feat: add custom modulate operator and ConvRot fusion (#2258)

CMakeLists.txt

@@ -378,7 +378,7 @@ if (NOT SD_USE_UPSTREAM_GGML AND (SD_CUDA OR SD_HIPBLAS OR SD_MUSA))
 endif()
 if (SD_VULKAN AND NOT SD_USE_UPSTREAM_GGML)
     add_subdirectory(src/core/vulkan)
-    list(APPEND SD_BACKEND_REGISTRATIONS sd_ggml_register_vulkan_rope sd_ggml_register_vulkan_i8)
+    list(APPEND SD_BACKEND_REGISTRATIONS sd_ggml_register_vulkan_rope sd_ggml_register_vulkan_i8 sd_ggml_register_vulkan_modulate)
 endif()
 if(SD_BACKEND_REGISTRATIONS)
     string(REPLACE ";" ",\n    " SD_BACKEND_REGISTRATIONS "${SD_BACKEND_REGISTRATIONS}")

cmake/backend_extensions.h.in

@@ -5,6 +5,7 @@
 
 void sd_ggml_register_vulkan_rope(ggml_backend_dev_t device);
 void sd_ggml_register_vulkan_i8(ggml_backend_dev_t device);
+void sd_ggml_register_vulkan_modulate(ggml_backend_dev_t device);
 
 static void (*const sd_backend_extensions[])(ggml_backend_dev_t) = {
     @SD_BACKEND_REGISTRATIONS@

docs/ggml_extensions.md

@@ -135,6 +135,25 @@ CPU kernels and fusion rules live in `src/core/cpu/`, with registration collecte
 in `cpu/registration.cpp`. CUDA-style and Vulkan implementations live in their
 respective backend directories.
 
+`ggml_ext_modulate(ctx, x, shift, scale)` emits `sd.modulate.v1` for F32 inputs
+when the runner context uses CPU, CUDA/HIP/MUSA, or Vulkan.
+Its explicit inputs are `x`, `scale`, and optional `shift`; it computes
+`(x + x * scale) + shift` with separate FP32 multiply/add rounding. The output
+is contiguous and out of place. Inputs may be strided and use ggml's
+dimension-wise repetition rules. Callers prepare token, batch, or frame layouts
+before constructing the operation; normalization and residual gates remain
+separate operations. CPU, CUDA/HIP/MUSA, and Vulkan have standalone kernels.
+Other backends (including Metal, SYCL, and OpenCL), upstream ggml builds, and
+other input types retain the original MUL/ADD graph.
+
+ConvRot fusion matches the Modulate node directly, including scale-only
+modulation, instead of recognizing its former MUL/ADD/ADD decomposition.
+CUDA accepts contiguous per-channel parameters for `Modulate -> ConvRot` and
+`RMSNorm -> Mul(weight) -> Modulate -> ConvRot`. Vulkan limits the first fusion
+to widths 4096/6144/5376 and the RMSNorm fusion to width 5376.
+Other layouts execute the standalone operation followed by ConvRot. The
+gated-residual MUL/ADD fusion remains available for residual branches.
+
 `src/core/cpu/rope.cpp` registers the CPU implementation of
 `sd.rope_apply.v1` and a
 `RMSNorm -> Mul -> RoPE` fusion. The fusion accepts a one-dimensional weight
@@ -269,8 +288,9 @@ unsupported layouts retain ordinary operator execution. Fusion and force-cuBLAS
 disable settings are honored.
 
 `src/core/cuda/convrot.cu` registers CUDA fusion for `Scale -> ConvRot`,
-`Norm -> Mul -> ConvRot`, `RMSNorm -> Mul(weight) -> Mul(modulation) -> Add ->
+`Norm -> Mul -> ConvRot`, `Modulate -> ConvRot`,
-Add(shift) -> ConvRot`, `SwiGLU -> ConvRot`, and `SiLU/Sigmoid -> Mul -> ConvRot`.
+`RMSNorm -> Mul(weight) -> Modulate -> ConvRot`, `SwiGLU -> ConvRot`, and
+`SiLU/Sigmoid -> Mul -> ConvRot`.
 ConvRot outputs are custom nodes. SwiGLU accepts packed,
 swapped, or separate gate/up inputs. Independent modulation and up-branch
 producers are moved ahead of their consumers through the safe graph-delay API.

src/core/cpu/modulate.cpp

@@ -0,0 +1,82 @@
+#include "core/cpu/registration.h"
+#include "core/ggml_extensions.h"
+
+#ifndef SD_USE_UPSTREAM_GGML
+#include <cstring>
+
+namespace sd_cpu {
+    static bool modulate_supports(ggml_backend_dev_t, const ggml_tensor* node, void*) {
+        return sd_ggml_modulate_describe(node);
+    }
+
+    static float modulate_value(const ggml_tensor* t, const int64_t* index) {
+        size_t offset = 0;
+        for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+            offset += (index[d] % t->ne[d]) * t->nb[d];
+        }
+        return *reinterpret_cast<const float*>(static_cast<const char*>(t->data) + offset);
+    }
+
+    static void modulate_compute(ggml_tensor* dst, int ith, int nth, void*) {
+        auto* output       = static_cast<float*>(dst->data);
+        const int64_t rows = ggml_nrows(dst);
+        for (int64_t row = ith; row < rows; row += nth) {
+            int64_t index[] = {0, row % dst->ne[1], (row / dst->ne[1]) % dst->ne[2], row / (dst->ne[1] * dst->ne[2])};
+            for (int64_t col = 0; col < dst->ne[0]; ++col) {
+                index[0]      = col;
+                const float x = modulate_value(dst->src[0], index);
+                // Match separate ggml MUL/ADD rounding even with contraction enabled.
+                volatile float product         = x * modulate_value(dst->src[1], index);
+                volatile float scaled          = x + product;
+                output[row * dst->ne[0] + col] = dst->src[2] ? scaled + modulate_value(dst->src[2], index) : scaled;
+            }
+        }
+    }
+
+    static ggml_status modulate_prepare(ggml_backend_t, ggml_tensor* const* nodes, int, void**, size_t* bytes, void*) {
+        *bytes = sd_ggml_modulate_workspace_size(nodes[0]);
+        return GGML_STATUS_SUCCESS;
+    }
+
+    static void modulate_destroy(void*, void*) {}
+
+    static ggml_status modulate_execute(const ggml_ext_execution_context* context, ggml_tensor* const* nodes, int, void*, void*) {
+        auto reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(context->backend));
+        auto run = reinterpret_cast<ggml_ext_cpu_parallel_fn>(ggml_backend_reg_get_proc_address(reg, "ggml_ext_cpu_parallel_v1"));
+        GGML_ASSERT(run);
+        auto* dst          = nodes[0];
+        const auto address = reinterpret_cast<uintptr_t>(dst->data);
+        bool alias         = false;
+        for (const auto* input : dst->src) {
+            if (input) {
+                const auto src = reinterpret_cast<uintptr_t>(input->data);
+                alias |= address < src + ggml_nbytes(input) && src < address + ggml_nbytes(dst);
+            }
+        }
+        if (!alias) {
+            return run(context->backend, dst, modulate_compute, nullptr);
+        }
+        ggml_tensor temporary = *dst;
+        temporary.data        = ggml_backend_buffer_get_base(context->workspace);
+        const auto status     = run(context->backend, &temporary, modulate_compute, nullptr);
+        if (status == GGML_STATUS_SUCCESS) {
+            std::memcpy(dst->data, temporary.data, ggml_nbytes(dst));
+        }
+        return status;
+    }
+
+    void register_modulate(ggml_backend_dev_t device) {
+        ggml_ext_kernel_desc desc = {};
+        desc.abi_version          = GGML_EXT_ABI_VERSION;
+        desc.name                 = "sd.modulate.cpu.v1";
+        desc.op                   = GGML_OP_CUSTOM;
+        desc.custom_op            = sd_ggml_modulate_op();
+        desc.priority             = 100;
+        desc.supports             = modulate_supports;
+        desc.prepare              = modulate_prepare;
+        desc.destroy              = modulate_destroy;
+        desc.execute              = modulate_execute;
+        GGML_ASSERT(ggml_ext_register_kernel(device, &desc));
+    }
+}
+#endif

src/core/cpu/registration.cpp

@@ -6,6 +6,7 @@ namespace sd_cpu {
     void register_extensions(ggml_backend_dev_t device) {
         register_i8(device);
         register_rope(device);
+        register_modulate(device);
     }
 
 }

src/core/cpu/registration.h

@@ -7,6 +7,7 @@
 namespace sd_cpu {
     void register_extensions(ggml_backend_dev_t device);
     void register_rope(ggml_backend_dev_t device);
+    void register_modulate(ggml_backend_dev_t device);
     void register_i8(ggml_backend_dev_t device);
 }
 #endif

src/core/cuda/CMakeLists.txt

@@ -2,7 +2,7 @@ cmake_policy(SET CMP0079 NEW)
 
 function(sd_cuda_sources VENDOR LANGUAGE)
     set(SOURCES)
-    foreach(OP rope int8 convrot affine registration)
+    foreach(OP rope int8 convrot affine modulate registration)
         set(SOURCE "${CMAKE_CURRENT_BINARY_DIR}/${OP}_${VENDOR}.cu")
         string(TOUPPER "${VENDOR}" UPPER_VENDOR)
         file(GENERATE OUTPUT "${SOURCE}" CONTENT "#define SD_CUDA_VENDOR_${UPPER_VENDOR} 1\n#include \"core/cuda/${OP}.cu\"\n")

