whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -53,6 +53,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/gated_delta_net.cuh"
 #include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
@@ -70,17 +71,18 @@
 #include <condition_variable>
 #include <cstddef>
 #include <cstdint>
-#include <float.h>
+#include <cfloat>
 #include <initializer_list>
 #include <limits>
 #include <map>
 #include <memory>
 #include <mutex>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
 #include <string>
 #include <vector>
+#include <unordered_set>
 
 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
@@ -122,7 +124,10 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
         err = cudaMallocManaged(ptr, size);
 #if defined(GGML_USE_HIP)
         if (err == hipSuccess) {
-            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+            // hipMemAdviseSetCoarseGrain is an optional performance hint;
+            // ignore errors (e.g. hipErrorInvalidValue on some APU/iGPU configs).
+            cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device);
+            (void)hipGetLastError(); // clear any error
         }
 
         // fall back to cudaMalloc if not supported (e.g. on Windows)
@@ -203,7 +208,14 @@ static ggml_cuda_device_info ggml_cuda_init() {
     GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);
 
     int64_t total_vram = 0;
-    GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+    for (int id = 0; id < info.device_count; ++id) {
+        cudaDeviceProp prop;
+        CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
+        total_vram += prop.totalGlobalMem;
+    }
+    GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices (Total VRAM: %zu MiB):\n",
+                  __func__, info.device_count, (size_t)(total_vram / (1024 * 1024)));
+    total_vram = 0;
 
     std::vector<std::pair<int, std::string>> turing_devices_without_mma;
     for (int id = 0; id < info.device_count; ++id) {
@@ -241,6 +253,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
 #else
         info.devices[id].supports_cooperative_launch = false;
 #endif // !(GGML_USE_MUSA)
+
 #if defined(GGML_USE_HIP)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
@@ -255,22 +268,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
                 info.devices[id].cc += prop.minor * 0x10;
             }
         }
-        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d\n",
+        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d, VRAM: %zu MiB\n",
                       id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
-                      device_vmm ? "yes" : "no", prop.warpSize);
+                      device_vmm ? "yes" : "no", prop.warpSize,
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)));
 #elif defined(GGML_USE_MUSA)
         // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
         info.devices[id].warp_size = 32;
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100;
         info.devices[id].cc += prop.minor * 0x10;
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
-                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB\n",
+                      id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)));
 #else
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
-                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB\n",
+                      id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)));
         std::string device_name(prop.name);
         if (device_name == "NVIDIA GeForce MX450") {
             turing_devices_without_mma.push_back({ id, device_name });
@@ -285,6 +301,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         // TODO: Check for future drivers the default scheduling strategy and
         // remove this call again when cudaDeviceScheduleSpin is default.
         if (prop.major == 12 && prop.minor == 1) {
+            CUDA_CHECK(cudaSetDevice(id));
             CUDA_CHECK(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
         }
 
@@ -1224,6 +1241,34 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     }
 }
 
+struct cublas_force_compute_type {
+    bool fp32 = false;
+    bool fp16 = false;
+};
+
+static const cublas_force_compute_type & ggml_cuda_cublas_get_force_compute_type() {
+    static const cublas_force_compute_type compute_type = [] {
+        cublas_force_compute_type result;
+
+        const bool ggml_cuda_force_cublas_compute_32f_env = getenv("GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F") != nullptr;
+        const bool ggml_cuda_force_cublas_compute_16f_env = getenv("GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F") != nullptr;
+
+        GGML_ASSERT(ggml_cuda_force_cublas_compute_16f_env == false || ggml_cuda_force_cublas_compute_32f_env == false);
+
+        if (ggml_cuda_force_cublas_compute_32f_env) {
+            GGML_LOG_INFO("Detected GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F\n");
+            result.fp32 = true;
+        } else if (ggml_cuda_force_cublas_compute_16f_env) {
+            GGML_LOG_INFO("Detected GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F\n");
+            result.fp16 = true;
+        }
+
+        return result;
+    }();
+
+    return compute_type;
+}
+
 static void ggml_cuda_op_mul_mat_cublas(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
@@ -1306,7 +1351,13 @@ static void ggml_cuda_op_mul_mat_cublas(
 
         CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
 
-        if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+        const auto & force_compute_type = ggml_cuda_cublas_get_force_compute_type();
+
+        if (!force_compute_type.fp16 && (GGML_CUDA_CC_IS_CDNA(cc)
+                                        || GGML_CUDA_CC_IS_RDNA4(cc)
+                                        || cc == GGML_CUDA_CC_VOLTA
+                                        || force_compute_type.fp32))
+        {
             const float alpha = 1.0f;
             const float beta = 0.0f;
             CUBLAS_CHECK(
@@ -1905,10 +1956,23 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
     cudaDataType_t cu_data_type_b = traits::data_type;
     const void * alpha = traits::get_alpha();
     const void * beta = traits::get_beta();
-    const float alpha_f32 = 1.0f;
-    const float beta_f32 = 0.0f;
 
