refactor: use custom RoPE Apply nodes exclusively

docs/ggml_extensions.md

@@ -136,13 +136,15 @@ in `cpu/registration.cpp`. CUDA-style and Vulkan implementations live in their
 respective backend directories.
 
 `src/core/cpu/rope.cpp` registers the CPU implementation of
-`sd.rope_apply.v1`, an override for existing `GGML_OP_ROPE_APPLY` nodes, and a
+`sd.rope_apply.v1` and a
 `RMSNorm -> Mul -> RoPE` fusion. The fusion accepts a one-dimensional weight
 vector. Both interleaved and split-half RoPE, partial rotation, and strided input
 rows are supported. Model construction uses the custom operator on CPU, CUDA,
-HIP, MUSA, and Vulkan, and retains existing backend implementations elsewhere.
+HIP, MUSA, and Vulkan. Other backends can use the scheduler's CPU fallback.
+The former `ggml_rope_apply` API, builtin opcode, and CPU implementation have
+been removed from ggml.
 
-`src/core/cuda/rope.cu` supplies standalone CUDA RoPE, the same builtin override,
+`src/core/cuda/rope.cu` supplies standalone CUDA RoPE
 and `RMSNorm -> Mul -> RoPE [-> Scale -> FP16 Copy]` fusion. CuTe attention shares
 its RMSNorm/RoPE kernel. Partial rotation and arbitrary float-aligned input and
 position-encoding strides are supported by the standalone kernel; RMSNorm fusion
@@ -180,7 +182,7 @@ Independent affine operands, including broadcasted constants, are computed befor
 their normalization/projection
 chain when graph dependencies allow it. The attention optimizer can then move
 the complete five-node projection when safe.
-These rules accept both builtin and custom RoPE nodes. They preserve the original
+These rules accept custom RoPE nodes. They preserve the original
 large-matrix selection thresholds, full interleaved head dimension 128, packed
 activation layout, and separate weight scales. External consumers prevent fusion;
 aliasing terminal copies use scheduler workspace. The CUDA projection kernels
@@ -379,7 +381,7 @@ Callback errors propagate without retry. Evaluation callbacks, direct backend
 calls, synchronous devices, serialized diagnostics, and timestamp profiling
 retain fragment execution.
 
-`src/core/vulkan/rope.cpp` registers custom and builtin RoPE implementations,
+`src/core/vulkan/rope.cpp` registers the custom RoPE implementation,
 plus `RMSNorm -> Mul -> RoPE` and `RMSNorm -> Mul -> RoPE -> Scale -> FP16 Copy`
 fusion. These kernels retain the original strides, partial-rotation behavior,
 128-channel subgroup reduction, and separate arithmetic rounding. Their SPIR-V
@@ -390,8 +392,8 @@ External intermediate consumers retain ordinary operators, and overlapping
 outputs use plan workspace. `GGML_VK_DISABLE_FUSION` is honored.
 
 The corresponding Vulkan shader sources and native fusion rules have been
-removed from ggml. Existing builtin RoPE graphs use the registered override;
-unsupported cases retain scheduler fallback to another supported backend.
+removed from ggml. Unsupported custom nodes retain scheduler fallback to
+another supported backend.
 
 Remaining model-specific CUDA paths and Vulkan INT8/ConvRot implementations still
 need auditing or migration.