src/core/cuda/convrot.cu

@@ -27,6 +27,7 @@ namespace SD_CUDA_NAMESPACE {
         }
         enum Kind { SCALE,
                     NORM,
+                    MOD,
                     RMS_MOD,
                     SWIGLU,
                     SILU,
@@ -65,16 +66,20 @@ namespace SD_CUDA_NAMESPACE {
                 op.eps        = param_f32(first, 0);
                 if (op.k < 1024 || !vector(op.modulation, op.k))
                     return false;
-            } else if (count == 6 && first->op == GGML_OP_RMS_NORM && nodes[1]->op == GGML_OP_MUL && nodes[1]->src[0] == first &&
+            } else if (count == 2 && sd_ggml_modulate_describe(first) && value == first) {
-                       nodes[2]->op == GGML_OP_MUL && nodes[2]->src[0] == nodes[1] &&
+                op.kind       = MOD;
-                       nodes[3]->op == GGML_OP_ADD && nodes[3]->src[0] == nodes[1] && nodes[3]->src[1] == nodes[2] &&
+                op.modulation = first->src[1];
-                       nodes[4]->op == GGML_OP_ADD && nodes[4]->src[0] == nodes[3] && value == nodes[4]) {
+                op.shift      = first->src[2];
+                if (!vector(op.modulation, op.k) || (op.shift && !vector(op.shift, op.k)))
+                    return false;
+            } else if (count == 4 && first->op == GGML_OP_RMS_NORM && nodes[1]->op == GGML_OP_MUL && nodes[1]->src[0] == first &&
+                       sd_ggml_modulate_describe(nodes[2]) && nodes[2]->src[0] == nodes[1] && value == nodes[2]) {
                 op.kind       = RMS_MOD;
                 op.weight     = nodes[1]->src[1];
                 op.modulation = nodes[2]->src[1];
-                op.shift      = nodes[4]->src[1];
+                op.shift      = nodes[2]->src[2];
                 op.eps        = param_f32(first, 0);
-                if (op.k < 1024 || !vector(op.weight, op.k) || !vector(op.modulation, op.k) || !vector(op.shift, op.k))
+                if (op.k < 1024 || !vector(op.weight, op.k) || !vector(op.modulation, op.k) || (op.shift && !vector(op.shift, op.k)))
                     return false;
             } else if (count == 2 && first->op == GGML_OP_GLU && ggml_get_glu_op(first) == GGML_GLU_OP_SWIGLU && value == first) {
                 op.kind = SWIGLU;
@@ -119,7 +124,7 @@ namespace SD_CUDA_NAMESPACE {
             auto* g             = const_cast<ggml_cgraph*>(graph);
             const int available = ggml_graph_n_nodes(g) - first;
             const auto* node    = ggml_graph_node(g, first);
-            const int count     = node->op == GGML_OP_RMS_NORM ? 6 : node->op == GGML_OP_NORM || node->op == GGML_OP_UNARY ? 3
+            const int count     = node->op == GGML_OP_RMS_NORM ? 4 : node->op == GGML_OP_NORM || node->op == GGML_OP_UNARY ? 3
                                                                                                                            : 2;
             if (available < count)
                 return 0;
@@ -132,7 +137,7 @@ namespace SD_CUDA_NAMESPACE {
         static void optimize(ggml_backend_dev_t device, ggml_cgraph* graph, void*) {
             if (fusion_disabled())
                 return;
-            for (int i = 2; i < ggml_graph_n_nodes(graph); ++i) {
+            for (int i = 1; i < ggml_graph_n_nodes(graph); ++i) {
                 auto* quant = ggml_graph_node(graph, i);
                 int group;
                 if (!sd_ggml_convrot_describe(quant, &group) || group != 256)
@@ -140,17 +145,18 @@ namespace SD_CUDA_NAMESPACE {
                 ggml_tensor* chain[6] = {};
                 int count             = 0;
                 auto* value           = quant->src[0];
-                if (value->op == GGML_OP_ADD && value->src[0]->op == GGML_OP_ADD) {
+                if (sd_ggml_modulate_describe(value)) {
-                    auto* add    = value->src[0];
+                    auto* weight = value->src[0];
-                    auto* weight = add->src[0];
+                    if (weight->op == GGML_OP_MUL && weight->src[0]->op == GGML_OP_RMS_NORM) {
-                    if (weight->op == GGML_OP_MUL) {
                         chain[0] = weight->src[0];
                         chain[1] = weight;
-                        chain[2] = add->src[1];
+                        chain[2] = value;
-                        chain[3] = add;
+                        chain[3] = quant;
-                        chain[4] = value;
+                        count    = 4;
-                        chain[5] = quant;
+                    } else {
-                        count    = 6;
+                        chain[0] = value;
+                        chain[1] = quant;
+                        count    = 2;
                     }
                 } else if (value->op == GGML_OP_MUL) {
                     for (int src = 0; src < 2; ++src)
@@ -210,13 +216,13 @@ namespace SD_CUDA_NAMESPACE {
             const auto x = reinterpret_cast<uintptr_t>(a->data), y = reinterpret_cast<uintptr_t>(b->data);
             return x <= y ? y - x < ggml_nbytes(a) : x - y < ggml_nbytes(b);
         }
-        template <bool Swiglu = false, bool Normalize = false, bool Scaled = false, bool Sigmoid = false, bool RMS = false>
+        template <bool Swiglu = false, bool Normalize = false, bool Scaled = false, bool Sigmoid = false, bool RMS = false, bool Modulate = false>
         static CUDAError launch(Plan& p, const Operation& op, int8_t* output, CUDAStream stream) {
-            constexpr bool warp = Scaled || RMS;
+            constexpr bool warp = Scaled || RMS || Modulate;
             constexpr int extra = Normalize ? 384 : RMS ? 256
                                                         : 128;
             if (!p.configured) {
-                auto status = cuda_set_shared_memory(quantize<Swiglu, false, Normalize, Scaled, warp, float, Sigmoid, RMS>,
+                auto status = cuda_set_shared_memory(quantize<Swiglu, false, Normalize, Scaled, warp, float, Sigmoid, RMS, Modulate>,
                                                      int(p.device.shared_memory - extra));
                 if (status != cuda_success)
                     return status;
@@ -229,7 +235,7 @@ namespace SD_CUDA_NAMESPACE {
                                                                                         : 256;
             const int64_t padded = (op.rows + 3) / 4 * 4;
             auto data            = [](const ggml_tensor* t) { return t ? static_cast<const float*>(t->data) : nullptr; };
-            quantize<Swiglu, false, Normalize, Scaled, warp, float, Sigmoid, RMS><<<unsigned(op.rows), threads, size_t(op.k) * sizeof(float), stream>>>(
+            quantize<Swiglu, false, Normalize, Scaled, warp, float, Sigmoid, RMS, Modulate><<<unsigned(op.rows), threads, size_t(op.k) * sizeof(float), stream>>>(
                 data(op.input) + op.gate_offset, output, reinterpret_cast<float*>(output + op.k * padded), op.k, op.rows, op.k,
                 op.up ? data(op.up) + op.up_offset : nullptr, op.stride, op.up_stride, nullptr, nullptr, data(op.modulation), op.eps, op.scale, op.bias, data(op.weight), data(op.shift));
             return cuda_last_error();
@@ -260,6 +266,9 @@ namespace SD_CUDA_NAMESPACE {
                 case RMS_MOD:
                     error = launch<false, false, false, false, true>(plan, op, output, stream);
                     break;
+                case MOD:
+                    error = launch<false, false, false, false, false, true>(plan, op, output, stream);
+                    break;
                 case SWIGLU:
                 case SILU:
                     error = launch<true>(plan, op, output, stream);
@@ -277,12 +286,13 @@ namespace SD_CUDA_NAMESPACE {
     }
 
     void register_convrot(ggml_backend_dev_t device) {
-        for (auto op : {GGML_OP_SCALE, GGML_OP_NORM, GGML_OP_RMS_NORM, GGML_OP_GLU, GGML_OP_UNARY}) {
+        for (auto op : {GGML_OP_SCALE, GGML_OP_NORM, GGML_OP_RMS_NORM, GGML_OP_GLU, GGML_OP_UNARY, GGML_OP_CUSTOM}) {
             ggml_ext_kernel_desc desc = {};
             desc.abi_version          = GGML_EXT_ABI_VERSION;
             const std::string name    = std::string("sd.convrot_fusion.") + ggml_op_name(op) + "." SD_CUDA_NAME ".v1";
             desc.name                 = name.c_str();
             desc.op                   = op;
+            desc.custom_op            = op == GGML_OP_CUSTOM ? sd_ggml_modulate_op() : nullptr;
             desc.priority             = 225;
             desc.optimize             = op == GGML_OP_SCALE ? sd_convrot::optimize : nullptr;
             desc.match                = sd_convrot::match;

src/core/cuda/convrot_kernel.h

@@ -40,7 +40,7 @@ namespace SD_CUDA_NAMESPACE {
             return __fadd_rn(__fadd_rn(__fadd_rn(a, b), c), d);
         }
 
-        template <bool swiglu, bool dequantize = false, bool normalize = false, bool scaled = false, bool warp_rotation = false, typename T = float, bool sigmoid_gate = false, bool rms_modulation = false>
+        template <bool swiglu, bool dequantize = false, bool normalize = false, bool scaled = false, bool warp_rotation = false, typename T = float, bool sigmoid_gate = false, bool rms_modulation = false, bool modulate = false>
         __global__ void quantize(
             const T* src,
             int8_t* dst,
@@ -103,7 +103,12 @@ namespace SD_CUDA_NAMESPACE {
                             float value = src_row[i];
                             if constexpr (rms_modulation) {
                                 value = __fmul_rn(__fmul_rn(value, rms_scale), rms_weight[i]);
-                                value = __fadd_rn(__fadd_rn(value, __fmul_rn(value, modulation[i])), rms_shift[i]);
+                            }
+                            if constexpr (rms_modulation || modulate) {
+                                value = __fadd_rn(value, __fmul_rn(value, modulation[i]));
+                                if (rms_shift) {
+                                    value = __fadd_rn(value, rms_shift[i]);
+                                }
                             }
                             if constexpr (scaled) {
                                 value = __fmaf_rn(value, input_scale, input_bias);

src/core/cuda/modulate.cu

@@ -0,0 +1,95 @@
+#include <algorithm>
+#include "core/ggml_extensions.h"
+#include "runtime.h"
+
+namespace SD_CUDA_NAMESPACE {
+    namespace sd_modulate {
+        struct Parameters {
+            int64_t ne[4], scale_ne[4], shift_ne[4];
+            size_t x_stride[4], scale_stride[4], shift_stride[4];
+            int64_t count;
+        };
+
+        static __global__ void modulate(const float* x, const float* scale, const float* shift, float* dst, Parameters p) {
+            for (int64_t i = int64_t(blockIdx.x) * blockDim.x + threadIdx.x; i < p.count; i += int64_t(blockDim.x) * gridDim.x) {
+                int64_t index = i;
+                size_t xi = 0, si = 0, hi = 0;
+#pragma unroll
+                for (int d = 0; d < 4; ++d) {
+                    const int64_t coord = index % p.ne[d];
+                    index /= p.ne[d];
+                    xi += coord * p.x_stride[d];
+                    si += (coord % p.scale_ne[d]) * p.scale_stride[d];
+                    if (shift) {
+                        hi += (coord % p.shift_ne[d]) * p.shift_stride[d];
+                    }
+                }
+                float value = __fadd_rn(x[xi], __fmul_rn(x[xi], scale[si]));
+                dst[i]      = shift ? __fadd_rn(value, shift[hi]) : value;
+            }
+        }
+
+        static bool supports(ggml_backend_dev_t, const ggml_tensor* node, void*) {
+            return sd_ggml_modulate_describe(node);
+        }
+
+        static ggml_status prepare(ggml_backend_t, ggml_tensor* const* nodes, int, void**, size_t* bytes, void*) {
+            *bytes = sd_ggml_modulate_workspace_size(nodes[0]);
+            return GGML_STATUS_SUCCESS;
+        }
+
+        static void destroy(void*, void*) {}
+
+        static ggml_status execute(const ggml_ext_execution_context* e, ggml_tensor* const* nodes, int, void*, void*) {
+            ggml_ext_gpu_context context;
+            if (!cuda_context(e->backend, &context)) {
+                return GGML_STATUS_FAILED;
+            }
+            auto* dst          = nodes[0];
+            const auto* x      = dst->src[0];
+            const auto* scale  = dst->src[1];
+            const auto* shift  = dst->src[2];
+            Parameters p       = {};
+            p.count            = ggml_nelements(dst);
+            bool alias         = false;
+            const auto address = reinterpret_cast<uintptr_t>(dst->data);
+            for (const auto* input : {x, scale, shift}) {
+                if (input) {
+                    const auto src = reinterpret_cast<uintptr_t>(input->data);
+                    alias |= address < src + ggml_nbytes(input) && src < address + ggml_nbytes(dst);
+                }
+            }
+            for (int d = 0; d < 4; ++d) {
+                p.ne[d]           = x->ne[d];
+                p.scale_ne[d]     = scale->ne[d];
+                p.shift_ne[d]     = shift ? shift->ne[d] : 1;
+                p.x_stride[d]     = x->nb[d] / sizeof(float);
+                p.scale_stride[d] = scale->nb[d] / sizeof(float);
+                p.shift_stride[d] = shift ? shift->nb[d] / sizeof(float) : 0;
+            }
+            auto stream           = static_cast<CUDAStream>(context.stream);
+            auto* output          = static_cast<float*>(alias ? ggml_backend_buffer_get_base(e->workspace) : dst->data);
+            const unsigned blocks = unsigned(std::min<int64_t>((p.count + 255) / 256, 65535));
+            modulate<<<blocks, 256, 0, stream>>>(static_cast<const float*>(x->data), static_cast<const float*>(scale->data),
+                                                 shift ? static_cast<const float*>(shift->data) : nullptr, output, p);
+            if (cuda_last_error() != cuda_success || (alias && cuda_copy_async(dst->data, output, ggml_nbytes(dst), stream) != cuda_success)) {
+                return GGML_STATUS_FAILED;
+            }
+            return GGML_STATUS_SUCCESS;
+        }
+    }
+
+    void register_modulate(ggml_backend_dev_t device) {
+        ggml_ext_kernel_desc desc = {};
+        desc.abi_version          = GGML_EXT_ABI_VERSION;
+        desc.name                 = "sd.modulate." SD_CUDA_NAME ".v1";
+        desc.op                   = GGML_OP_CUSTOM;
+        desc.custom_op            = sd_ggml_modulate_op();
+        desc.priority             = 100;
+        desc.supports             = sd_modulate::supports;
+        desc.prepare              = sd_modulate::prepare;
+        desc.destroy              = sd_modulate::destroy;
+        desc.execute              = sd_modulate::execute;
+        GGML_ASSERT(ggml_ext_register_kernel_ex(device, &desc, GGML_EXT_KERNEL_CAPTURE_SAFE | GGML_EXT_KERNEL_REPLAY_REUSABLE));
+    }
+}