-    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+    const auto & force_compute_type = ggml_cuda_cublas_get_force_compute_type();
+
+    int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+    static constexpr bool is_src0_type_f16 = src0_type == GGML_TYPE_F16;
+
+    // bf16 and fp32 are already being computed in fp32 (ensure it using static_assert),
+    // so checking necessity of forced fp32 only for fp16 src0_type
+    static_assert(is_src0_type_f16 || traits::compute_type == CUBLAS_COMPUTE_32F);
+
+    const bool need_compute_32f = is_src0_type_f16 && !force_compute_type.fp16 && (GGML_CUDA_CC_IS_CDNA(cc)
+                                                                                  || GGML_CUDA_CC_IS_RDNA4(cc)
+                                                                                  || cc == GGML_CUDA_CC_VOLTA
+                                                                                  || force_compute_type.fp32);
+
+    if (dst->op_params[0] == GGML_PREC_DEFAULT && !need_compute_32f) {
         if constexpr (src0_type == GGML_TYPE_F32) {
             dst_t = (char *) dst_ddf;  // Direct F32 output
         } else {
@@ -1918,18 +1982,10 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
         }
     } else {
         dst_t = (char *) dst_ddf;
-        cu_compute_type = CUBLAS_COMPUTE_32F;
-        cu_data_type = CUDA_R_32F;
-        alpha = &alpha_f32;
-        beta = &beta_f32;
-    }
-
-    int id = ggml_cuda_get_device();
-    const int cc = ggml_cuda_info().devices[id].cc;
-    if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
-        cu_compute_type = CUBLAS_COMPUTE_32F;
-        alpha = &alpha_f32;
-        beta = &beta_f32;
+        cu_compute_type = batched_mul_mat_traits<GGML_TYPE_F32>::compute_type;
+        cu_data_type = batched_mul_mat_traits<GGML_TYPE_F32>::data_type;
+        alpha = batched_mul_mat_traits<GGML_TYPE_F32>::get_alpha();
+        beta = batched_mul_mat_traits<GGML_TYPE_F32>::get_beta();
     }
 
     GGML_ASSERT(ne12 % ne02 == 0);
@@ -2277,14 +2333,21 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
+    // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
     if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        if (ne2 == 1) {
+        static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
+        if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
             if (ggml_is_quantized(src0->type)) {
-                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                if (ne2 <= MMVQ_MMID_MAX_BATCH_SIZE) {
+                    ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                    return;
+                }
             } else {
-                ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                if (GGML_CUDA_CC_IS_AMD(cc)) {
+                    ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                    return;
+                }
             }
-            return;
         }
 
         if (ggml_cuda_should_use_mmq(src0->type, cc, ne12, /*n_experts=*/ne02)) {
@@ -2298,6 +2361,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         }
     }
 
+    // note: this path should not be reached when recording CUDA graphs, because it requires stream synchronization
+    // TODO: add asserts to verify this. should work with CUDA, HIP, etc.
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(nb12 % nb11 == 0);
@@ -2723,6 +2788,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GATED_LINEAR_ATTN:
             ggml_cuda_op_gated_linear_attn(ctx, dst);
             break;
+        case GGML_OP_GATED_DELTA_NET:
+            ggml_cuda_op_gated_delta_net(ctx, dst);
+            break;
         case GGML_OP_RWKV_WKV7:
             ggml_cuda_op_rwkv_wkv7(ctx, dst);
             break;
@@ -2858,14 +2926,6 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
     bool use_cuda_graph = true;
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
 
-    const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
-    const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
-    const std::string ffn_moe_gate_bias_prefix = "ffn_moe_gate_biased";
-    const std::string ffn_moe_up_bias_prefix = "ffn_moe_up_biased";
-    const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
-    const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out";
-    const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d";
-
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
@@ -2880,30 +2940,14 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 #endif
         }
 
-        if (node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
-            use_cuda_graph = false; // This node type is not supported by CUDA graph capture
-#ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
-#endif
-        }
-
-        if (node->op == GGML_OP_ADD &&
-            node->src[1] && node->src[1]->ne[1] > 1 &&
-            (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) &&
-            (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true) &&
-            strncmp(node->name, ffn_moe_gate_bias_prefix.c_str(), ffn_moe_gate_bias_prefix.size()) != 0 &&
-            strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
-            strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 &&
-            strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 &&
-            strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0) {
-            // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
-            // by means of matching node names. See
-            // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
-            // https://github.com/huggingface/transformers/blob/bda75b4011239d065de84aa3e744b67ebfa7b245/src/transformers/models/gemma3n/modeling_gemma3n.py#L1773,
-            // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.
+        // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
+        if (node->op == GGML_OP_MUL_MAT_ID && (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > MMVQ_MMID_MAX_BATCH_SIZE)) {
+            // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
+            // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+            // ref: https://github.com/ggml-org/llama.cpp/pull/18958
             use_cuda_graph = false;
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
         }
 