src/core/cpu/rope.cpp

@@ -23,7 +23,7 @@ namespace sd_cpu {
         const auto* x         = execution.norm ? execution.norm->src[0] : rope->src[0];
         const auto* pe        = rope->src[1];
         int32_t interleaved;
-        std::memcpy(&interleaved, ggml_ext_op_get(rope) ? ggml_ext_op_attributes(rope) : rope->op_params, sizeof(interleaved));
+        std::memcpy(&interleaved, ggml_ext_op_attributes(rope), sizeof(interleaved));
         float eps = 0.0f;
         if (execution.norm) {
             std::memcpy(&eps, execution.norm->op_params, sizeof(eps));
@@ -90,7 +90,7 @@ namespace sd_cpu {
         if (norm->op != GGML_OP_RMS_NORM || norm->type != GGML_TYPE_F32 || !norm->src[0] ||
             norm->src[0]->type != GGML_TYPE_F32 || norm->src[0]->nb[0] != sizeof(float) ||
             mul->op != GGML_OP_MUL || mul->type != GGML_TYPE_F32 || rope->src[0] != mul ||
-            (rope->op != GGML_OP_ROPE_APPLY && ggml_ext_op_get(rope) != sd_ggml_rope_op()) ||
+            ggml_ext_op_get(rope) != sd_ggml_rope_op() ||
             !sd_ggml_rope_describe(rope) || !ggml_are_same_shape(norm, mul)) {
             return 0;
         }
@@ -126,13 +126,9 @@ namespace sd_cpu {
         desc.execute              = rope_execute;
         GGML_ASSERT(ggml_ext_register_kernel(device, &desc));
 
-        desc.name      = "sd.rope_apply.cpu.override.v1";
-        desc.op        = GGML_OP_ROPE_APPLY;
-        desc.custom_op = nullptr;
-        GGML_ASSERT(ggml_ext_register_kernel(device, &desc));
-
         desc.name      = "sd.rms_norm_mul_rope.cpu.v1";
         desc.op        = GGML_OP_RMS_NORM;
+        desc.custom_op = nullptr;
         desc.priority  = 200;
         desc.supports  = nullptr;
         desc.match     = rope_fusion_match;

src/core/cuda/i8_projection.cu

@@ -211,7 +211,7 @@ bool sd_cuda_i8_rope_supports(ggml_backend_dev_t device, const ggml_tensor* mat,
         !sd_cuda::rope_fusion_supports(norm, mul, rope) || rope->src[1]->ne[2] != 64 || !aligned(mul->src[1]) || !aligned(rope->src[1]))
         return false;
     int32_t interleaved;
-    std::memcpy(&interleaved, ggml_ext_op_get(rope) ? ggml_ext_op_attributes(rope) : rope->op_params, sizeof(interleaved));
+    std::memcpy(&interleaved, ggml_ext_op_attributes(rope), sizeof(interleaved));
     return interleaved != 0;
 }
 

src/core/cuda/rope.cu

@@ -132,7 +132,7 @@ namespace SD_CUDA_NAMESPACE {
 
     static bool rope_interleaved(const ggml_tensor* rope) {
         int32_t value;
-        std::memcpy(&value, ggml_ext_op_get(rope) ? ggml_ext_op_attributes(rope) : rope->op_params, sizeof(value));
+        std::memcpy(&value, ggml_ext_op_attributes(rope), sizeof(value));
         return value != 0;
     }
 
@@ -143,7 +143,7 @@ namespace SD_CUDA_NAMESPACE {
     }
 
     static bool rope_is_node(const ggml_tensor* node) {
-        return node && (node->op == GGML_OP_ROPE_APPLY || ggml_ext_op_get(node) == sd_ggml_rope_op());
+        return node && ggml_ext_op_get(node) == sd_ggml_rope_op();
     }
 
     static bool rope_supports(ggml_backend_dev_t, const ggml_tensor* rope, void*) {
@@ -324,12 +324,9 @@ namespace SD_CUDA_NAMESPACE {
         desc.destroy              = rope_destroy;
         desc.execute              = rope_execute;
         GGML_ASSERT(ggml_ext_register_kernel_ex(device, &desc, GGML_EXT_KERNEL_CAPTURE_SAFE | GGML_EXT_KERNEL_REPLAY_REUSABLE));
-        desc.name      = "sd.rope_apply." SD_CUDA_NAME ".override.v1";
-        desc.op        = GGML_OP_ROPE_APPLY;
-        desc.custom_op = nullptr;
-        GGML_ASSERT(ggml_ext_register_kernel_ex(device, &desc, GGML_EXT_KERNEL_CAPTURE_SAFE | GGML_EXT_KERNEL_REPLAY_REUSABLE));
         desc.name      = "sd.rms_norm_mul_rope." SD_CUDA_NAME ".v1";
         desc.op        = GGML_OP_RMS_NORM;
+        desc.custom_op = nullptr;
         desc.priority  = 250;
         desc.supports  = nullptr;
         desc.match     = rope_match;

src/core/ggml_extensions.cpp

@@ -265,20 +265,10 @@ void sd_ggml_register_extensions(ggml_backend_t backend) {
 
 ggml_tensor* sd_ggml_rope_apply(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* x, ggml_tensor* pe, bool interleaved) {
     sd_ggml_register_extensions(backend);
-    if (backend) {
-        auto reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));
-        if (ggml_backend_reg_get_proc_address(reg, "ggml_ext_cpu_parallel_v1")
-#ifdef SD_BACKEND_EXTENSIONS
-            || ggml_backend_reg_get_proc_address(reg, "ggml_ext_gpu_context_v1") || ggml_backend_reg_get_proc_address(reg, "ggml_ext_vulkan_v1")
-#endif
-        ) {
     ggml_tensor* inputs[] = {x, pe};
     int32_t attributes    = interleaved ? 1 : 0;
     return ggml_ext_op_create(ctx, sd_ggml_rope_op(), inputs, 2, &attributes);
 }
-    }
-    return ggml_rope_apply(ctx, x, pe, interleaved);
-}
 
 #else
 void sd_ggml_register_extensions(ggml_backend_t) {}

src/core/ggml_extensions.h

@@ -10,6 +10,12 @@ void sd_ggml_register_extensions(ggml_backend_t backend);
 
 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:
+// a: F32 [head_dim, heads, tokens, batch], pe: F32 [2, 2, rot_dim/2, tokens].
+// pe stores [[cos, -sin], [sin, cos]] for each pair and token, shared across heads/batches.
+// Returns contiguous F32 [head_dim, tokens, heads*batch]. Input strides are preserved.
+// Interleaved pairs are (2*i, 2*i+1); otherwise pairs are (i, i+rot_dim/2).
+// Channels from rot_dim to head_dim are copied unchanged.
 ggml_tensor* sd_ggml_rope_apply(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* x, ggml_tensor* pe, bool interleaved);
 
 struct SDI8Operation {

src/core/vulkan/rope.cpp

@@ -32,7 +32,7 @@ namespace sd_vulkan_rope {
         return value;
     }
     static bool is_rope(const ggml_tensor* node) {
-        return node && (node->op == GGML_OP_ROPE_APPLY || ggml_ext_op_get(node) == sd_ggml_rope_op());
+        return node && ggml_ext_op_get(node) == sd_ggml_rope_op();
     }
     static bool aligned(const ggml_tensor* tensor, size_t alignment) {
         size_t offset = 0;
@@ -166,7 +166,7 @@ namespace sd_vulkan_rope {
         pc.x_offset   = buffers[0].shader_offset / 4;
         pc.pe_offset  = buffers[1].shader_offset / 4;
         pc.dst_offset = buffers[2].shader_offset / uint32_t(ggml_type_size(output->type));
-        std::memcpy(&pc.interleaved, ggml_ext_op_get(rope) ? ggml_ext_op_attributes(rope) : rope->op_params, 4);
+        std::memcpy(&pc.interleaved, ggml_ext_op_attributes(rope), 4);
         uint32_t groups[3]                = {1, 1, 1};
         ggml_ext_vulkan_pipeline pipeline = {};
         pipeline.name                     = "sd.rope_apply.vulkan";
@@ -213,12 +213,9 @@ void sd_ggml_register_vulkan_rope(ggml_backend_dev_t device) {
     desc.execute              = sd_vulkan_rope::execute;
     desc.destroy              = sd_vulkan_rope::destroy;
     GGML_ASSERT(ggml_ext_register_kernel(device, &desc));
-    desc.name      = "sd.rope_apply.vulkan.override.v1";
-    desc.op        = GGML_OP_ROPE_APPLY;
-    desc.custom_op = nullptr;
-    GGML_ASSERT(ggml_ext_register_kernel(device, &desc));
     desc.name      = "sd.rms_rope_apply.vulkan.v1";
     desc.op        = GGML_OP_RMS_NORM;
+    desc.custom_op = nullptr;
     desc.priority  = 200;
     desc.supports  = nullptr;
     desc.match     = sd_vulkan_rope::match;