src/core/cuda/registration.cu

@@ -11,6 +11,7 @@ namespace SD_CUDA_NAMESPACE {
     void register_i8(ggml_backend_dev_t device);
     void register_convrot(ggml_backend_dev_t device);
     void register_affine(ggml_backend_dev_t device);
+    void register_modulate(ggml_backend_dev_t device);
 #if !defined(SD_CUDA_VENDOR_HIP) && !defined(SD_CUDA_VENDOR_MUSA)
     void register_sage_attention(ggml_backend_dev_t device);
     void register_sol_attention(ggml_backend_dev_t device);
@@ -20,6 +21,7 @@ namespace SD_CUDA_NAMESPACE {
         if (cuda_device_info(device).device < 0)
             return;
         register_affine(device);
+        register_modulate(device);
         register_i8(device);
         register_convrot(device);
         register_rope(device);

src/core/ggml_extend.cpp

@@ -6,6 +6,7 @@
 #include <vector>
 
 #include "core/ggml_extend_backend.h"
+#include "core/ggml_runner.h"
 
 ggml_tensor* ggml_ext_mul_n_mode(ggml_context* ctx, ggml_tensor* a, ggml_tensor* b, int mode) {
     // reshape A
@@ -155,6 +156,26 @@ static bool ggml_ext_is_padded_1d(const ggml_tensor* x) {
            x->nb[3] == x->nb[2] * x->ne[2];
 }
 
+ggml_tensor* ggml_ext_modulate(GGMLRunnerContext* ctx,
+                               ggml_tensor* x,
+                               ggml_tensor* shift,
+                               ggml_tensor* scale) {
+#ifndef SD_USE_UPSTREAM_GGML
+    if (x->type == GGML_TYPE_F32 && scale->type == GGML_TYPE_F32 && (!shift || shift->type == GGML_TYPE_F32) &&
+        (sd_backend_is_cpu(ctx->backend) || sd_backend_is(ctx->backend, "CUDA") ||
+         sd_backend_is(ctx->backend, "ROCm") || sd_backend_is(ctx->backend, "MUSA") || sd_backend_is(ctx->backend, "Vulkan"))) {
+        ggml_tensor* inputs[] = {x, scale, shift};
+        return ggml_ext_op_create(ctx->ggml_ctx, sd_ggml_modulate_op(), inputs, shift ? 3 : 2, nullptr);
+    }
+#endif
+    // Preserve separate multiply/add nodes for rounding and backend fusion matching.
+    x = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, x, scale));
+    if (shift != nullptr) {
+        x = ggml_add(ctx->ggml_ctx, x, shift);
+    }
+    return x;
+}
+
 ggml_tensor* ggml_ext_scale(ggml_context* ctx,
                             ggml_tensor* x,
                             float factor,

src/core/ggml_extend.h

@@ -8,6 +8,8 @@
 #include "ggml-backend.h"
 #include "ggml.h"
 
+struct GGMLRunnerContext;
+
 #define EPS 1e-05f
 
 static_assert(GGML_MAX_NAME >= 160, "GGML_MAX_NAME must be at least 160");
@@ -58,6 +60,13 @@ ggml_tensor* ggml_ext_scale(ggml_context* ctx,
                             float factor,
                             bool inplace = false);
 
+// scale and optional shift must already be broadcastable to x in ggml layout.
+// Computes (x + x * scale) + shift; F32 uses sd.modulate.v1 when supported by ctx->backend.
+ggml_tensor* ggml_ext_modulate(GGMLRunnerContext* ctx,
+                               ggml_tensor* x,
+                               ggml_tensor* shift,
+                               ggml_tensor* scale);
+
 ggml_tensor* ggml_ext_gelu(ggml_context* ctx,
                            ggml_tensor* x,
                            bool inplace = false);

src/core/ggml_extensions.cpp

@@ -15,6 +15,83 @@
 #include "sd_backend_extensions.h"
 #endif
 
+static bool modulate_infer(ggml_tensor* const* inputs, int count, const void*, ggml_ext_tensor_desc* out) {
+    if (count != 2 && count != 3) {
+        return false;
+    }
+    const auto* x = inputs[0];
+    for (int i = 0; i < count; ++i) {
+        const auto* t = inputs[i];
+        if (!t || t->type != GGML_TYPE_F32) {
+            return false;
+        }
+        for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+            if (t->ne[d] <= 0 || x->ne[d] % t->ne[d] != 0 || t->nb[d] % sizeof(float)) {
+                return false;
+            }
+        }
+    }
+    out->type = GGML_TYPE_F32;
+    std::memcpy(out->ne, x->ne, sizeof(out->ne));
+    return true;
+}
+
+ggml_ext_op_t sd_ggml_modulate_op() {
+    static const auto op = [] {
+        const ggml_ext_op_desc desc = {GGML_EXT_ABI_VERSION, "sd.modulate.v1", 0, modulate_infer};
+        auto result                 = ggml_ext_register_op(&desc);
+        GGML_ASSERT(result);
+        return result;
+    }();
+    return op;
+}
+
+bool sd_ggml_modulate_describe(const ggml_tensor* node) {
+    if (!node || ggml_ext_op_get(node) != sd_ggml_modulate_op() || !ggml_is_contiguous(node) || node->src[3]) {
+        return false;
+    }
+    ggml_ext_tensor_desc out;
+    return modulate_infer(node->src, node->src[2] ? 3 : 2, nullptr, &out) &&
+           node->type == out.type && std::memcmp(node->ne, out.ne, sizeof(out.ne)) == 0;
+}
+
+size_t sd_ggml_modulate_workspace_size(const ggml_tensor* node) {
+    // Owning custom outputs cannot reuse live input storage. Segment outputs may
+    // already be allocated, so reserve scratch only if their ranges can overlap.
+    if (!node->data && !node->view_src) {
+        return 0;
+    }
+    if (node->view_src || !node->buffer) {
+        return ggml_nbytes(node);
+    }
+    for (const auto* input : node->src) {
+        if (!input) {
+            continue;
+        }
+        const size_t bytes = ggml_nbytes(input);
+        size_t offset      = 0;
+        while (!input->data && input->view_src) {
+            offset += input->view_offs;
+            input = input->view_src;
+        }
+        if (!input->data) {
+            continue;
+        }
+        if (!input->buffer) {
+            return ggml_nbytes(node);
+        }
+        if (node->buffer != input->buffer) {
+            continue;
+        }
+        const auto dst = reinterpret_cast<uintptr_t>(node->data);
+        const auto src = reinterpret_cast<uintptr_t>(input->data) + offset;
+        if (dst < src + bytes && src < dst + ggml_nbytes(node)) {
+            return ggml_nbytes(node);
+        }
+    }
+    return 0;
+}
+
 static bool rope_infer(ggml_tensor* const* inputs, int count, const void*, ggml_ext_tensor_desc* out) {
     if (count != 2) {
         return false;

src/core/ggml_extensions.h

@@ -8,6 +8,10 @@ void sd_ggml_register_extensions(ggml_backend_t backend);
 #ifndef SD_USE_UPSTREAM_GGML
 #include "ggml-ext.h"
 
+ggml_ext_op_t sd_ggml_modulate_op();
+bool sd_ggml_modulate_describe(const ggml_tensor* node);
+size_t sd_ggml_modulate_workspace_size(const ggml_tensor* node);
+
 ggml_ext_op_t sd_ggml_rope_op();
 bool sd_ggml_rope_describe(const ggml_tensor* node);
 // Apply precomputed rotary matrices, with shapes in ne[] order:

src/core/vulkan/CMakeLists.txt

@@ -48,6 +48,7 @@ endfunction()
 # Preserve the native compilation policy: optimization changes RoPE reductions
 # and is disabled for KHR cooperative matrix shaders.
 sd_vulkan_shader(rope_apply rope_apply ON)
+sd_vulkan_shader(modulate modulate ON -O)
 sd_vulkan_shader(rms_rope_apply rms_rope_apply ON)
 sd_vulkan_shader(rms_rope_cast rms_rope_apply ON -DCAST_F16)
 sd_vulkan_shader(quantize_i8_convrot quantize_i8_convrot ${SD_VK_DOT} -O)
@@ -65,5 +66,5 @@ sd_vulkan_shader(mul_mat_i8_swiglu_split_cm2 mul_mat_i8_tensorwise_cm2 ${SD_VK_C
 sd_vulkan_shader(mul_mat_i8_residual_cm2 mul_mat_i8_tensorwise_cm2 ${SD_VK_CM2} -O -DRESIDUAL=1)
 add_custom_target(sd-vulkan-shaders DEPENDS ${SD_VULKAN_HEADERS})
 add_dependencies(${SD_LIB} sd-vulkan-shaders)
-target_sources(${SD_LIB} PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/int8.cpp")
+target_sources(${SD_LIB} PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/int8.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/modulate.cpp")
 target_include_directories(${SD_LIB} PRIVATE "${SD_VULKAN_GENERATED}")

src/core/vulkan/int8.cpp

@@ -165,22 +165,15 @@ namespace sd_vulkan_i8 {
         const ggml_tensor* nodes[4]{};
 
         const ggml_tensor* mod = nullptr;
-        const bool affine      = quant->src[0]->op == GGML_OP_ADD;
+        const bool affine      = sd_ggml_modulate_describe(quant->src[0]);
-        const int count        = affine ? 4 : 3;
+        const int count        = affine ? 2 : 3;
         if (affine) {
-            const ggml_tensor* shift = quant->src[0];
+            nodes[0]                  = quant->src[0];
-            const ggml_tensor* add   = shift->src[0];
+            nodes[1]                  = quant;
-            if (add->op != GGML_OP_ADD || add->src[1]->op != GGML_OP_MUL || add->src[1]->src[0] != add->src[0]) {
-                return 0;
-            }
-            nodes[0]                  = add->src[1];
-            nodes[1]                  = add;
-            nodes[2]                  = shift;
-            nodes[3]                  = quant;
             mod                       = nodes[0]->src[1];
-            const ggml_tensor* vector = shift->src[1];
+            const ggml_tensor* vector = nodes[0]->src[2];
-            if (vector->type != GGML_TYPE_F32 || !ggml_is_contiguous(vector) ||
+            if (vector && (vector->type != GGML_TYPE_F32 || !ggml_is_contiguous(vector) ||
-                vector->ne[0] != add->ne[0] || ggml_nelements(vector) != add->ne[0]) {
+                           vector->ne[0] != nodes[0]->ne[0] || ggml_nelements(vector) != nodes[0]->ne[0])) {
                 return 0;
             }
         } else {
@@ -215,7 +208,7 @@ namespace sd_vulkan_i8 {
     }
 
     static int match_rms_modulate_convrot(const Device& device, const ggml_cgraph* graph, int q, int* indices) {
-        if (!can_modulate(device, 5376, true) || match_modulate_convrot(device, graph, q, indices + 2) != 4) {
+        if (!can_modulate(device, 5376, true) || match_modulate_convrot(device, graph, q, indices + 2) != 2) {
             return 0;
         }
         const auto weighted = node(graph, indices[2])->src[0];
@@ -236,27 +229,27 @@ namespace sd_vulkan_i8 {
             }
         }
 