@@ -2916,21 +2960,27 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 }
 
 static void ggml_cuda_graph_node_set_properties(ggml_cuda_graph_node_properties * props, ggml_tensor * node) {
-    props->node_address = node->data;
+    memset(props, 0, sizeof(ggml_cuda_graph_node_properties));
+    props->node_data = node->data;
     props->node_op = node->op;
+    props->node_type = node->type;
+    props->flags = node->flags;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         props->ne[i] = node->ne[i];
         props->nb[i] = node->nb[i];
     }
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        props->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
+        if (!node->src[i]) {
+            continue;
+        }
+
+        props->src_data[i] = node->src[i]->data;
     }
     memcpy(props->op_params, node->op_params, GGML_MAX_OP_PARAMS);
 }
 
 static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_graph_node_properties * props) {
-    if (node->data != props->node_address &&
-          node->op != GGML_OP_VIEW) {
+    if (node->data != props->node_data && node->op != GGML_OP_VIEW) {
         return false;
     }
 
@@ -2938,6 +2988,10 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         return false;
     }
 
+    if (node->type != props->node_type) {
+        return false;
+    }
+
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (node->ne[i] != props->ne[i]) {
             return false;
@@ -2947,73 +3001,104 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         }
     }
 
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (node->src[i] &&
-            node->src[i]->data != props->src_address[i] &&
-            node->op != GGML_OP_VIEW
-        ) {
-            return false;
+    if (node->op != GGML_OP_VIEW) {
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (!node->src[i]) {
+                if (props->src_data[i] != nullptr) {
+                    return false;
+                }
+                continue;
+            }
+
+            if (node->src[i]->data != props->src_data[i]) {
+                return false;
+            }
         }
     }
 
-    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
-        memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+    if (memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+        return false;
+    }
+
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) != (props->flags & GGML_TENSOR_FLAG_COMPUTE)) {
         return false;
     }
 
     return true;
 }
 
-static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
+static const void * ggml_cuda_graph_get_key(ggml_cgraph * cgraph) {
+    return cgraph->nodes[0];
+}
 
+static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
     bool res = false;
 
-    if (cuda_ctx->cuda_graph->instance == nullptr) {
-        res = true;
-    }
+    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
     // Check if the graph size has changed
-    if (cuda_ctx->cuda_graph->props.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
+    if (graph->props.size() != (size_t)cgraph->n_nodes) {
         res = true;
-        cuda_ctx->cuda_graph->props.resize(cgraph->n_nodes + cgraph->n_leafs);
+        graph->props.resize(cgraph->n_nodes);
     }
 
     // Loop over nodes in GGML graph to determine if CUDA graph update is required
     // and store properties to allow this comparison for the next token
+    std::unordered_set<ggml_tensor *> seen_node;
+    std::vector<ggml_tensor *> srcs_extra;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         bool props_match = true;
+
+        seen_node.insert(cgraph->nodes[i]);
+
         if (!res) {
-            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &cuda_ctx->cuda_graph->props[i]);
+            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &graph->props[i]);
         }
         if (!props_match) {
             res = true;
         }
-        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[i], cgraph->nodes[i]);
+        ggml_cuda_graph_node_set_properties(&graph->props[i], cgraph->nodes[i]);
+
+        for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+            ggml_tensor * src = cgraph->nodes[i]->src[src_idx];
+            if (src && seen_node.find(src) == seen_node.end()) {
+                srcs_extra.push_back(src);
+            }
+        }
     }
 
-    for (int i = 0; i < cgraph->n_leafs; i++) {
-        bool props_match= true;
+    if (graph->extra.size() != (size_t) srcs_extra.size()) {
+        res = true;
+        graph->extra.resize(srcs_extra.size());
+    }
+
+    for (size_t i = 0; i < srcs_extra.size(); ++i) {
+        bool props_match = true;
+
         if (!res) {
-            props_match = ggml_cuda_graph_node_properties_match(cgraph->leafs[i], &cuda_ctx->cuda_graph->props[cgraph->n_nodes + i]);
+            props_match = ggml_cuda_graph_node_properties_match(srcs_extra[i], &graph->extra[i]);
         }
+
         if (!props_match) {
             res = true;
         }
-        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[cgraph->n_nodes + i], cgraph->leafs[i]);
+        ggml_cuda_graph_node_set_properties(&graph->extra[i], srcs_extra[i]);
     }
 
