shaders(true, matmul_id, true, false, true);
+            // Coopmat, fp32acc and fp16acc
+            matmul_shaders(true, matmul_id_type, true, false, false);
+            matmul_shaders(true, matmul_id_type, true, false, true);
 #endif
 
 #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-        // Coopmat2, fp32acc and fp16acc
-        matmul_shaders(true, matmul_id, false, true, false);
-        matmul_shaders(true, matmul_id, false, true, true);
+            // Coopmat2, fp32acc and fp16acc
+            matmul_shaders(true, matmul_id_type, false, true, false);
+            matmul_shaders(true, matmul_id_type, false, true, true);
 #endif
+        }
     }
 
     // flash attention
@@ -471,6 +485,9 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
 
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
@@ -498,6 +515,7 @@ void process_shaders() {
     string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
@@ -533,13 +551,15 @@ void process_shaders() {
         s += std::string(dst_f16 ? "_f16" : "_f32");
         return s;
     };
-    for (std::string op : {"add", "sub", "mul", "div"}) {
+    for (std::string op : {"add", "sub", "mul", "div", "add_rms", }) {
     for (auto src0_f16 : {false, true}) {
     for (auto src1_f16 : {false, true}) {
     for (auto dst_f16  : {false, true}) {
     for (auto rte      : {false, true}) {
+        auto source = op == "add_rms" ? std::string("add") : op;
         auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
-        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
+        auto add_rms = op == "add_rms" ? "1" : "0";
+        string_to_spv(name.c_str(), source + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}, {"ADD_RMS" , add_rms}});
     }
     }
     }
@@ -565,6 +585,8 @@ void process_shaders() {
 
     string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
+    string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+
     string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -579,6 +601,8 @@ void process_shaders() {
 
     string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
+    string_to_spv("exp_f16",        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("exp_f32",        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_erf_f16",   "gelu_erf.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
@@ -602,6 +626,8 @@ void process_shaders() {
         string_to_spv("reglu_f32" + suffix,      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("swiglu_f16" + suffix,     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("swiglu_f32" + suffix,     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_oai_f16" + suffix, "swiglu_oai.comp",  {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_oai_f32" + suffix, "swiglu_oai.comp",  {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_erf_f16" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_erf_f32" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
@@ -654,14 +680,31 @@ void process_shaders() {
     string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+    string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+
+    string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+    string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
-    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+#endif
 
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
+    string_to_spv("conv2d_dw_whcn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
+    string_to_spv("conv2d_dw_cwhn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 
     string_to_spv("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "0"}});
+    string_to_spv("multi_add_rms_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "1"}});
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -718,7 +761,7 @@ void write_output_files() {
     }
 
     std::string suffixes[2] = {"_f32", "_f16"};
-    for (const char *op : {"add", "sub", "mul", "div"}) {
+    for (const char *op : {"add", "sub", "mul", "div", "add_rms"}) {
         fprintf(hdr, "extern unsigned char *%s_data[2][2][2][2];\n", op);
         fprintf(hdr, "extern uint64_t %s_len[2][2][2][2];\n", op);
         std::string data = "unsigned char *" + std::string(op) + "_data[2][2][2][2] = ";
@@ -770,6 +813,18 @@ void write_output_files() {
         fputs(data.c_str(), src);
         fputs(len.c_str(), src);
     }
+
+    for (const std::string& btype : {"f16", "f32"}) {
+    for (const auto& tname : type_names) {
+        fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[2];\n", tname.c_str(), btype.c_str());
+        fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[2];\n", tname.c_str(), btype.c_str());
+        std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[2] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data};\n";
+        std::string len =  "uint64_t arr_dmmv_"       + tname + "_" + btype + "_f32_len[2] =  {mul_mat_vec_" + tname + "_" + btype + "_f32_len,  mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len};\n";
+        fputs(data.c_str(), src);
+        fputs(len.c_str(), src);
+    }
+    }
+
     fclose(hdr);
     fclose(src);
 }

package/cpp/llama.cpp/ggml/src/ggml-webgpu/CMakeLists.txt

@@ -20,8 +20,8 @@ add_custom_command(
     COMMAND ${CMAKE_COMMAND} -E make_directory ${SHADER_OUTPUT_DIR}
     COMMAND ${CMAKE_COMMAND} -E env PYTHONIOENCODING=utf-8
         ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
-            --input "${SHADER_DIR}"
-            --output "${SHADER_HEADER}"
+            --input_dir "${SHADER_DIR}"
+            --output_file "${SHADER_HEADER}"
     DEPENDS ${WGSL_SHADER_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
     VERBATIM
 )