-        return next < 0 ? 6 : 0;
+        return next < 0 ? 4 : 0;
     }
 
     static int can_fuse_modulate_convrot(const Device& device, const ggml_cgraph* graph, int i) {
         const ggml_op op = node(graph, i)->op;
         if (op == GGML_OP_RMS_NORM) {
-            int indices[6];
+            int indices[4];
-            if (match_rms_modulate_convrot(device, graph, i + 5, indices) != 6) {
+            if (match_rms_modulate_convrot(device, graph, i + 3, indices) != 4) {
                 return 0;
             }
-            for (int j = 0; j < 6; ++j) {
+            for (int j = 0; j < 4; ++j) {
                 if (indices[j] != i + j) {
                     return 0;
                 }
             }
-            return 6;
+            return 4;
         }
-        if (op != GGML_OP_NORM && op != GGML_OP_MUL) {
+        if (op != GGML_OP_NORM && !sd_ggml_modulate_describe(node(graph, i))) {
             return 0;
         }
-        const int count = op == GGML_OP_NORM ? 3 : 4;
+        const int count = op == GGML_OP_NORM ? 3 : 2;
         int indices[4];
         if (match_modulate_convrot(device, graph, i + count - 1, indices) != count) {
             return 0;
@@ -491,23 +484,24 @@ namespace sd_vulkan_i8 {
     static ggml_status quantize(const ggml_ext_execution_context* e, const Device& d, ggml_tensor* const* nodes, int count) {
         const auto* output                          = nodes[count - 1];
         const auto* first                           = nodes[0];
-        const bool modulate                         = count > 1 && (first->op == GGML_OP_NORM || first->op == GGML_OP_MUL || first->op == GGML_OP_RMS_NORM);
+        const bool modulate                         = count > 1 && (first->op == GGML_OP_NORM || sd_ggml_modulate_describe(first) || first->op == GGML_OP_RMS_NORM);
         const bool rms                              = modulate && first->op == GGML_OP_RMS_NORM;
-        const bool affine                           = modulate && (rms || first->op == GGML_OP_MUL);
+        const bool affine                           = modulate && (rms || sd_ggml_modulate_describe(first));
         const bool sigmoid                          = count > 1 && first->op == GGML_OP_UNARY;
         const auto* x                               = count == 1 ? output->src[0] : first->src[0];
         const auto* up                              = count == 1 || modulate ? nullptr : sigmoid ? nodes[1]->src[0]
                                                                                                  : first->src[1];
         const int extra                             = rms ? 2 : 0;
         const auto* mod                             = modulate ? nodes[extra + (affine ? 0 : 1)]->src[1] : nullptr;
-        const auto* shift                           = modulate ? (affine ? nodes[extra + 2]->src[1] : mod) : nullptr;
+        const auto* shift                           = modulate ? (affine ? nodes[extra]->src[2] : mod) : nullptr;
         const auto* weight                          = rms ? nodes[1]->src[1] : x;
         std::vector<ggml_ext_vulkan_binding> inputs = {{x}};
         if (up)
             inputs.push_back({up});
         if (modulate) {
             inputs.push_back({mod});
-            inputs.push_back({shift});
+            if (shift)
+                inputs.push_back({shift});
             if (rms)
                 inputs.push_back({weight});
         }
@@ -521,7 +515,7 @@ namespace sd_vulkan_i8 {
         const uint32_t groups[] = {std::min(rows, d.info.max_group_count[0]), 1, 1};
         ggml_status status;
         if (modulate) {
-            std::vector<ggml_ext_vulkan_binding> bindings = {{x}, {mod}, {shift}, packed, scales};
+            std::vector<ggml_ext_vulkan_binding> bindings = {{x}, {mod}, {shift ? shift : mod}, packed, scales};
             if (rms)
                 bindings.push_back({weight});
             ModulateParameters pc = {rows, 0, 0, 0, affine && !rms ? 0.f : param_f32(first, 0), k, 0};
@@ -534,9 +528,9 @@ namespace sd_vulkan_i8 {
                     *offsets[i] = info.shader_offset / 4;
                 }
             auto pipeline              = rms ? SD_VK_PIPELINE(rms_modulate_convrot, 6, ModulateParameters) : SD_VK_PIPELINE(modulate_convrot, 5, ModulateParameters);
-            const uint32_t constants[] = {k, uint32_t(affine)};
+            const uint32_t constants[] = {k, uint32_t(affine), uint32_t(shift != nullptr)};
             pipeline.constants         = constants;
-            pipeline.constant_count    = 2;
+            pipeline.constant_count    = 3;
             status                     = d.api->dispatch(e->backend, &pipeline, bindings.data(), &pc, groups);
         } else {
             ggml_ext_vulkan_binding bindings[] = {{x}, packed, scales, {up}};
@@ -788,6 +782,9 @@ void sd_ggml_register_vulkan_i8(ggml_backend_dev_t device) {
     d.custom_op = sd_ggml_i8_op();
     GGML_ASSERT(ggml_ext_register_kernel(device, &d));
     d.optimize  = nullptr;
+    d.name      = "sd.modulate_convrot.vulkan.v1";
+    d.custom_op = sd_ggml_modulate_op();
+    GGML_ASSERT(ggml_ext_register_kernel(device, &d));
     d.custom_op = nullptr;
     const struct {
         ggml_op op;
@@ -795,7 +792,6 @@ void sd_ggml_register_vulkan_i8(ggml_backend_dev_t device) {
     } fusions[] = {
         {GGML_OP_NORM, "sd.norm_convrot.vulkan.v1"},
         {GGML_OP_RMS_NORM, "sd.rms_modulate_convrot.vulkan.v1"},
-        {GGML_OP_MUL, "sd.modulate_convrot.vulkan.v1"},
         {GGML_OP_GLU, "sd.swiglu_convrot.vulkan.v1"},
         {GGML_OP_UNARY, "sd.sigmoid_convrot.vulkan.v1"}};
     for (const auto& fusion : fusions) {

src/core/vulkan/modulate.cpp

@@ -0,0 +1,134 @@
+#include "modulate.h"
+#include <algorithm>
+#include <climits>
+#include "core/ggml_extensions.h"
+#include "ggml-ext-vulkan.h"
+
+namespace sd_vulkan_modulate {
+    struct Parameters {
+        uint32_t ne[4], scale_ne[4], shift_ne[4];
+        uint32_t x_stride[4], scale_stride[4], shift_stride[4];
+        uint32_t offsets[4];
+        uint32_t count, has_shift;
+    };
+    static_assert(sizeof(Parameters) == 120, "Modulate push constant layout mismatch");
+
+    static const ggml_ext_vulkan_api* api(ggml_backend_dev_t device) {
+        auto query = reinterpret_cast<ggml_ext_vulkan_api_fn>(ggml_backend_reg_get_proc_address(
+            ggml_backend_dev_backend_reg(device), "ggml_ext_vulkan_v1"));
+        return query ? query() : nullptr;
+    }
+
+    static bool supports(ggml_backend_dev_t device, const ggml_tensor* node, void*) {
+        const auto* service = api(device);
+        ggml_ext_vulkan_device_info info;
+        if (!service || !service->device_info(device, &info) || !sd_ggml_modulate_describe(node) ||
+            info.max_group_size[0] < 256 || info.max_group_invocations < 256 || info.max_push_constants < sizeof(Parameters) || ggml_nelements(node) > UINT32_MAX - 255) {
+            return false;
+        }
+        const ggml_tensor* tensors[] = {node, node->src[0], node->src[1], node->src[2]};
+        for (const auto* t : tensors) {
+            if (!t) {
+                continue;
+            }
+            size_t offset     = 0;
+            const auto* owner = t;
+            while (!owner->data && owner->view_src) {
+                offset += owner->view_offs;
+                owner = owner->view_src;
+            }
+            if ((reinterpret_cast<uintptr_t>(owner->data) + offset) % 4 ||
+                ggml_nbytes(t) > info.max_buffer_range - (info.buffer_alignment - 1)) {
+                return false;
+            }
+            for (int d = 0; d < 4; ++d) {
+                if (t->nb[d] / 4 > UINT32_MAX || t->ne[d] > UINT32_MAX) {
+                    return false;
+                }
+            }
+        }
+        return (uint64_t(ggml_nelements(node)) + 255) / 256 <= uint64_t(info.max_group_count[0]) * info.max_group_count[1];
+    }
+
+    static ggml_status prepare(ggml_backend_t, ggml_tensor* const* nodes, int, void**, size_t* bytes, void*) {
+        *bytes = sd_ggml_modulate_workspace_size(nodes[0]);
+        return GGML_STATUS_SUCCESS;
+    }
+
+    static void destroy(void*, void*) {}
+
+    static ggml_status execute(const ggml_ext_execution_context* e, ggml_tensor* const* nodes, int, void*, void*) {
+        auto device         = ggml_backend_get_device(e->backend);
+        const auto* service = api(device);
+        ggml_ext_vulkan_device_info info;
+        if (!service || !service->device_info(device, &info)) {
+            return GGML_STATUS_FAILED;
+        }
+        const auto* dst                    = nodes[0];
+        const auto* x                      = dst->src[0];
+        const auto* scale                  = dst->src[1];
+        const auto* shift                  = dst->src[2];
+        ggml_ext_vulkan_binding bindings[] = {{x}, {scale}, {shift ? shift : scale}, {dst}};
+        ggml_ext_vulkan_buffer_info buffers[4];
+        for (int i = 0; i < 4; ++i) {
+            if (!service->buffer_info(e->backend, &bindings[i], &buffers[i])) {
+                return GGML_STATUS_FAILED;
+            }
+        }
+        bool alias = false;
+        for (int i = 0; i < 3; ++i) {
+            alias |= buffers[i].storage == buffers[3].storage && buffers[i].offset < buffers[3].offset + buffers[3].size &&
+                     buffers[3].offset < buffers[i].offset + buffers[i].size;
+        }
+        const auto destination = bindings[3];
+        if (alias) {
+            bindings[3] = {nullptr, e->workspace, 0, ggml_nbytes(dst)};
+            if (!service->buffer_info(e->backend, &bindings[3], &buffers[3])) {
+                return GGML_STATUS_FAILED;
+            }
+        }
+        Parameters p = {};
+        p.count      = uint32_t(ggml_nelements(dst));
+        p.has_shift  = shift != nullptr;
+        for (int d = 0; d < 4; ++d) {
+            p.ne[d]           = uint32_t(x->ne[d]);
+            p.scale_ne[d]     = uint32_t(scale->ne[d]);
+            p.shift_ne[d]     = shift ? uint32_t(shift->ne[d]) : 1;
+            p.x_stride[d]     = uint32_t(x->nb[d] / 4);
+            p.scale_stride[d] = uint32_t(scale->nb[d] / 4);
+            p.shift_stride[d] = shift ? uint32_t(shift->nb[d] / 4) : 0;
+            p.offsets[d]      = buffers[d].shader_offset / 4;
+        }
+        const uint32_t count              = (p.count + 255) / 256;
+        const uint32_t width              = std::min(count, info.max_group_count[0]);
+        const uint32_t groups[]           = {width, (count + width - 1) / width, 1};
+        ggml_ext_vulkan_pipeline pipeline = {};
+        pipeline.name                     = "sd.modulate.vulkan";
+        pipeline.spirv                    = sd_vulkan_shaders::modulate;
+        pipeline.spirv_size               = sizeof(sd_vulkan_shaders::modulate);
+        pipeline.binding_count            = 4;
+        pipeline.push_constant_size       = sizeof(Parameters);
+        auto status                       = service->dispatch(e->backend, &pipeline, bindings, &p, groups);
+        if (status == GGML_STATUS_SUCCESS && alias) {
+            status = service->copy(e->backend, &bindings[3], &destination);
+        }
+        return status;
+    }
+}
+
+void sd_ggml_register_vulkan_modulate(ggml_backend_dev_t device) {
+    if (!sd_vulkan_modulate::api(device)) {
+        return;
+    }
+    ggml_ext_kernel_desc desc = {};
+    desc.abi_version          = GGML_EXT_ABI_VERSION;
+    desc.name                 = "sd.modulate.vulkan.v1";
+    desc.op                   = GGML_OP_CUSTOM;
+    desc.custom_op            = sd_ggml_modulate_op();
+    desc.priority             = 100;
+    desc.supports             = sd_vulkan_modulate::supports;
+    desc.prepare              = sd_vulkan_modulate::prepare;
+    desc.destroy              = sd_vulkan_modulate::destroy;
+    desc.execute              = sd_vulkan_modulate::execute;
+    GGML_ASSERT(ggml_ext_register_kernel(device, &desc));
+}