     return res;
 }
 
-static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx) {
+static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
 #if CUDART_VERSION >= 12000
     cudaGraphExecUpdateResultInfo result_info;
-    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &result_info);
 #else
     cudaGraphNode_t errorNode;
     cudaGraphExecUpdateResult result_info;
-    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &errorNode, &result_info);
 #endif // CUDART_VERSION >= 12000
 
     if (stat == cudaErrorGraphExecUpdateFailure) {
@@ -3024,14 +3109,14 @@ static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_c
         // The pre-existing graph exec cannot be updated due to violated constraints
         // so instead clear error and re-instantiate
         (void)cudaGetLastError();
-        CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
-        cuda_ctx->cuda_graph->instance = nullptr;
-        CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        CUDA_CHECK(cudaGraphExecDestroy(graph->instance));
+        graph->instance = nullptr;
+        CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
     } else {
         GGML_ASSERT(stat == cudaSuccess);
     }
 }
-#endif
+#endif // USE_CUDA_GRAPH
 
 static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
                                                 const ggml_tensor * view,
@@ -3067,63 +3152,166 @@ static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
     return true;
 }
 
-static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops, std::initializer_list<enum ggml_unary_op> unary_ops) {
-#ifndef NDEBUG
-    const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
-    GGML_ASSERT(unary_ops.size() == num_unary);
-#endif
+static bool ggml_cuda_topk_moe_fusion(const struct ggml_cgraph * cgraph, int node_idx, ggml_cuda_topk_moe_args & args) {
+    args.sigmoid         = false;
+    args.softmax         = false;
+    args.delayed_softmax = false;
+    args.prob_bias       = false;
+    args.norm            = false;
 
-    //TODO: remove special case once ggml_can_fuse can handle empty nodes
-    std::initializer_list<enum ggml_op> topk_moe_ops =
-        ggml_cuda_topk_moe_ops(/*with_norm*/ false, /*delayed_softmax=*/false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm =
-        ggml_cuda_topk_moe_ops(/*with_norm=*/true, /*delayed_softmax=*/false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_delayed_softmax =
-        ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
+    const int      n_nodes = cgraph->n_nodes;
+    ggml_tensor ** nodes   = cgraph->nodes;
 
-    const auto is_equal = [](const std::initializer_list<enum ggml_op> & list1,
-                             const std::initializer_list<enum ggml_op> & list2) {
-        return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end());
-    };
-
-    if (is_equal(topk_moe_ops_with_norm, ops) &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
-        ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx + 9];
-        ggml_tensor * get_rows = cgraph->nodes[node_idx + 4];
-        ggml_tensor * argsort = cgraph->nodes[node_idx + 2];
-        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
+    if (nodes[node_idx]->op == GGML_OP_SOFT_MAX) {
+        args.softmax = true;
+    }
 
-        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
-            return true;
+    if (nodes[node_idx]->op == GGML_OP_UNARY) {
+        if (ggml_get_unary_op(nodes[node_idx]) != GGML_UNARY_OP_SIGMOID) {
+            return false;
         }
+        args.sigmoid = true;
     }
 
-    if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
-        ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx + 4];
-        ggml_tensor * get_rows = cgraph->nodes[node_idx + 4];
-        ggml_tensor * argsort = cgraph->nodes[node_idx + 2];
-        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
+    if (nodes[node_idx]->op == GGML_OP_ARGSORT) {
+        args.delayed_softmax = true;
+    }
 
-        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
-            return true;
+    node_idx++;
+
+    if (args.sigmoid || args.softmax) {
+        // SOFTMAX -> RESHAPE
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_RESHAPE ||
+                nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        }
+        ggml_tensor * probs_reshaped = nodes[node_idx];
+        node_idx++;
+
+        if (node_idx >= n_nodes) {
+            return false;
+        }
+
+        // src of bias add is the unreshaped probs (-2 instead of -1)
+        if (nodes[node_idx]->op == GGML_OP_ADD && nodes[node_idx]->src[0] == nodes[node_idx - 2]) {
+            args.prob_bias = true;
+            node_idx++;
+        }
+        // RESHAPE/ADD -> ARGSORT
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_ARGSORT) {
+            return false;
+        }
+
+        if (args.prob_bias && nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        } else if (!args.prob_bias && nodes[node_idx]->src[0] != nodes[node_idx - 2]) {
+            return false;
+        }
+
+        node_idx++;
+
+        // ARGSORT-> VIEW
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_VIEW ||
+                nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        }
+        node_idx++;
+
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_GET_ROWS) {
+            return false;
+        }
+
+        // GET_ROWS
+        if (nodes[node_idx]->src[0] != probs_reshaped || nodes[node_idx]->src[1] != nodes[node_idx - 1]) {
+            return false;
+        }
+        node_idx++;
+    } else if (args.delayed_softmax) {
+        if (node_idx - 2 < 0) {
+            return false;
         }
+        ggml_tensor * probs_reshaped = nodes[node_idx - 2];
+
+        // VIEW->ARGSORT
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_VIEW ||
+            nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        }
+        node_idx++;
+
+        // GET_ROWS
+        if (node_idx >= n_nodes || nodes[node_idx]->src[1] != nodes[node_idx - 1] ||
+                nodes[node_idx]->src[0] != probs_reshaped) {
+            return false;
+        }
+        node_idx++;
+
+        static const std::vector<ggml_op> remaining_ops = { GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE };
+
+        for (const ggml_op op : remaining_ops) {
+            if (node_idx >= n_nodes || nodes[node_idx]->op != op || nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+                return false;
+            }
+            node_idx++;
+        }
+    }
+
+    // At this point we can check for norm + scale. Everything is now at least valid till the norm
+    if (node_idx >= n_nodes) {
+        return true;
     }
 
-    if (is_equal(topk_moe_ops_delayed_softmax, ops) &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
-        ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
-        ggml_tensor * weights = cgraph->nodes[node_idx + 5];
-        ggml_tensor * get_rows = cgraph->nodes[node_idx + 2];
-        ggml_tensor * argsort = cgraph->nodes[node_idx + 0];
-        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
+    if (nodes[node_idx]->op == GGML_OP_RESHAPE) {
+        //check RESHAPE->SUM_ROWS->CLAMP->DIV->RESHAPE
+        static const std::vector<ggml_op> norm_ops = { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP };
 
-        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
+        args.norm = true;
+        for (const ggml_op op : norm_ops) {
+            if (nodes[node_idx]->op == op && nodes[node_idx]->src[0] == nodes[node_idx - 1]) {
+                node_idx++;
+            } else {
+                args.norm = false;
+                return true;
+            }
+        }
+
+        // DIV <- CLAMP, RESHAPE
+        if (nodes[node_idx]->op != GGML_OP_DIV || nodes[node_idx]->src[1] != nodes[node_idx - 1] ||
+            nodes[node_idx]->src[0] != nodes[node_idx - 3]) {
+            args.norm = false;
             return true;
         }
+        node_idx++;
+
+        if (nodes[node_idx]->op != GGML_OP_RESHAPE || nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            args.norm = false;
+            return true;
+        }
+
+        node_idx++;
+    }
+
+    if (nodes[node_idx]->op == GGML_OP_SCALE && nodes[node_idx]->src[0] == nodes[node_idx - 1]) {
+        args.scale = true;
     }
 
+    return true;
+}
+
+static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
+                               int                                       node_idx,
+                               std::initializer_list<enum ggml_op>       ops,
+                               std::initializer_list<enum ggml_unary_op> unary_ops) {
+#ifndef NDEBUG
+    const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
+    GGML_ASSERT(unary_ops.size() == num_unary);
+#endif
+
+    const auto is_equal = [](const std::initializer_list<enum ggml_op> & list1,
+                             const std::initializer_list<enum ggml_op> & list2) {
+        return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end());
+    };
+
     std::initializer_list<enum ggml_op> mul_mat_bias_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_GLU };
     std::initializer_list<enum ggml_op> mul_mat_id_bias_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
 
@@ -3200,7 +3388,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
             return false;
         }
 
-        //rms_norm kernel assumes contigous rows
+        //rms_norm kernel assumes contiguous rows
         if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
             return false;
         }
@@ -3212,6 +3400,46 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         return true;
     }
 
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_SSM_CONV && ops.begin()[1] == GGML_OP_UNARY
+     && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {
+        const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];
+        const ggml_tensor * silu     = cgraph->nodes[node_idx+1];
+
+        if (ssm_conv->type != GGML_TYPE_F32 || silu->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        return true;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_UNARY && ops.begin()[1] == GGML_OP_MUL
+     && unary_ops.size() == 1 && (unary_ops.begin()[0] == GGML_UNARY_OP_SILU || unary_ops.begin()[0] == GGML_UNARY_OP_SIGMOID || unary_ops.begin()[0] == GGML_UNARY_OP_SOFTPLUS)) {
+        const ggml_tensor * unary = cgraph->nodes[node_idx];
+        const ggml_tensor * mul   = cgraph->nodes[node_idx+1];
+
+        if (ggml_get_unary_op(unary) != unary_ops.begin()[0]) {
+            return false;
+        }
+
+        if (unary->type != GGML_TYPE_F32 && unary->type != GGML_TYPE_F16) {
+            return false;
+        }
+
+        if (unary->type != mul->type) {
+            return false;
+        }
+
+        const ggml_tensor * other = (mul->src[0] == unary) ? mul->src[1] : mul->src[0];
+        if (other->type != unary->type) {
+            return false;
+        }
+        if (!ggml_is_contiguous_1(other) || !ggml_is_contiguous_1(unary->src[0]) || !ggml_are_same_shape(other, unary)) {
+            return false;
+        }
+
+        return true;
+    }
+
     if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SCALE && ops.begin()[1] == GGML_OP_UNARY && ops.begin()[2] == GGML_OP_SCALE
      && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_TANH) {
         const ggml_tensor *scale  = cgraph->nodes[node_idx];
@@ -3236,7 +3464,70 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     return false;
 }
 