src/core/vulkan/shaders/modulate.comp

@@ -0,0 +1,39 @@
+#version 450
+
+layout(local_size_x = 256) in;
+layout(binding = 0) readonly buffer X { float x[]; };
+layout(binding = 1) readonly buffer S { float scale[]; };
+layout(binding = 2) readonly buffer H { float shift[]; };
+layout(binding = 3) writeonly buffer D { float dst[]; };
+layout(push_constant) uniform Parameters {
+    uint ne[4], scale_ne[4], shift_ne[4];
+    uint x_stride[4], scale_stride[4], shift_stride[4];
+    uint offsets[4];
+    uint count, has_shift;
+} p;
+
+void main() {
+    uint group = gl_WorkGroupID.y * gl_NumWorkGroups.x + gl_WorkGroupID.x;
+    if (group >= (p.count + 255) / 256) {
+        return;
+    }
+    uint i = group * 256 + gl_LocalInvocationID.x;
+    if (i >= p.count) {
+        return;
+    }
+    uint index = i;
+    uint xi = p.offsets[0], si = p.offsets[1], hi = p.offsets[2];
+    for (uint d = 0; d < 4; ++d) {
+        uint coord = index % p.ne[d];
+        index /= p.ne[d];
+        xi += coord * p.x_stride[d];
+        si += (coord % p.scale_ne[d]) * p.scale_stride[d];
+        hi += (coord % p.shift_ne[d]) * p.shift_stride[d];
+    }
+    precise float product = x[xi] * scale[si];
+    precise float value   = x[xi] + product;
+    if (p.has_shift != 0) {
+        value += shift[hi];
+    }
+    dst[p.offsets[3] + i] = value;
+}

src/core/vulkan/shaders/modulate_convrot.comp

@@ -4,6 +4,7 @@
 
 layout(constant_id = 0) const uint K      = 4096;
 layout(constant_id = 1) const bool AFFINE = false;
+layout(constant_id = 2) const bool HAS_SHIFT = true;
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer X {
     float data_x[];
@@ -111,7 +112,9 @@ void main() {
             if (AFFINE) {
                 precise float modulated = values[col] * data_m[p.m_offset + col];
                 value                   = values[col] + modulated;
-                value += data_h[p.h_offset + col];
+                if (HAS_SHIFT) {
+                    value += data_h[p.h_offset + col];
+                }
             } else {
                 value = (values[col] - mean) * inv_std;
                 value *= data_m[p.m_offset + col];

src/model/diffusion/anima.hpp

@@ -129,7 +129,7 @@ namespace Anima {
             auto gate       = emb_chunks[2];
 
             auto x = norm->forward(ctx, hidden_states);
-            x      = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            x      = Flux::modulate(ctx, x, shift, scale);
 
             return {x, gate};
         }
@@ -169,7 +169,7 @@ namespace Anima {
             auto scale      = emb_chunks[1];
 
             auto x = norm->forward(ctx, hidden_states);
-            x      = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            x      = Flux::modulate(ctx, x, shift, scale);
             return x;
         }
     };

src/model/diffusion/boogu.hpp

@@ -128,9 +128,9 @@ namespace Boogu {
         }
     };
 
-    __STATIC_INLINE__ ggml_tensor* scale_modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* scale) {
+    __STATIC_INLINE__ ggml_tensor* scale_modulate(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* scale) {
-        scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);
+        scale = ggml_reshape_3d(ctx->ggml_ctx, scale, scale->ne[0], 1, scale->ne[1]);
-        return ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        return ggml_ext_modulate(ctx, x, nullptr, scale);
     }
 
     __STATIC_INLINE__ ggml_tensor* gate_residual(ggml_context* ctx, ggml_tensor* residual, ggml_tensor* x, ggml_tensor* gate) {
@@ -186,7 +186,7 @@ namespace Boogu {
             auto scale_mlp = mods[2];
             auto gate_mlp  = mods[3];
 
-            x = scale_modulate(ctx->ggml_ctx, norm->forward(ctx, x), scale_msa);
+            x = scale_modulate(ctx, norm->forward(ctx, x), scale_msa);
             return {x, gate_msa, scale_mlp, gate_mlp};
         }
     };
@@ -231,7 +231,7 @@ namespace Boogu {
             auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
 
             auto emb = linear_1->forward(ctx, ggml_silu(ctx->ggml_ctx, conditioning_embedding));
-            x        = scale_modulate(ctx->ggml_ctx, norm->forward(ctx, x), emb);
+            x        = scale_modulate(ctx, norm->forward(ctx, x), emb);
             x        = linear_2->forward(ctx, x);
             return x;
         }
@@ -334,7 +334,7 @@ namespace Boogu {
                 auto attn_output = attn->forward(ctx, norm_hidden_states, norm_hidden_states, rotary_emb, attention_mask);
                 hidden_states    = gate_residual(ctx->ggml_ctx, hidden_states, norm2->forward(ctx, attn_output), gate_msa);
 
-                auto mlp_input  = scale_modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, hidden_states), scale_mlp);
+                auto mlp_input  = scale_modulate(ctx, ffn_norm1->forward(ctx, hidden_states), scale_mlp);
                 auto mlp_output = feed_forward->forward(ctx, mlp_input);
                 hidden_states   = gate_residual(ctx->ggml_ctx, hidden_states, ffn_norm2->forward(ctx, mlp_output), gate_mlp);
             } else {
@@ -512,14 +512,14 @@ namespace Boogu {
             img_hidden_states = gate_residual(ctx->ggml_ctx, img_hidden_states, img_attn_norm->forward(ctx, img_attn_out), img_gate_msa);
             img_hidden_states = gate_residual(ctx->ggml_ctx, img_hidden_states, img_self_attn_norm->forward(ctx, img_self_attn_out), img_gate_self);
 
-            auto img_mlp_input = scale_modulate(ctx->ggml_ctx, img_norm2_out, img_scale_mlp);
+            auto img_mlp_input = scale_modulate(ctx, img_norm2_out, img_scale_mlp);
             img_shift_mlp      = ggml_reshape_3d(ctx->ggml_ctx, img_shift_mlp, img_shift_mlp->ne[0], 1, img_shift_mlp->ne[1]);
             img_mlp_input      = ggml_add(ctx->ggml_ctx, img_mlp_input, img_shift_mlp);
             auto img_mlp_out   = img_feed_forward->forward(ctx, img_ffn_norm1->forward(ctx, img_mlp_input));
             img_hidden_states  = gate_residual(ctx->ggml_ctx, img_hidden_states, img_ffn_norm2->forward(ctx, img_mlp_out), img_gate_mlp);
 
             instruct_hidden_states  = gate_residual(ctx->ggml_ctx, instruct_hidden_states, instruct_attn_norm->forward(ctx, instruct_attn_out), instruct_gate_msa);
-            auto instruct_mlp_input = scale_modulate(ctx->ggml_ctx, instruct_norm2_out, instruct_scale_mlp);
+            auto instruct_mlp_input = scale_modulate(ctx, instruct_norm2_out, instruct_scale_mlp);
             instruct_shift_mlp      = ggml_reshape_3d(ctx->ggml_ctx, instruct_shift_mlp, instruct_shift_mlp->ne[0], 1, instruct_shift_mlp->ne[1]);
             instruct_mlp_input      = ggml_add(ctx->ggml_ctx, instruct_mlp_input, instruct_shift_mlp);
             auto instruct_mlp_out   = instruct_feed_forward->forward(ctx, instruct_ffn_norm1->forward(ctx, instruct_mlp_input));

src/model/diffusion/ernie_image.hpp

@@ -246,13 +246,13 @@ namespace ErnieImage {
 
             auto residual = x;
             x             = adaLN_sa_ln->forward(ctx, x);
-            x             = Flux::modulate(ctx->ggml_ctx, x, shift_msa, scale_msa, true);
+            x             = Flux::modulate(ctx, x, shift_msa, scale_msa, true);
             auto attn_out = self_attention->forward(ctx, x, pe, attention_mask);
             x             = ggml_add(ctx->ggml_ctx, residual, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
 
             residual = x;
             x        = adaLN_mlp_ln->forward(ctx, x);
-            x        = Flux::modulate(ctx->ggml_ctx, x, shift_mlp, scale_mlp, true);
+            x        = Flux::modulate(ctx, x, shift_mlp, scale_mlp, true);
             x        = ggml_add(ctx->ggml_ctx, residual, ggml_mul(ctx->ggml_ctx, mlp->forward(ctx, x), gate_mlp));
             return x;
         }
@@ -274,7 +274,7 @@ namespace ErnieImage {
             auto shift = mods[1];
 
             x = norm->forward(ctx, x);
-            x = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            x = Flux::modulate(ctx, x, shift, scale);
             return x;
         }
     };

src/model/diffusion/fibo2.hpp

@@ -262,7 +262,7 @@ namespace Fibo2 {
                 }
                 return m;
             };
-            auto h = ZImage::modulate(c, norm("attention_norm1", x), select(0, false));
+            auto h = ZImage::modulate(ctx, norm("attention_norm1", x), select(0, false));
             h      = std::dynamic_pointer_cast<Attention>(blocks["attention"])->forward(ctx, h, h, pe);
             x      = ggml_add(c, x, ggml_mul(c, norm("attention_norm2", h), select(1, true)));
             if (injection) {
@@ -270,7 +270,7 @@ namespace Fibo2 {
                 h = ggml_mul(c, norm("cross_attn_norm2", h), ggml_tanh(c, params["cross_attn_gate"]));
                 x = ggml_add(c, x, h);
             }
-            h = ZImage::modulate(c, norm("ffn_norm1", x), select(2, false));
+            h = ZImage::modulate(ctx, norm("ffn_norm1", x), select(2, false));
             h = std::dynamic_pointer_cast<FeedForward>(blocks["feed_forward"])->forward(ctx, h);
             return ggml_add(c, x, ggml_mul(c, norm("ffn_norm2", h), select(3, true)));
         }

src/model/diffusion/flux.hpp

@@ -413,7 +413,7 @@ namespace Flux {
         }
     };
 
-    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* modulate(GGMLRunnerContext* ctx,
                                             ggml_tensor* x,
                                             ggml_tensor* shift,
                                             ggml_tensor* scale,
@@ -422,12 +422,10 @@ namespace Flux {
         // scale: [N, C]
         // shift: [N, C]
         if (!skip_reshape) {
-            scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
+            scale = ggml_reshape_3d(ctx->ggml_ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
-            shift = ggml_reshape_3d(ctx, shift, shift->ne[0], 1, shift->ne[1]);  // [N, 1, C]
+            shift = ggml_reshape_3d(ctx->ggml_ctx, shift, shift->ne[0], 1, shift->ne[1]);  // [N, 1, C]
         }
-        x = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        return ggml_ext_modulate(ctx, x, shift, scale);
-        x = ggml_add(ctx, x, shift);
-        return x;
     }
 
     struct DoubleStreamBlock : public GGMLBlock {
@@ -540,7 +538,7 @@ namespace Flux {
 
             // prepare image for attention
             auto img_modulated = img_norm1->forward(ctx, img);
-            img_modulated      = Flux::modulate(ctx->ggml_ctx, img_modulated, img_mod1.shift, img_mod1.scale);
+            img_modulated      = Flux::modulate(ctx, img_modulated, img_mod1.shift, img_mod1.scale);
             auto img_qkv       = img_attn->pre_attention(ctx, img_modulated);  // q,k,v: [N, n_img_token, n_head, d_head]
             auto img_q         = img_qkv[0];
             auto img_k         = img_qkv[1];
@@ -548,7 +546,7 @@ namespace Flux {
 
             // prepare txt for attention
             auto txt_modulated = txt_norm1->forward(ctx, txt);
-            txt_modulated      = Flux::modulate(ctx->ggml_ctx, txt_modulated, txt_mod1.shift, txt_mod1.scale);
+            txt_modulated      = Flux::modulate(ctx, txt_modulated, txt_mod1.shift, txt_mod1.scale);
             auto txt_qkv       = txt_attn->pre_attention(ctx, txt_modulated);  // q,k,v: [N, n_txt_token, n_head, d_head]
             auto txt_q         = txt_qkv[0];
             auto txt_k         = txt_qkv[1];
@@ -580,14 +578,14 @@ namespace Flux {
             // calculate the img bloks
             img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_attn->post_attention(ctx, img_attn_out), img_mod1.gate));
 
-            auto img_mlp_out = img_mlp->forward(ctx, Flux::modulate(ctx->ggml_ctx, img_norm2->forward(ctx, img), img_mod2.shift, img_mod2.scale));
+            auto img_mlp_out = img_mlp->forward(ctx, Flux::modulate(ctx, img_norm2->forward(ctx, img), img_mod2.shift, img_mod2.scale));
 
             img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_mlp_out, img_mod2.gate));
 
             // calculate the txt bloks
             txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn->post_attention(ctx, txt_attn_out), txt_mod1.gate));
 
-            auto txt_mlp_out = txt_mlp->forward(ctx, Flux::modulate(ctx->ggml_ctx, txt_norm2->forward(ctx, txt), txt_mod2.shift, txt_mod2.scale));
+            auto txt_mlp_out = txt_mlp->forward(ctx, Flux::modulate(ctx, txt_norm2->forward(ctx, txt), txt_mod2.shift, txt_mod2.scale));
             txt              = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_mlp_out, txt_mod2.gate));
 
             return {img, txt};
@@ -670,7 +668,7 @@ namespace Flux {
                 }
             }
 
-            auto x_mod   = Flux::modulate(ctx->ggml_ctx, pre_norm->forward(ctx, x), mod.shift, mod.scale);
+            auto x_mod   = Flux::modulate(ctx, pre_norm->forward(ctx, x), mod.shift, mod.scale);
             auto qkv_mlp = linear1->forward(ctx, x_mod);  // [N, n_token, hidden_size * 3 + mlp_hidden_dim*mlp_mult_factor]
 
             int64_t head_dim = hidden_size / num_heads;
@@ -752,7 +750,7 @@ namespace Flux {
                 scale      = m_vec[1];
             }
 
-            x = Flux::modulate(ctx->ggml_ctx, norm_final->forward(ctx, x), shift, scale);
+            x = Flux::modulate(ctx, norm_final->forward(ctx, x), shift, scale);
             x = linear->forward(ctx, x);
 
             return x;

src/model/diffusion/flux3.hpp

@@ -83,7 +83,7 @@ namespace Flux3 {
             int64_t offset     = 0;
             for (size_t i = 0; i < lengths.size(); ++i) {
                 auto part = token_slice(gc, x, offset, lengths[i]);
-                part      = Flux::modulate(gc, ggml_norm(gc, part, 1e-6f), mods[i].shift, mods[i].scale);
+                part      = Flux::modulate(ctx, ggml_norm(gc, part, 1e-6f), mods[i].shift, mods[i].scale);
                 x_mod     = x_mod ? ggml_concat(gc, x_mod, part, 1) : part;
                 offset += lengths[i];
             }

src/model/diffusion/ideogram4.hpp

@@ -103,9 +103,9 @@ namespace Ideogram4 {
         return ggml_reshape_3d(ctx, x, hidden_size * LLM_HIDDEN_STATE_LAYERS, token_count, batch_count);
     }
 
-    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* scale) {
+    __STATIC_INLINE__ ggml_tensor* modulate(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* scale) {
-        scale = to_token_modulation(ctx, scale);
+        scale = to_token_modulation(ctx->ggml_ctx, scale);
-        return ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        return ggml_ext_modulate(ctx, x, nullptr, scale);
     }
 
     __STATIC_INLINE__ ggml_tensor* patchify(ggml_context* ctx, ggml_tensor* x, const Ideogram4Config& config) {
@@ -285,13 +285,13 @@ namespace Ideogram4 {
             auto gate_mlp  = to_token_modulation(ctx->ggml_ctx, ggml_tanh(ctx->ggml_ctx, mods[3]));
 
             auto attn_out = attention_norm1->forward(ctx, x);
-            attn_out      = modulate(ctx->ggml_ctx, attn_out, scale_msa);
+            attn_out      = modulate(ctx, attn_out, scale_msa);
             attn_out      = attention->forward(ctx, attn_out, pe, mask);
             attn_out      = attention_norm2->forward(ctx, attn_out);
             x             = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
 
             auto ffn_out = ffn_norm1->forward(ctx, x);
-            ffn_out      = modulate(ctx->ggml_ctx, ffn_out, scale_mlp);
+            ffn_out      = modulate(ctx, ffn_out, scale_mlp);
             ffn_out      = feed_forward->forward(ctx, ffn_out);
             ffn_out      = ffn_norm2->forward(ctx, ffn_out);
             x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, ffn_out, gate_mlp));
@@ -337,7 +337,7 @@ namespace Ideogram4 {
 
             auto scale = adaln_modulation->forward(ctx, ggml_silu(ctx->ggml_ctx, c));
             x          = norm_final->forward(ctx, x);
-            x          = modulate(ctx->ggml_ctx, x, scale);
+            x          = modulate(ctx, x, scale);
             x          = linear->forward(ctx, x);
             return x;
         }

src/model/diffusion/krea2.hpp

@@ -465,8 +465,8 @@ namespace Krea2 {
                 auto pre_x_main = ggml_view_3d(ctx->ggml_ctx, pre_x, D, len_main, B, nb1, nb2, 0);
                 auto pre_x_refs = ggml_view_3d(ctx->ggml_ctx, pre_x, D, len_refs, B, nb1, nb2, len_main * nb1);
 
-                auto attn_in_main = Flux::modulate(ctx->ggml_ctx, pre_x_main, mods_main[1], mods_main[0], true);
+                auto attn_in_main = Flux::modulate(ctx, pre_x_main, mods_main[1], mods_main[0], true);
-                auto attn_in_refs = Flux::modulate(ctx->ggml_ctx, pre_x_refs, mods_refs[1], mods_refs[0], true);
+                auto attn_in_refs = Flux::modulate(ctx, pre_x_refs, mods_refs[1], mods_refs[0], true);
 
                 auto attn_input = ggml_concat(ctx->ggml_ctx, attn_in_main, attn_in_refs, 1);
 
@@ -487,8 +487,8 @@ namespace Krea2 {
                 auto post_x_main = ggml_view_3d(ctx->ggml_ctx, post_x, D, len_main, B, post_x->nb[1], post_x->nb[2], 0);
                 auto post_x_refs = ggml_view_3d(ctx->ggml_ctx, post_x, D, len_refs, B, post_x->nb[1], post_x->nb[2], len_main * post_x->nb[1]);
 
-                auto mlp_in_main = Flux::modulate(ctx->ggml_ctx, post_x_main, mods_main[4], mods_main[3], true);
+                auto mlp_in_main = Flux::modulate(ctx, post_x_main, mods_main[4], mods_main[3], true);
-                auto mlp_in_refs = Flux::modulate(ctx->ggml_ctx, post_x_refs, mods_refs[4], mods_refs[3], true);
+                auto mlp_in_refs = Flux::modulate(ctx, post_x_refs, mods_refs[4], mods_refs[3], true);
 
                 auto mlp_input = ggml_concat(ctx->ggml_ctx, mlp_in_main, mlp_in_refs, 1);
                 auto mlp_out   = mlp->forward(ctx, mlp_input);
@@ -503,7 +503,7 @@ namespace Krea2 {
                 x            = ggml_add(ctx->ggml_ctx, x, mlp_res);
             } else {
                 auto mods       = mod->forward(ctx, vec);
-                auto attn_input = Flux::modulate(ctx->ggml_ctx,
+                auto attn_input = Flux::modulate(ctx,
                                                  prenorm->forward(ctx, x),
                                                  mods[1],
                                                  mods[0],
@@ -511,7 +511,7 @@ namespace Krea2 {
                 auto attn_out   = attn->forward(ctx, attn_input, pe);
                 x               = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, mods[2]));
 
-                auto mlp_input = Flux::modulate(ctx->ggml_ctx,
+                auto mlp_input = Flux::modulate(ctx,
                                                 postnorm->forward(ctx, x),
                                                 mods[4],
                                                 mods[3],
@@ -595,7 +595,7 @@ namespace Krea2 {
             auto modulation = std::dynamic_pointer_cast<KreaFinalModulation>(blocks["modulation"]);
 
             auto mods = modulation->forward(ctx, vec);
-            x         = Flux::modulate(ctx->ggml_ctx,
+            x         = Flux::modulate(ctx,
                                        norm->forward(ctx, x),
                                        mods[1],
                                        mods[0],

src/model/diffusion/lens.hpp

@@ -237,9 +237,9 @@ namespace Lens {
             auto txt_mod_vec    = ggml_ext_chunk(ctx->ggml_ctx, txt_mod_params, 6, 0);
 
             auto img_normed    = img_norm1->forward(ctx, img);
-            auto img_modulated = Flux::modulate(ctx->ggml_ctx, img_normed, img_mod_vec[0], img_mod_vec[1]);
+            auto img_modulated = Flux::modulate(ctx, img_normed, img_mod_vec[0], img_mod_vec[1]);
             auto txt_normed    = txt_norm1->forward(ctx, txt);
-            auto txt_modulated = Flux::modulate(ctx->ggml_ctx, txt_normed, txt_mod_vec[0], txt_mod_vec[1]);
+            auto txt_modulated = Flux::modulate(ctx, txt_normed, txt_mod_vec[0], txt_mod_vec[1]);
 
             auto [img_attn_output, txt_attn_output] = attn->forward(ctx, img_modulated, txt_modulated, pe);
 
@@ -247,9 +247,9 @@ namespace Lens {
             txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn_output, txt_mod_vec[2]));
 
             auto img_normed2    = img_norm2->forward(ctx, img);
-            auto img_modulated2 = Flux::modulate(ctx->ggml_ctx, img_normed2, img_mod_vec[3], img_mod_vec[4]);
+            auto img_modulated2 = Flux::modulate(ctx, img_normed2, img_mod_vec[3], img_mod_vec[4]);
             auto txt_normed2    = txt_norm2->forward(ctx, txt);
-            auto txt_modulated2 = Flux::modulate(ctx->ggml_ctx, txt_normed2, txt_mod_vec[3], txt_mod_vec[4]);
+            auto txt_modulated2 = Flux::modulate(ctx, txt_normed2, txt_mod_vec[3], txt_mod_vec[4]);
 
             img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_mlp->forward(ctx, img_modulated2), img_mod_vec[5]));
             txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_mlp->forward(ctx, txt_modulated2), txt_mod_vec[5]));
@@ -272,7 +272,7 @@ namespace Lens {
             auto scale  = mods[0];
             auto shift  = mods[1];
             x           = ggml_norm(ctx->ggml_ctx, x, eps);
-            return Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            return Flux::modulate(ctx, x, shift, scale);
         }
     };
 

src/model/diffusion/lingbot_video.hpp

@@ -485,12 +485,12 @@ namespace LingBotVideo {
             auto scale_mlp = mods[4];
             auto gate_mlp  = ggml_tanh(ctx->ggml_ctx, mods[5]);
 
-            auto attn_in  = Flux::modulate(ctx->ggml_ctx, norm1->forward(ctx, x), shift_msa, scale_msa, true);
+            auto attn_in  = Flux::modulate(ctx, norm1->forward(ctx, x), shift_msa, scale_msa, true);
             auto attn_out = attn->forward(ctx, attn_in, pe, attention_mask);
             attn_out      = norm_post_attn->forward(ctx, attn_out);
             x             = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, gate_msa, attn_out));
 
-            auto ffn_in  = Flux::modulate(ctx->ggml_ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp, true);
+            auto ffn_in  = Flux::modulate(ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp, true);
             auto ffn_out = ffn->forward(ctx, ffn_in);
             ffn_out      = norm_post_ffn->forward(ctx, ffn_out);
             x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, gate_mlp, ffn_out));
@@ -587,7 +587,7 @@ namespace LingBotVideo {
                                              2,
                                              0);
             hidden_states   = norm_out->forward(ctx, hidden_states);
-            hidden_states   = Flux::modulate(ctx->ggml_ctx, hidden_states, final_mods[0], final_mods[1], true);
+            hidden_states   = Flux::modulate(ctx, hidden_states, final_mods[0], final_mods[1], true);
             hidden_states   = proj_out->forward(ctx, hidden_states);
             hidden_states   = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, 0, n_img);
 

src/model/diffusion/ltxv.hpp

@@ -38,12 +38,12 @@ namespace LTXV {
         return mod;
     }
 
-    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* modulate(GGMLRunnerContext* ctx,
                                             ggml_tensor* x,
                                             ggml_tensor* shift,
                                             ggml_tensor* scale) {
-        shift = align_token_modulation(ctx, x, shift);
+        shift = align_token_modulation(ctx->ggml_ctx, x, shift);
-        scale = align_token_modulation(ctx, x, scale);
+        scale = align_token_modulation(ctx->ggml_ctx, x, scale);
         return Flux::modulate(ctx, x, shift, scale, true);
     }
 
@@ -821,7 +821,7 @@ namespace LTXV {
             auto gate_mlp  = mods[5];
 
             auto x_norm = rms_norm(ctx->ggml_ctx, x);
-            x_norm      = LTXV::modulate(ctx->ggml_ctx, x_norm, shift_msa, scale_msa);
+            x_norm      = LTXV::modulate(ctx, x_norm, shift_msa, scale_msa);
             auto msa    = attn1->forward(ctx, x_norm, nullptr, self_attention_mask, pe);
             x           = ggml_add(ctx->ggml_ctx, x, apply_gate(ctx->ggml_ctx, msa, gate_msa));
 
@@ -831,12 +831,12 @@ namespace LTXV {
                 auto gate_q  = mods[8];
 
                 auto q = rms_norm(ctx->ggml_ctx, x);
-                q      = LTXV::modulate(ctx->ggml_ctx, q, shift_q, scale_q);
+                q      = LTXV::modulate(ctx, q, shift_q, scale_q);
 
                 auto context_mod = context;
                 if (prompt_timestep != nullptr) {
                     auto prompt_mods = get_prompt_scale_shift_values(ctx, prompt_timestep);
-                    context_mod      = LTXV::modulate(ctx->ggml_ctx, context_mod, prompt_mods[0], prompt_mods[1]);
+                    context_mod      = LTXV::modulate(ctx, context_mod, prompt_mods[0], prompt_mods[1]);
                 }
 
                 auto mca = attn2->forward(ctx, q, context_mod, attention_mask, nullptr, nullptr);
@@ -847,7 +847,7 @@ namespace LTXV {
             }
 
             auto y       = rms_norm(ctx->ggml_ctx, x);
-            y            = LTXV::modulate(ctx->ggml_ctx, y, shift_mlp, scale_mlp);
+            y            = LTXV::modulate(ctx, y, shift_mlp, scale_mlp);
             auto mlp_out = ff->forward(ctx, y);
             x            = ggml_add(ctx->ggml_ctx, x, apply_gate(ctx->ggml_ctx, mlp_out, gate_mlp));
             return x;
@@ -1199,11 +1199,11 @@ namespace LTXV {
             if (cross_attention_adaln) {
                 auto q_mods      = get_ada_values(ctx, table, timestep, dim, 9, 6, 3);
                 auto q           = rms_norm(ctx->ggml_ctx, x);
-                q                = LTXV::modulate(ctx->ggml_ctx, q, q_mods[0], q_mods[1]);
+                q                = LTXV::modulate(ctx, q, q_mods[0], q_mods[1]);
                 auto context_mod = context;
                 if (prompt_timestep != nullptr && prompt_table != nullptr) {
                     auto p_mods = get_ada_values(ctx, prompt_table, prompt_timestep, dim, 2);
-                    context_mod = LTXV::modulate(ctx->ggml_ctx, context_mod, p_mods[0], p_mods[1]);
+                    context_mod = LTXV::modulate(ctx, context_mod, p_mods[0], p_mods[1]);
                 }
                 auto out = attn->forward(ctx, q, context_mod, attention_mask, nullptr, nullptr);
                 return apply_gate(ctx->ggml_ctx, out, q_mods[2]);
@@ -1250,7 +1250,7 @@ namespace LTXV {
 
             auto v_mods = get_ada_values(ctx, v_table, v_timestep, v_dim, cross_attention_adaln ? 9 : 6);
             auto v_norm = rms_norm(ctx->ggml_ctx, vx);
-            v_norm      = LTXV::modulate(ctx->ggml_ctx, v_norm, v_mods[0], v_mods[1]);
+            v_norm      = LTXV::modulate(ctx, v_norm, v_mods[0], v_mods[1]);
             auto v_sa   = attn1->forward(ctx, v_norm, nullptr, self_attention_mask, v_pe);
             vx          = ggml_add(ctx->ggml_ctx, vx, apply_gate(ctx->ggml_ctx, v_sa, v_mods[2]));
             auto v_txt  = apply_text_cross_attention(ctx,
@@ -1268,7 +1268,7 @@ namespace LTXV {
             if (run_ax) {
                 auto a_mods = get_ada_values(ctx, a_table, a_timestep, a_dim, cross_attention_adaln ? 9 : 6);
                 auto a_norm = rms_norm(ctx->ggml_ctx, ax);
-                a_norm      = LTXV::modulate(ctx->ggml_ctx, a_norm, a_mods[0], a_mods[1]);
+                a_norm      = LTXV::modulate(ctx, a_norm, a_mods[0], a_mods[1]);
                 auto a_sa   = audio_attn1->forward(ctx, a_norm, nullptr, nullptr, a_pe);
                 ax          = ggml_add(ctx->ggml_ctx, ax, apply_gate(ctx->ggml_ctx, a_sa, a_mods[2]));
                 auto a_txt  = apply_text_cross_attention(ctx,
@@ -1291,8 +1291,8 @@ namespace LTXV {
                     auto a2v_video_table = ggml_ext_slice(ctx->ggml_ctx, params["scale_shift_table_a2v_ca_video"], 1, 0, 4);
                     auto a2v_audio       = get_ada_values(ctx, a2v_audio_table, a_cross_scale_shift_timestep, a_dim, 4);
                     auto a2v_video       = get_ada_values(ctx, a2v_video_table, v_cross_scale_shift_timestep, v_dim, 4);
-                    auto vx_scaled       = LTXV::modulate(ctx->ggml_ctx, vx_norm3, a2v_video[1], a2v_video[0]);
+                    auto vx_scaled       = LTXV::modulate(ctx, vx_norm3, a2v_video[1], a2v_video[0]);
-                    auto ax_scaled       = LTXV::modulate(ctx->ggml_ctx, ax_norm3, a2v_audio[1], a2v_audio[0]);
+                    auto ax_scaled       = LTXV::modulate(ctx, ax_norm3, a2v_audio[1], a2v_audio[0]);
                     auto a2v_out         = audio_to_video_attn->forward(ctx, vx_scaled, ax_scaled, nullptr, v_cross_pe, a_cross_pe);
                     auto a2v_gate_table  = ggml_ext_slice(ctx->ggml_ctx, params["scale_shift_table_a2v_ca_video"], 1, 4, 5);
                     auto a2v_gate        = get_ada_values(ctx, a2v_gate_table, v_cross_gate_timestep, v_dim, 1)[0];
@@ -1304,8 +1304,8 @@ namespace LTXV {
                     auto v2a_video_table = ggml_ext_slice(ctx->ggml_ctx, params["scale_shift_table_a2v_ca_video"], 1, 0, 4);
                     auto v2a_audio       = get_ada_values(ctx, v2a_audio_table, a_cross_scale_shift_timestep, a_dim, 4);
                     auto v2a_video       = get_ada_values(ctx, v2a_video_table, v_cross_scale_shift_timestep, v_dim, 4);
-                    auto ax_scaled       = LTXV::modulate(ctx->ggml_ctx, ax_norm3, v2a_audio[3], v2a_audio[2]);
+                    auto ax_scaled       = LTXV::modulate(ctx, ax_norm3, v2a_audio[3], v2a_audio[2]);
-                    auto vx_scaled       = LTXV::modulate(ctx->ggml_ctx, vx_norm3, v2a_video[3], v2a_video[2]);
+                    auto vx_scaled       = LTXV::modulate(ctx, vx_norm3, v2a_video[3], v2a_video[2]);
                     auto v2a_out         = video_to_audio_attn->forward(ctx, ax_scaled, vx_scaled, nullptr, a_cross_pe, v_cross_pe);
                     auto v2a_gate_table  = ggml_ext_slice(ctx->ggml_ctx, params["scale_shift_table_a2v_ca_audio"], 1, 4, 5);
                     auto v2a_gate        = get_ada_values(ctx, v2a_gate_table, a_cross_gate_timestep, a_dim, 1)[0];
@@ -1313,14 +1313,14 @@ namespace LTXV {
                 }
                 auto a_ff_mods = get_ada_values(ctx, a_table, a_timestep, a_dim, cross_attention_adaln ? 9 : 6, 3, 3);
                 auto ax_scaled = rms_norm(ctx->ggml_ctx, ax);
-                ax_scaled      = LTXV::modulate(ctx->ggml_ctx, ax_scaled, a_ff_mods[0], a_ff_mods[1]);
+                ax_scaled      = LTXV::modulate(ctx, ax_scaled, a_ff_mods[0], a_ff_mods[1]);
                 auto a_ff_out  = audio_ff->forward(ctx, ax_scaled);
                 ax             = ggml_add(ctx->ggml_ctx, ax, apply_gate(ctx->ggml_ctx, a_ff_out, a_ff_mods[2]));
             }
 
             auto v_ff_mods = get_ada_values(ctx, v_table, v_timestep, v_dim, cross_attention_adaln ? 9 : 6, 3, 3);
             auto vx_scaled = rms_norm(ctx->ggml_ctx, vx);
-            vx_scaled      = LTXV::modulate(ctx->ggml_ctx, vx_scaled, v_ff_mods[0], v_ff_mods[1]);
+            vx_scaled      = LTXV::modulate(ctx, vx_scaled, v_ff_mods[0], v_ff_mods[1]);
             auto v_ff_out  = ff->forward(ctx, vx_scaled);
             vx             = ggml_add(ctx->ggml_ctx, vx, apply_gate(ctx->ggml_ctx, v_ff_out, v_ff_mods[2]));
 
@@ -1720,14 +1720,14 @@ namespace LTXV {
 
             auto v_shift_scale = get_output_scale_shift(ctx, params["scale_shift_table"], v_embedded_time, config.hidden_size);
             vx                 = norm_out->forward(ctx, vx);
-            vx                 = LTXV::modulate(ctx->ggml_ctx, vx, v_shift_scale[0], v_shift_scale[1]);
+            vx                 = LTXV::modulate(ctx, vx, v_shift_scale[0], v_shift_scale[1]);
             vx                 = proj_out->forward(ctx, vx);
             vx                 = unpatchify_video(ctx, vx, width, height, frames);
 
             if (ax != nullptr && audio_time > 0) {
                 auto a_shift_scale = get_output_scale_shift(ctx, params["audio_scale_shift_table"], a_embedded_time, config.audio_hidden_size);
                 ax                 = audio_norm_out->forward(ctx, ax);
-                ax                 = LTXV::modulate(ctx->ggml_ctx, ax, a_shift_scale[0], a_shift_scale[1]);
+                ax                 = LTXV::modulate(ctx, ax, a_shift_scale[0], a_shift_scale[1]);
                 ax                 = audio_proj_out->forward(ctx, ax);
                 ax                 = unpatchify_audio(ctx, ax, audio_time);
             }

src/model/diffusion/minimax_h3.hpp

@@ -445,9 +445,7 @@ namespace MiniMaxH3 {
                 inputs.push_back(ggml_ext_slice(ctx->ggml_ctx, x, 1, segment.start, segment.end, false));
             }
             auto modulate = [&](ggml_tensor* value, const std::vector<ggml_tensor*>& row, int shift, int scale) {
-                return ggml_add(ctx->ggml_ctx,
+                return ggml_ext_modulate(ctx, value, row[shift], row[scale]);
-                                ggml_add(ctx->ggml_ctx, value, ggml_mul(ctx->ggml_ctx, value, row[scale])),
-                                row[shift]);
             };
             ggml_tensor* qkv = nullptr;
             for (size_t i = 0; i < segments.size(); ++i) {

src/model/diffusion/mmdit.hpp

@@ -372,18 +372,16 @@ public:
     }
 };
 
-__STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
+__STATIC_INLINE__ ggml_tensor* modulate(GGMLRunnerContext* ctx,
                                         ggml_tensor* x,
                                         ggml_tensor* shift,
                                         ggml_tensor* scale) {
     // x: [N, L, C]
     // scale: [N, C]
     // shift: [N, C]
-    scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
+    scale = ggml_reshape_3d(ctx->ggml_ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
-    shift = ggml_reshape_3d(ctx, shift, shift->ne[0], 1, shift->ne[1]);  // [N, 1, C]
+    shift = ggml_reshape_3d(ctx->ggml_ctx, shift, shift->ne[0], 1, shift->ne[1]);  // [N, 1, C]
-    x     = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+    return ggml_ext_modulate(ctx, x, shift, scale);
-    x     = ggml_add(ctx, x, shift);
-    return x;
 }
 
 struct DismantledBlock : public GGMLBlock {
@@ -455,10 +453,10 @@ public:
 
         auto x_norm = norm1->forward(ctx, x);
 
-        auto attn_in = modulate(ctx->ggml_ctx, x_norm, shift_msa, scale_msa);
+        auto attn_in = modulate(ctx, x_norm, shift_msa, scale_msa);
         auto qkv     = attn->pre_attention(ctx, attn_in);
 
-        auto attn2_in = modulate(ctx->ggml_ctx, x_norm, shift_msa2, scale_msa2);
+        auto attn2_in = modulate(ctx, x_norm, shift_msa2, scale_msa2);
         auto qkv2     = attn2->pre_attention(ctx, attn2_in);
 
         return {qkv, qkv2, {x, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2}};
@@ -488,13 +486,13 @@ public:
             auto scale_mlp = m_vec[4];  // [N, hidden_size]
             auto gate_mlp  = m_vec[5];  // [N, hidden_size]
 
-            auto attn_in = modulate(ctx->ggml_ctx, norm1->forward(ctx, x), shift_msa, scale_msa);
+            auto attn_in = modulate(ctx, norm1->forward(ctx, x), shift_msa, scale_msa);
 
             auto qkv = attn->pre_attention(ctx, attn_in);
 
             return {qkv, {x, gate_msa, shift_mlp, scale_mlp, gate_mlp}};
         } else {
-            auto attn_in = modulate(ctx->ggml_ctx, norm1->forward(ctx, x), shift_msa, scale_msa);
+            auto attn_in = modulate(ctx, norm1->forward(ctx, x), shift_msa, scale_msa);
             auto qkv     = attn->pre_attention(ctx, attn_in);
 
             return {qkv, {nullptr, nullptr, nullptr, nullptr, nullptr}};
@@ -533,7 +531,7 @@ public:
 
         x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
         x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn2_out, gate_msa2));
-        auto mlp_out = mlp->forward(ctx, modulate(ctx->ggml_ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp));
+        auto mlp_out = mlp->forward(ctx, modulate(ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp));
         x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, mlp_out, gate_mlp));
 
         return x;
@@ -565,7 +563,7 @@ public:
         attn_out = attn->post_attention(ctx, attn_out);
 
         x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
-        auto mlp_out = mlp->forward(ctx, modulate(ctx->ggml_ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp));
+        auto mlp_out = mlp->forward(ctx, modulate(ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp));
         x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, mlp_out, gate_mlp));
 
         return x;
@@ -756,7 +754,7 @@ public:
         auto shift = m_vec[0];  // [N, hidden_size]
         auto scale = m_vec[1];  // [N, hidden_size]
 
-        x = modulate(ctx->ggml_ctx, norm_final->forward(ctx, x), shift, scale);
+        x = modulate(ctx, norm_final->forward(ctx, x), shift, scale);
         x = linear->forward(ctx, x);
 
         return x;

src/model/diffusion/pixart.hpp

@@ -198,11 +198,11 @@ namespace PixArt {
                 return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, y, g));
             };
 
-            auto h = modulate(ctx->ggml_ctx, norm(ctx->ggml_ctx, x), mv[0], mv[1]);
+            auto h = modulate(ctx, norm(ctx->ggml_ctx, x), mv[0], mv[1]);
             x      = gate(attn1->forward(ctx, h, h), mv[2]);
             // ada_norm_single: no norm before cross-attention (PixArtMS.py)
             x = ggml_add(ctx->ggml_ctx, x, attn2->forward(ctx, x, context, context_mask));
-            h = modulate(ctx->ggml_ctx, norm(ctx->ggml_ctx, x), mv[3], mv[4]);
+            h = modulate(ctx, norm(ctx->ggml_ctx, x), mv[3], mv[4]);
             h = proj->forward(ctx, h);
             h = ggml_ext_gelu(ctx->ggml_ctx, h, true);
             h = fc2->forward(ctx, h);
@@ -288,7 +288,7 @@ namespace PixArt {
                                         ggml_reshape_2d(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, ss), config.hidden_size * 2, N),
                                         2, 0);
             h          = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer.norm_final"])->forward(ctx, h);
-            h          = modulate(ctx->ggml_ctx, h, parts[0], parts[1]);
+            h          = modulate(ctx, h, parts[0], parts[1]);
             h          = std::dynamic_pointer_cast<Linear>(blocks["final_layer.linear"])->forward(ctx, h);  // [N, hp*wp, p*p*out_ch]
             h          = DiT::unpatchify(ctx->ggml_ctx, h, hp, wp, static_cast<int>(p), static_cast<int>(p), false);
             return h;  // [N, out_channels, H, W]

src/model/diffusion/qwen_image.hpp

@@ -350,11 +350,11 @@ namespace Qwen {
             auto txt_mod_param_vec = get_mod_params_vec(ctx->ggml_ctx, txt_mod_params);
 
             auto img_normed    = img_norm1->forward(ctx, img);
-            auto img_modulated = Flux::modulate(ctx->ggml_ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1], modulate_index != nullptr);
+            auto img_modulated = Flux::modulate(ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1], modulate_index != nullptr);
             auto img_gate1     = img_mod_param_vec[2];
 
             auto txt_normed    = txt_norm1->forward(ctx, txt);
-            auto txt_modulated = Flux::modulate(ctx->ggml_ctx, txt_normed, txt_mod_param_vec[0], txt_mod_param_vec[1]);
+            auto txt_modulated = Flux::modulate(ctx, txt_normed, txt_mod_param_vec[0], txt_mod_param_vec[1]);
             auto txt_gate1     = txt_mod_param_vec[2];
 
             auto [img_attn_output, txt_attn_output] = attn->forward(ctx, img_modulated, txt_modulated, pe);
@@ -363,11 +363,11 @@ namespace Qwen {
             txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn_output, txt_gate1));
 
             auto img_normed2    = img_norm2->forward(ctx, img);
-            auto img_modulated2 = Flux::modulate(ctx->ggml_ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4], modulate_index != nullptr);
+            auto img_modulated2 = Flux::modulate(ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4], modulate_index != nullptr);
             auto img_gate2      = img_mod_param_vec[5];
 
             auto txt_normed2    = txt_norm2->forward(ctx, txt);
-            auto txt_modulated2 = Flux::modulate(ctx->ggml_ctx, txt_normed2, txt_mod_param_vec[3], txt_mod_param_vec[4]);
+            auto txt_modulated2 = Flux::modulate(ctx, txt_normed2, txt_mod_param_vec[3], txt_mod_param_vec[4]);
             auto txt_gate2      = txt_mod_param_vec[5];
 
             auto img_mlp_out = img_mlp->forward(ctx, img_modulated2);
@@ -407,7 +407,7 @@ namespace Qwen {
             auto shift = mods[1];
 
             x = norm->forward(ctx, x);
-            x = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            x = Flux::modulate(ctx, x, shift, scale);
 
             return x;
         }

src/model/diffusion/z_image.hpp

@@ -294,17 +294,16 @@ namespace ZImage {
         }
     };
 
-    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* modulate(GGMLRunnerContext* ctx,
                                             ggml_tensor* x,
                                             ggml_tensor* scale) {
         // x: [N, L, C]
         // scale: [N, C], or [N, L, C] when the caller modulates per token (LLaDA-Image editing
         // feeds a per-token timestep embedding so each segment carries its own modulation).
         if (scale->ne[1] != x->ne[1]) {
-            scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
+            scale = ggml_reshape_3d(ctx->ggml_ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
         }
-        x = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        return ggml_ext_modulate(ctx, x, nullptr, scale);
-        return x;
     }
 
     struct JointTransformerBlock : public GGMLBlock {
@@ -361,14 +360,14 @@ namespace ZImage {
                 auto gate_mlp  = mods[3];
 
                 auto residual = x;
-                x             = modulate(ctx->ggml_ctx, attention_norm1->forward(ctx, x), scale_msa);
+                x             = modulate(ctx, attention_norm1->forward(ctx, x), scale_msa);
                 x             = attention->forward(ctx, x, pe, mask);
                 x             = attention_norm2->forward(ctx, x);
                 x             = ggml_mul(ctx->ggml_ctx, x, ggml_tanh(ctx->ggml_ctx, gate_msa));
                 x             = ggml_add(ctx->ggml_ctx, x, residual);
 
                 residual = x;
-                x        = modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, x), scale_mlp);
+                x        = modulate(ctx, ffn_norm1->forward(ctx, x), scale_mlp);
                 x        = feed_forward->forward(ctx, x);
                 x        = ffn_norm2->forward(ctx, x);
                 x        = ggml_mul(ctx->ggml_ctx, x, ggml_tanh(ctx->ggml_ctx, gate_mlp));
@@ -505,14 +504,14 @@ namespace ZImage {
             auto gate_mlp  = mods[3];
 
             auto residual = c;
-            c             = modulate(ctx->ggml_ctx, attention_norm1->forward(ctx, c), scale_msa);
+            c             = modulate(ctx, attention_norm1->forward(ctx, c), scale_msa);
             c             = forward_attention(ctx, c, pe);
             c             = attention_norm2->forward(ctx, c);
             c             = ggml_mul(ctx->ggml_ctx, c, ggml_tanh(ctx->ggml_ctx, gate_msa));
             c             = ggml_add(ctx->ggml_ctx, c, residual);
 
             residual = c;
-            c        = modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, c), scale_mlp);
+            c        = modulate(ctx, ffn_norm1->forward(ctx, c), scale_mlp);
             c        = feed_forward->forward(ctx, c);
             c        = ffn_norm2->forward(ctx, c);
             c        = ggml_mul(ctx->ggml_ctx, c, ggml_tanh(ctx->ggml_ctx, gate_mlp));
@@ -544,7 +543,7 @@ namespace ZImage {
 
             auto scale = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, hidden_size]
             x          = norm_final->forward(ctx, x);
-            x          = modulate(ctx->ggml_ctx, x, scale);
+            x          = modulate(ctx, x, scale);
             x          = linear->forward(ctx, x);
 
             return x;