-static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required) {
+// returns whether the write (out) nodes overwrite the read nodes in operation
+static bool ggml_cuda_check_fusion_memory_ranges(ggml_cgraph * cgraph,
+                                                 int           node_idx,
+                                                 int           node_count,
+                                                 int *         out_nodes,
+                                                 int           out_count) {
+    auto nodes_overlap = [&](const ggml_tensor * a, const ggml_tensor * b) {
+        const int64_t a_start = (int64_t) a->data;
+        const int64_t a_end   = a_start + ggml_nbytes(a);
+
+        const int64_t b_start = (int64_t) b->data;
+        const int64_t b_end   = b_start + ggml_nbytes(b);
+
+        if ((b_start <= a_start && a_start < b_end) || (a_start <= b_start && b_start < a_end)) {
+            return true;
+        }
+
+        return false;
+    };
+
+    bool is_ok = true;
+    // for nrows=1, all fusion operations correctly read the src before writing dst or do it elementwise, so we should be ok
+    if (ggml_nrows(cgraph->nodes[node_idx]) == 1) {
+        return true;
+    }
+
+    for (int i = 0; i < out_count; ++i) {
+        const ggml_tensor * dst = cgraph->nodes[out_nodes[i]];
+
+        for (int j = node_idx; j < node_idx + node_count; ++j) {
+            // Loop over all srcs of all nodes in the fusion. If the src overlaps
+            // the destination and the src is not an intermediate node that's being
+            // elided, then disable fusion.
+
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+                const ggml_tensor * src = cgraph->nodes[j]->src[src_idx];
+
+                if (!src || src->op == GGML_OP_NONE) {
+                    continue;
+                }
+
+                if (nodes_overlap(dst, src)) {
+                    bool found = false;
+
+                    for (int k = node_idx; k < j; ++k) {
+                        if (cgraph->nodes[k] == src) {
+                            found = true;
+                            break;
+                        }
+                    }
+
+                    if (!found) {
+                        is_ok = false;
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    return is_ok;
+}
+
+static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required, const void * graph_key) {
     bool graph_evaluated_or_captured = false;
 
     // flag used to determine whether it is an integrated_gpu
@@ -3378,39 +3669,84 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                     continue;
                 }
 
+                if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                    continue;
+                }
 
                 // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
+                    ggml_cuda_topk_moe_args args;
+
+                    if (cgraph->nodes[i]->op == GGML_OP_UNARY || cgraph->nodes[i]->op == GGML_OP_SOFT_MAX ||
+                        cgraph->nodes[i]->op == GGML_OP_ARGSORT) {
+                        const bool can_fuse = ggml_cuda_topk_moe_fusion(cgraph, i, args);
+
+                        std::vector<ggml_op> ops;
+
+                        if (can_fuse) {
+                            const ggml_tensor * logits  = node->src[0];
+                            ggml_tensor *       weights = nullptr;
+                            ggml_tensor *       ids     = nullptr;
+                            const ggml_tensor * bias    = nullptr;
+                            const ggml_tensor * clamp   = nullptr;
+                            const ggml_tensor * scale   = nullptr;
+
+                            if (!args.delayed_softmax) {
+                                ggml_op gating_op = args.sigmoid ? GGML_OP_UNARY : GGML_OP_SOFT_MAX;
+                                int     out_nodes[2];  // nodes which can't be elided
+
+                                if (args.prob_bias) {
+                                    bias = cgraph->nodes[i + 2]->src[1];
+                                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT,
+                                                            GGML_OP_VIEW, GGML_OP_GET_ROWS });
+                                    out_nodes[0] = i + 4;
+                                    ids          = cgraph->nodes[i + 4];
+                                } else {
+                                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW,
+                                                            GGML_OP_GET_ROWS });
+                                    out_nodes[0] = i + 3;
+                                    ids          = cgraph->nodes[i + 3];
+                                }
 
-                    if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
-                        ggml_tensor * weights          = cgraph->nodes[i + 9];
-                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
-                        ggml_tensor * clamp            = cgraph->nodes[i + 7];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ true,
-                                              /*delayed softmax*/ false, clamp);
-                        i += 9;
-                        continue;
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) {
-                        ggml_tensor * weights          = cgraph->nodes[i + 4];
-                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ false,
-                                              /*delayed softmax*/ false);
-                        i += 4;
-                        continue;
-                    }
+                                if (args.norm) {
+                                    ops.insert(ops.end(), { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP,
+                                                            GGML_OP_DIV, GGML_OP_RESHAPE });
+                                    clamp = cgraph->nodes[i + ops.size() - 3];
+                                }
+                                if (args.scale) {
+                                    ops.insert(ops.end(), { GGML_OP_SCALE });
+                                    scale = cgraph->nodes[i + ops.size() - 1];
+                                }
 
-                    if (ggml_cuda_can_fuse(cgraph, i,
-                                           ggml_cuda_topk_moe_ops(/*with norm*/ false, /*delayed softmax*/ true), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i + 5];
-                        ggml_tensor * ids     = cgraph->nodes[i + 1];
+                                weights      = cgraph->nodes[i + ops.size() - 1];
+                                out_nodes[1] = i + ops.size() - 1;
 
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, ids, /*with norm*/ false,
-                                              /*delayed_softmax*/ true);
-                        i += 5;
-                        continue;
+                                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
+                                    ggml_cuda_should_use_topk_moe(node, logits, weights, ids) &&
+                                    ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2)) {
+                                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
+                                    i += ops.size() - 1;
+                                    continue;
+                                }
+                            } else if (!args.norm && !args.prob_bias) {
+                                //special case gpt-oss, no norm, no bias.
+                                ops.insert(ops.end(), { GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS,
+                                                        GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE });
+                                weights                     = cgraph->nodes[i + 5];
+                                ids                         = cgraph->nodes[i + 1];
+                                const ggml_tensor * softmax = cgraph->nodes[i + 4];
+
+                                int out_nodes[2] = { i + 1, i + 5 };
+                                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
+                                    ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids) &&
+                                    ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2)) {
+                                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
+                                    i += ops.size() - 1;
+                                    continue;
+                                }
+                            }
+                        }
                     }
 
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
@@ -3442,11 +3778,13 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         n_fuse++;
 
                         if (n_fuse > 1) {
+                            ggml_tensor fused_add_node;
+                            memcpy(&fused_add_node, node, sizeof(ggml_tensor));
                             for (int j = 0; j < n_fuse - 1; ++j) {
-                                node->src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
+                                fused_add_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
                             }
-                            cgraph->nodes[i + n_fuse - 1]->data = node->data;
-                            ggml_cuda_op_fused_add(*cuda_ctx, node, n_fuse);
+                            fused_add_node.data = cgraph->nodes[i + n_fuse - 1]->data;
+                            ggml_cuda_op_fused_add(*cuda_ctx, &fused_add_node, n_fuse);
                             i += n_fuse - 1;
 
                             continue;
@@ -3655,6 +3993,20 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         continue;
                     }
 
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
+                        ggml_cuda_op_ssm_conv(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SILU }) ||
+                        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SIGMOID }) ||
+                        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SOFTPLUS })) {
+                        ggml_cuda_op_unary_mul(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
                         i += 2;
                         ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i], node);
@@ -3687,13 +4039,14 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
         }
 
 #ifdef USE_CUDA_GRAPH
+        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
         if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
-            if (cuda_ctx->cuda_graph->graph != nullptr) {
-                CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
-                cuda_ctx->cuda_graph->graph = nullptr;
+            if (graph->graph != nullptr) {
+                CUDA_CHECK(cudaGraphDestroy(graph->graph));
+                graph->graph = nullptr;
             }
 
-            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
+            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &graph->graph));
             graph_evaluated_or_captured = true; // CUDA graph has been captured
 
             std::lock_guard<std::mutex> lock(ggml_cuda_lock);
@@ -3706,41 +4059,38 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
     }
 
     if (use_cuda_graph) {
-        if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
-            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+        if (graph->instance == nullptr) { // Create executable graph from captured graph.
+            CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
         }
         if (cuda_graph_update_required) { // Update graph executable
-            ggml_cuda_graph_update_executable(cuda_ctx);
+            ggml_cuda_graph_update_executable(cuda_ctx, graph_key);
         }
         // Launch graph
-        CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
+        CUDA_CHECK(cudaGraphLaunch(graph->instance, cuda_ctx->stream()));
 #else
+        GGML_UNUSED(graph_key);
         graph_evaluated_or_captured = true;
 #endif  // USE_CUDA_GRAPH
     }
 }
 
-static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx) {
-
 #ifdef USE_CUDA_GRAPH
+static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
-    if (cuda_ctx->cuda_graph == nullptr) {
-        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
-    }
-
-    if (cuda_ctx->cuda_graph->graph == nullptr) {
+    if (graph->graph == nullptr) {
         if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
-            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+            if (!graph->disable_due_to_gpu_arch) {
+                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+            }
+            graph->disable_due_to_gpu_arch = true;
         }
     }
 
-    return cuda_ctx->cuda_graph->is_enabled();
-#else
-    GGML_UNUSED(cuda_ctx);
-    return false;
-#endif // USE_CUDA_GRAPH
+    return graph->is_enabled();
 }
+#endif // USE_CUDA_GRAPH
 
 static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
@@ -3749,15 +4099,40 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
 
     bool use_cuda_graph             = false;
     bool cuda_graph_update_required = false;
+    const void * graph_key = nullptr;
 
 #ifdef USE_CUDA_GRAPH
-    use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx);
-
-    if (cuda_ctx->cuda_graph->is_enabled()) {
-        cuda_graph_update_required = ggml_cuda_graph_update_required(cuda_ctx, cgraph);
-        use_cuda_graph             = ggml_cuda_graph_check_compability(cgraph);
-
-        cuda_ctx->cuda_graph->record_update(use_cuda_graph, cuda_graph_update_required);
+    graph_key = ggml_cuda_graph_get_key(cgraph);
+
+    ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+    if (graph->is_enabled()) {
+        const bool graph_compatible = ggml_cuda_graph_check_compability(cgraph);
+        if (graph_compatible) {
+            const bool properties_changed = ggml_cuda_graph_update_required(cuda_ctx, cgraph);
+
+            if (!graph->warmup_complete) {
+                // Warmup: need at least 2 calls with no property change on the 2nd call
+                if (!properties_changed) {
+                    graph->warmup_complete = true;
+                    GGML_LOG_DEBUG("%s: CUDA graph warmup complete\n", __func__);
+                    use_cuda_graph = true;
+                    cuda_graph_update_required = true;
+                }
+                // else: properties changed or first call - execute directly (use_cuda_graph stays false)
+            } else {
+                // Post-warmup: normal CUDA graph operation
+                if (properties_changed) {
+                    // Properties changed - reset warmup, execute directly until stable again
+                    graph->warmup_complete = false;
+                    GGML_LOG_DEBUG("%s: CUDA graph warmup reset\n", __func__);
+                } else {
+                    use_cuda_graph = true;
+                    cuda_graph_update_required = graph->instance == nullptr;
+                }
+            }
+        }
     }
 #endif // USE_CUDA_GRAPH
 
@@ -3771,7 +4146,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required);
+    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required, graph_key);
 
     return GGML_STATUS_SUCCESS;
 }
@@ -3804,7 +4179,14 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
 static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
 
-    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx);
+#ifdef USE_CUDA_GRAPH
+    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
+    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+#else
+    const bool use_cuda_graph = false;
+    GGML_UNUSED(cuda_ctx);
+    GGML_UNUSED(cgraph);
+#endif
 
     static bool enable_graph_optimization = [] {
         const char * env     = getenv("GGML_CUDA_GRAPH_OPT");
@@ -4335,6 +4717,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
                 case GGML_UNARY_OP_TRUNC:
+                    // TODO: should become:
+                    //return ggml_is_contiguous_rows(op->src[0]);
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -4551,7 +4935,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_L2_NORM:
             return true;
         case GGML_OP_RMS_NORM_BACK:
-            return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
+            return ggml_is_contiguous(op->src[0]);
             break;
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
@@ -4613,8 +4997,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
-        case GGML_OP_ACC:
             return true;
+        case GGML_OP_ACC:
+            // TODO: extend support like so:
+            //return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]);
+            return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
         case GGML_OP_SUM:
             return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_TOP_K:
@@ -4627,8 +5014,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_PAD:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_PAD:
+            return true;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
@@ -4638,6 +5026,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_RWKV_WKV7:
             return true;
+        case GGML_OP_GATED_DELTA_NET:
+            //TODO: enable once MUSA compiler is solved https://github.com/ggml-org/llama.cpp/pull/19504#issuecomment-4018634327
+#ifdef GGML_USE_MUSA
+            return false;
+#else
+            return true;
+#endif // GGML_USE_MUSA
         case GGML_OP_FLASH_ATTN_EXT:
             return ggml_cuda_flash_attn_ext_supported(dev_ctx->device, op);
         case GGML_OP_CROSS_ENTROPY_LOSS: