whispercpp 1.3.4 → 1.3.5

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

@@ -19,7 +19,9 @@
 #include "ggml-cuda/count-equal.cuh"
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
+#include "ggml-cuda/cumsum.cuh"
 #include "ggml-cuda/diagmask.cuh"
+#include "ggml-cuda/diag.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
@@ -43,6 +45,7 @@
 #include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/top-k.cuh"
 #include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/topk-moe.cuh"
@@ -50,8 +53,13 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
+#include "ggml-cuda/solve_tri.cuh"
+#include "ggml-cuda/tri.cuh"
+#include "ggml-cuda/cumsum.cuh"
+#include "ggml-cuda/fill.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -195,16 +203,6 @@ static ggml_cuda_device_info ggml_cuda_init() {
     GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);
 
     int64_t total_vram = 0;
-#ifdef GGML_CUDA_FORCE_MMQ
-    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    yes\n", __func__);
-#else
-    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    no\n", __func__);
-#endif // GGML_CUDA_FORCE_MMQ
-#ifdef GGML_CUDA_FORCE_CUBLAS
-    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: yes\n", __func__);
-#else
-    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__);
-#endif // GGML_CUDA_FORCE_CUBLAS
     GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
 
     std::vector<std::pair<int, std::string>> turing_devices_without_mma;
@@ -231,10 +229,18 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
-        info.devices[id].integrated = prop.integrated;
+        info.devices[id].integrated = false; // Temporarily disabled due to issues with corrupted output (e.g. #15034)
         info.devices[id].nsm        = prop.multiProcessorCount;
         info.devices[id].smpb       = prop.sharedMemPerBlock;
         info.devices[id].warp_size  = prop.warpSize;
+
+#ifndef GGML_USE_MUSA
+        int supports_coop_launch = 0;
+        CUDA_CHECK(cudaDeviceGetAttribute(&supports_coop_launch, cudaDevAttrCooperativeLaunch, id));
+        info.devices[id].supports_cooperative_launch = !!supports_coop_launch;
+#else
+        info.devices[id].supports_cooperative_launch = false;
+#endif // !(GGML_USE_MUSA)
 #if defined(GGML_USE_HIP)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
@@ -273,6 +279,15 @@ static ggml_cuda_device_info ggml_cuda_init() {
         } else if (device_name.substr(0, 21) == "NVIDIA GeForce GTX 16") {
             turing_devices_without_mma.push_back({ id, device_name });
         }
+
+        // Temporary performance fix:
+        // Setting device scheduling strategy for iGPUs with cc121 to "spinning" to avoid delays in cuda synchronize calls.
+        // 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(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
+        }
+
 #endif  // defined(GGML_USE_HIP)
     }
 
@@ -511,7 +526,8 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
 };
 #endif // defined(GGML_USE_VMM)
 
-std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int                  device,
+                                                                               [[maybe_unused]] int stream_no) {
 #if defined(GGML_USE_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
@@ -1948,8 +1964,15 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
 
         size_t src1_stride_size = sizeof(cuda_t);
 
-        dim3 block_dims(ne13, ne12);
-        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
+        const int threads_x = 16;
+        const int threads_y = 16;
+        dim3 block_dims(threads_x, threads_y);
+
+        dim3 grid_dims(
+            (ne13 + threads_x - 1) / threads_x,
+            (ne12 + threads_y - 1) / threads_y
+        );
+        k_compute_batched_ptrs<<<grid_dims, block_dims, 0, main_stream>>>(
                 src0_ptr, src1_ptr, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
@@ -1998,6 +2021,164 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     }
 }
 
+static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
+                                          const ggml_tensor * ffn_gate,
+                                          const ggml_tensor * glu,
+                                          const ggml_tensor * ffn_up_bias = nullptr,
+                                          const ggml_tensor * ffn_gate_bias = nullptr) {
+    const bool has_bias = ffn_up_bias != nullptr || ffn_gate_bias != nullptr;
+
+    if (has_bias && (!ffn_up_bias || !ffn_gate_bias)) {
+        return false;
+    }
+
+    const bool is_mul_mat     = ffn_up->op == GGML_OP_MUL_MAT     && ffn_gate->op == GGML_OP_MUL_MAT     && glu->op == GGML_OP_GLU;
+    const bool is_mul_mat_id  = ffn_up->op == GGML_OP_MUL_MAT_ID  && ffn_gate->op == GGML_OP_MUL_MAT_ID  && glu->op == GGML_OP_GLU;
+
+    GGML_ASSERT(ffn_up && ffn_gate && glu);
+
+    if (!is_mul_mat && !is_mul_mat_id) {
+        return false;
+    }
+
+    const ggml_op expected_bias_op = is_mul_mat ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+    if (has_bias) {
+        if (ffn_up_bias->op != expected_bias_op || ffn_gate_bias->op != expected_bias_op) {
+            return false;
+        }
+
+        if (glu->src[0] != ffn_gate_bias || glu->src[1] != ffn_up_bias) {
+            return false;
+        }
+
+        if (expected_bias_op == GGML_OP_ADD) {
+            const bool up_has_mul   = ffn_up_bias->src[0] == ffn_up || ffn_up_bias->src[1] == ffn_up;
+            const bool gate_has_mul = ffn_gate_bias->src[0] == ffn_gate || ffn_gate_bias->src[1] == ffn_gate;
+            if (!up_has_mul || !gate_has_mul) {
+                return false;
+            }
+        } else { // GGML_OP_ADD_ID
+            if (ffn_up_bias->src[0] != ffn_up || ffn_gate_bias->src[0] != ffn_gate) {
+                return false;
+            }
+            if (ffn_up_bias->src[2] != ffn_up->src[2] || ffn_gate_bias->src[2] != ffn_gate->src[2]) {
+                return false;
+            }
+        }
+    } else {
+        if (glu->src[0] != ffn_gate && glu->src[1] != ffn_up) {
+            return false;
+        }
+    }
+
+    if (ffn_up->src[0]->type != ffn_gate->src[0]->type || !ggml_are_same_shape(ffn_up->src[0], ffn_gate->src[0]) ||
+        !ggml_are_same_stride(ffn_up->src[0], ffn_gate->src[0])) {
+        return false;
+    }
+
+    if (ffn_up->src[1] != ffn_gate->src[1]) {
+        return false;
+    }
+
+    if (ffn_up->src[2] && (ffn_up->src[2] != ffn_gate->src[2])) {
+        return false;
+    }
+
+    static constexpr std::array<ggml_glu_op, 3> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI };
+
+    if (std::find(valid_glu_ops.begin(), valid_glu_ops.end(), ggml_get_glu_op(glu)) == valid_glu_ops.end()) {
+        return false;
+    }
+
+    if (const bool swapped = ggml_get_op_params_i32(glu, 1); swapped) {
+        return false;
+    }
+
+    const bool split = ggml_backend_buft_is_cuda_split(ffn_up->src[0]->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(ffn_gate->src[0]->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return true;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool is_mul_mat_id = tensor->op == GGML_OP_MUL_MAT_ID;
+
+    bool use_mul_mat_vec_f =
+        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) &&
+        src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    return use_mul_mat_vec_f;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE &&
+                                   ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) &&
+                                   src0->view_src;
+
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 &&
+                             dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    // fusion is not universally faster on Pascal
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    if (cc <= GGML_CUDA_CC_PASCAL) {
+        return false;
+    }
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return use_mul_mat_vec_q;
+}
+
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
@@ -2030,17 +2211,17 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
 
             const int cc            = ggml_cuda_info().devices[id].cc;
             const int warp_size     = ggml_cuda_info().devices[id].warp_size;
-            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1], /*n_experts=*/0);
+            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
         const int cc            = ggml_cuda_info().devices[ctx.device].cc;
         const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
-        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1], /*n_experts=*/0);
+        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
@@ -2106,12 +2287,12 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             return;
         }
 
-        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12, /*n_experts=*/ne02)) {
             ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
             return;
         }
 
-        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2], /*mul_mat_id=*/true)) {
+        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
             ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
             return;
         }
@@ -2259,6 +2440,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SET_ROWS:
             ggml_cuda_op_set_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cuda_op_set(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2334,6 +2518,27 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_ELU:
                     ggml_cuda_op_elu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_XIELU:
+                    ggml_cuda_op_xielu(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_FLOOR:
+                    ggml_cuda_op_floor(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_CEIL:
+                    ggml_cuda_op_ceil(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ROUND:
+                    ggml_cuda_op_round(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TRUNC:
+                    ggml_cuda_op_trunc(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_EXPM1:
+                    ggml_cuda_op_expm1(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_cuda_op_softplus(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2437,6 +2642,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
                 break;
+        case GGML_OP_DIAG:
+            ggml_cuda_op_diag(ctx, dst);
+            break;
         case GGML_OP_DIAG_MASK_INF:
             ggml_cuda_op_diag_mask_inf(ctx, dst);
             break;
@@ -2479,6 +2687,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SUM:
             ggml_cuda_op_sum(ctx, dst);
             break;
+        case GGML_OP_CUMSUM:
+            ggml_cuda_op_cumsum(ctx, dst);
+            break;
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
@@ -2491,6 +2702,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SSM_SCAN:
             ggml_cuda_op_ssm_scan(ctx, dst);
             break;
+        case GGML_OP_TOP_K:
+            ggml_cuda_op_top_k(ctx, dst);
+            break;
         case GGML_OP_ARGSORT:
             ggml_cuda_op_argsort(ctx, dst);
             break;
@@ -2500,6 +2714,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CROSS_ENTROPY_LOSS:
             ggml_cuda_cross_entropy_loss(ctx, dst);
             break;
+        case GGML_OP_TRI:
+            ggml_cuda_op_tri(ctx, dst);
+            break;
         case GGML_OP_RWKV_WKV6:
             ggml_cuda_op_rwkv_wkv6(ctx, dst);
             break;
@@ -2518,6 +2735,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_SGD:
             ggml_cuda_opt_step_sgd(ctx, dst);
             break;
+        case GGML_OP_SOLVE_TRI:
+            ggml_cuda_op_solve_tri(ctx, dst);
+            break;
+        case GGML_OP_FILL:
+            ggml_cuda_op_fill(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2630,11 +2853,10 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
 }
 
 #ifdef USE_CUDA_GRAPH
-static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
-    bool use_cuda_graph) {
+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
-    cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
     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";
@@ -2685,118 +2907,105 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_CPY) {
-
-            // Store the pointers which are updated for each token, such that these can be sent
-            // to the device and accessed using indirection from CUDA graph
-            cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data);
-
-            // store a pointer to each copy op CUDA kernel to identify it later
-            void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
-            if (!ptr) {
-                use_cuda_graph = false;
-#ifndef NDEBUG
-                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
-#endif
-            }
-        }
-
         if (!use_cuda_graph) {
             break;
         }
     }
 
-    if (use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = true;
-        // copy pointers to GPU so they can be accessed via indirection within CUDA graph
-        ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream());
-    }
-
     return use_cuda_graph;
 }
 
-static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
-    graph_node_properties->node_address = node->data;
-    graph_node_properties->node_op = node->op;
+static void ggml_cuda_graph_node_set_properties(ggml_cuda_graph_node_properties * props, ggml_tensor * node) {
+    props->node_address = node->data;
+    props->node_op = node->op;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        graph_node_properties->ne[i] = node->ne[i];
-        graph_node_properties->nb[i] = node->nb[i];
+        props->ne[i] = node->ne[i];
+        props->nb[i] = node->nb[i];
     }
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
+        props->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
     }
-    memcpy(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS);
+    memcpy(props->op_params, node->op_params, GGML_MAX_OP_PARAMS);
 }
 
-static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
-    if (node->data != graph_node_properties->node_address &&
-          node->op != GGML_OP_CPY &&
+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) {
         return false;
     }
 
-    if (node->op != graph_node_properties->node_op) {
+    if (node->op != props->node_op) {
         return false;
     }
 
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (node->ne[i] != graph_node_properties->ne[i]) {
+        if (node->ne[i] != props->ne[i]) {
             return false;
         }
-        if (node->nb[i] != graph_node_properties->nb[i]) {
+        if (node->nb[i] != props->nb[i]) {
             return false;
         }
     }
 
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (node->src[i] &&
-            node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_CPY &&
+            node->src[i]->data != props->src_address[i] &&
             node->op != GGML_OP_VIEW
         ) {
             return false;
         }
     }
 
-    if (node->op == GGML_OP_SCALE &&
-        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
+        memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
         return false;
     }
 
     return true;
 }
 
-static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
+static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
 
-    bool cuda_graph_update_required = false;
+    bool res = false;
 
     if (cuda_ctx->cuda_graph->instance == nullptr) {
-        cuda_graph_update_required = true;
+        res = true;
     }
 
     // Check if the graph size has changed
-    if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
-        cuda_graph_update_required = true;
-        cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+    if (cuda_ctx->cuda_graph->props.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
+        res = true;
+        cuda_ctx->cuda_graph->props.resize(cgraph->n_nodes + cgraph->n_leafs);
     }
 
     // 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
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        bool has_matching_properties = true;
-        if (!cuda_graph_update_required) {
-            has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
+        bool props_match = true;
+        if (!res) {
+            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &cuda_ctx->cuda_graph->props[i]);
         }
-        if (!has_matching_properties) {
-            cuda_graph_update_required = true;
+        if (!props_match) {
+            res = true;
         }
-        set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
+        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[i], cgraph->nodes[i]);
     }
 
-    return cuda_graph_update_required;
+    for (int i = 0; i < cgraph->n_leafs; 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]);
+        }
+        if (!props_match) {
+            res = true;
+        }
+        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[cgraph->n_nodes + i], cgraph->leafs[i]);
+    }
+
+    return res;
 }
 
-static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
+static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx) {
 
 #if CUDART_VERSION >= 12000
     cudaGraphExecUpdateResultInfo result_info;
@@ -2824,6 +3033,40 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif
 
+static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
+                                                const ggml_tensor * view,
+                                                const ggml_tensor * set_rows) {
+
+    if (rope->op != GGML_OP_ROPE || view->op != GGML_OP_VIEW || set_rows->op != GGML_OP_SET_ROWS) {
+        return false;
+    }
+    // ne3 not tested
+    if (rope->src[0]->ne[3] != 1) {
+        return false;
+    }
+
+    if (set_rows->type != GGML_TYPE_F32 && set_rows->type != GGML_TYPE_F16) {
+        return false;
+    }
+
+    if (set_rows->src[1]->type != GGML_TYPE_I64) {
+        return false;
+    }
+
+    // The view should flatten two dims of rope into one dim
+    if (!ggml_is_contiguous(view) || view->ne[0] != rope->ne[0] * rope->ne[1]) {
+        return false;
+    }
+
+    // Only norm/neox shaders have the fusion code
+    const int mode = ((const int32_t *) rope->op_params)[2];
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+        return false;
+    }
+
+    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);
@@ -2831,39 +3074,94 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
 #endif
 
     //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(false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm = ggml_cuda_topk_moe_ops(true);
+    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 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 (ops.size() == topk_moe_ops_with_norm.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops_with_norm.begin())) {
+    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 (node_idx + topk_moe_ops_with_norm.size() > (size_t)cgraph->n_nodes) {
-            return false;
+        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
+            return true;
         }
+    }
 
-        for (size_t i = 0; i < topk_moe_ops_with_norm.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops_with_norm.begin()[i]) return false;
-        }
+    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+8];
+        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 (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
             return true;
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops.begin())) {
+    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 (node_idx + topk_moe_ops.size() > (size_t)cgraph->n_nodes) {
-            return false;
+        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
+            return true;
         }
+    }
+
+    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 };
+
+    std::initializer_list<enum ggml_op> mul_mat_id_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_MUL_MAT_ID, GGML_OP_GLU };
+    std::initializer_list<enum ggml_op> mul_mat_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_MUL_MAT,    GGML_OP_GLU };
 
-        for (size_t i = 0; i < topk_moe_ops.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops.begin()[i]) return false;
+    if ((is_equal(mul_mat_bias_glu_ops, ops) || is_equal(mul_mat_id_bias_glu_ops, ops)) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) {
+        const ggml_tensor * ffn_gate      = cgraph->nodes[node_idx];
+        const ggml_tensor * ffn_gate_bias = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * ffn_up        = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * ffn_up_bias   = cgraph->nodes[node_idx + 3];
+        const ggml_tensor * glu           = cgraph->nodes[node_idx + 4];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu, ffn_up_bias, ffn_gate_bias)) {
+            return true;
         }
+    }
 
-        ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx+4];
-        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+    if ((is_equal(mul_mat_id_glu_ops, ops) || is_equal(mul_mat_glu_ops, ops)) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+        const ggml_tensor * ffn_gate = cgraph->nodes[node_idx];
+        const ggml_tensor * ffn_up   = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * glu      = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu)) {
+            return true;
+        }
+    }
+
+    std::initializer_list<enum ggml_op> rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };
+
+    if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+        const ggml_tensor * rope     = cgraph->nodes[node_idx];
+        const ggml_tensor * view     = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
             return true;
         }
     }
@@ -2898,7 +3196,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
 
         //if rms norm is the B operand, then we don't handle broadcast
-        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
 
@@ -2938,42 +3236,192 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     return false;
 }
 
-static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
-    bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
+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) {
+    bool graph_evaluated_or_captured = false;
+
     // flag used to determine whether it is an integrated_gpu
-    const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
+    const bool integrated            = ggml_cuda_info().devices[cuda_ctx->device].integrated;
+
+    ggml_cuda_stream_context & stream_ctx = cuda_ctx->stream_context();
+    bool                         is_concurrent_event_active = false;
+    ggml_cuda_concurrent_event * concurrent_event           = nullptr;
+    bool                         should_launch_concurrent_events = false;
+
+    const auto try_launch_concurrent_event = [&](const ggml_tensor * node) {
+        if (stream_ctx.concurrent_events.find(node) != stream_ctx.concurrent_events.end()) {
+            concurrent_event = &stream_ctx.concurrent_events[node];
+
+            is_concurrent_event_active = true;
+
+            GGML_LOG_DEBUG("Launching %d streams at %s\n", concurrent_event->n_streams, node->name);
+
+            cudaStream_t main_stream = cuda_ctx->stream();  // this should be stream 0
+            GGML_ASSERT(cuda_ctx->curr_stream_no == 0);
+            CUDA_CHECK(cudaEventRecord(concurrent_event->fork_event, main_stream));
+
+            for (int i = 1; i <= concurrent_event->n_streams; ++i) {
+                cudaStream_t stream = cuda_ctx->stream(cuda_ctx->device, i);
+                CUDA_CHECK(cudaStreamWaitEvent(stream, concurrent_event->fork_event));
+            }
+        }
+    };
 
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
+            [[maybe_unused]] int prev_i = 0;
+
+            if (stream_ctx.concurrent_events.size() > 0) {
+                should_launch_concurrent_events = true;
+                for (const auto & [tensor, event] : stream_ctx.concurrent_events) {
+                    should_launch_concurrent_events = should_launch_concurrent_events && event.is_valid();
+                }
+            }
+
+            if (should_launch_concurrent_events) {
+                // Restore original node order within each concurrent region to enable fusion within streams
+
+                std::unordered_map<const ggml_tensor *, int> node_to_idx;
+                node_to_idx.reserve(cgraph->n_nodes);
+                for (int i = 0; i < cgraph->n_nodes; ++i) {
+                    node_to_idx[cgraph->nodes[i]] = i;
+                }
+
+                for (auto & [fork_node, event] : stream_ctx.concurrent_events) {
+                    // Find positions of all nodes from this event in the current graph
+                    std::vector<int> positions;
+                    positions.reserve(event.original_order.size());
+
+                    bool all_found = true;
+                    for (const ggml_tensor * orig_node : event.original_order) {
+                        auto it = node_to_idx.find(orig_node);
+                        if (it != node_to_idx.end()) {
+                            positions.push_back(it->second);
+                        } else {
+                            all_found = false;
+                            break;
+                        }
+                    }
+
+                    if (!all_found || positions.size() != event.original_order.size()) {
+                        continue;
+                    }
+
+                    // Sort positions to get contiguous range
+                    std::vector<int> sorted_positions = positions;
+                    std::sort(sorted_positions.begin(), sorted_positions.end());
+
+                    bool is_contiguous = true;
+                    for (size_t i = 1; i < sorted_positions.size(); ++i) {
+                        if (sorted_positions[i] != sorted_positions[i-1] + 1) {
+                            is_contiguous = false;
+                            break;
+                        }
+                    }
+
+                    if (!is_contiguous) {
+                        continue;
+                    }
+
+                    // Restore original order at the sorted positions
+                    int start_pos = sorted_positions[0];
+                    for (size_t i = 0; i < event.original_order.size(); ++i) {
+                        cgraph->nodes[start_pos + i] = const_cast<ggml_tensor *>(event.original_order[i]);
+                    }
+                }
+            } else {
+                stream_ctx.concurrent_events.clear();
+            }
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
+                if (is_concurrent_event_active) {
+                    GGML_ASSERT(concurrent_event);
+
+                    if (node == concurrent_event->join_node) {
+                        cuda_ctx->curr_stream_no = 0;
+                        for (int i = 1; i <= concurrent_event->n_streams; ++i) {
+                            // Wait on join events of forked streams in the main stream
+                            CUDA_CHECK(cudaEventRecord(concurrent_event->join_events[i - 1],
+                                                       cuda_ctx->stream(cuda_ctx->device, i)));
+                            CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), concurrent_event->join_events[i - 1]));
+                        }
+
+                        is_concurrent_event_active = false;
+                        concurrent_event           = nullptr;
+                    } else {
+                        GGML_ASSERT (concurrent_event->stream_mapping.find(node) != concurrent_event->stream_mapping.end());
+                        cuda_ctx->curr_stream_no = concurrent_event->stream_mapping[node];
+                        GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
+                    }
+                } else if (i - prev_i > 1) {
+                    //the previous node was fused
+                    const ggml_tensor * prev_node = cgraph->nodes[i - 1];
+                    try_launch_concurrent_event(prev_node);
+
+                    if (is_concurrent_event_active) {
+                        cuda_ctx->curr_stream_no = concurrent_event->stream_mapping[node];
+                        GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
+                    }
+                }
+
+#ifdef GGML_CUDA_DEBUG
+                const int nodes_fused = i - prev_i - 1;
+                if (nodes_fused > 0) {
+                    GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
+                }
+#endif
+                prev_i = i;
 
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
                 }
 
+
+                // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+8];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ true);
-                        i += 8;
+                        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, weights, selected_experts, /*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 (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];
+
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, ids, /*with norm*/ false,
+                                              /*delayed_softmax*/ true);
+                        i += 5;
+                        continue;
+                    }
+
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
+                        ggml_tensor * rope = cgraph->nodes[i];
+                        ggml_tensor * set_rows = cgraph->nodes[i + 2];
+
+                        ggml_cuda_op_rope_fused(*cuda_ctx, rope, set_rows);
+                        i += 2;
+                        continue;
+                    }
+
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];
@@ -3005,6 +3453,195 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         }
                     }
 
+                    bool fused_mul_mat_vec = false;
+                    int fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu         = cgraph->nodes[i + 4];
+                            ggml_tensor * gate_bias_n = glu->src[0];
+                            ggml_tensor * up_bias_n   = glu->src[1];
+
+                            //we don't assume the order for {gate, up}. Instead infer it from the bias tensor
+                            ggml_tensor * gate_n      = nullptr;
+                            ggml_tensor * up_n        = nullptr;
+
+                            if (gate_bias_n->src[0] == cgraph->nodes[i] || gate_bias_n->src[1] == cgraph->nodes[i]) {
+                                gate_n = cgraph->nodes[i];
+                                up_n   = cgraph->nodes[i + 2];
+                            } else if (gate_bias_n->src[0] == cgraph->nodes[i + 2] || gate_bias_n->src[1] == cgraph->nodes[i + 2]) {
+                                gate_n = cgraph->nodes[i + 2];
+                                up_n   = cgraph->nodes[i];
+                            } else {
+                                continue;
+                            }
+
+                            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
+                                if (op_bias == GGML_OP_ADD) {
+                                    if (bias_node->src[0] == mul_node) {
+                                        return bias_node->src[1];
+                                    }
+                                    if (bias_node->src[1] == mul_node) {
+                                        return bias_node->src[0];
+                                    }
+                                    return (ggml_tensor *) nullptr;
+                                }
+                                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
+                                GGML_ASSERT(bias_node->src[0] == mul_node);
+                                return bias_node->src[1];
+                            };
+
+                            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
+                            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
+
+                            if (!up_bias_tensor || !gate_bias_tensor) {
+                                continue;
+                            }
+
+                            // we don't support repeating adds
+                            if (bias_op == GGML_OP_ADD &&
+                                (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
+                                 !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
+                                continue;
+                            }
+
+                            const ggml_tensor * src0 = up_n->src[0];
+                            const ggml_tensor * src1 = up_n->src[1];
+                            const ggml_tensor * ids  = up_n->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+                        } else if (ggml_cuda_can_fuse(cgraph, i, { op, op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu  = cgraph->nodes[i + 2];
+                            ggml_tensor * gate = glu->src[0];
+                            ggml_tensor * up   = glu->src[1];
+
+                            bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1])
+                                || (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
+
+                            if (!ok) continue;
+
+                            const ggml_tensor * src0 = up->src[0];
+                            const ggml_tensor * src1 = up->src[1];
+                            const ggml_tensor * ids  = up->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
+
+                    fused_mul_mat_vec = false;
+                    fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (!ggml_can_fuse(cgraph, i, { op, bias_op })) {
+                            continue;
+                        }
+
+                        ggml_tensor * mm_node   = cgraph->nodes[i];
+                        ggml_tensor * bias_node = cgraph->nodes[i + 1];
+
+                        ggml_tensor * bias_tensor = nullptr;
+                        if (bias_op == GGML_OP_ADD) {
+                            if (bias_node->src[0] == mm_node) {
+                                bias_tensor = bias_node->src[1];
+                            } else if (bias_node->src[1] == mm_node) {
+                                bias_tensor = bias_node->src[0];
+                            } else {
+                                continue;
+                            }
+                        } else {
+                            if (bias_node->src[0] != mm_node) {
+                                continue;
+                            }
+                            bias_tensor = bias_node->src[1];
+                        }
+
+                        const ggml_tensor * src0 = mm_node->src[0];
+                        const ggml_tensor * src1 = mm_node->src[1];
+                        const ggml_tensor * ids  = mm_node->src[2];
+
+                        if (bias_op == GGML_OP_ADD_ID && bias_node->src[2] != ids) {
+                            continue;
+                        }
+
+                        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
+                            continue;
+                        }
+
+                        ggml_cuda_mm_fusion_args_host fusion_data{};
+                        fusion_data.x_bias = bias_tensor;
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
+                            ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
+                            ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
 
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD}, {})) {
                         ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
@@ -3035,13 +3672,17 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 }
 #else
                 GGML_UNUSED(integrated);
-#endif // NDEBUG
+#endif  // NDEBUG
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
                     GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                 }
                 GGML_ASSERT(ok);
+
+                if (!is_concurrent_event_active) {
+                    try_launch_concurrent_event(node);
+               }
             }
         }
 
@@ -3069,7 +3710,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
             CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
         }
         if (cuda_graph_update_required) { // Update graph executable
-            update_cuda_graph_executable(cuda_ctx);
+            ggml_cuda_graph_update_executable(cuda_ctx);
         }
         // Launch graph
         CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
@@ -3079,60 +3720,46 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
     }
 }
 
-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;
-
-    ggml_cuda_set_device(cuda_ctx->device);
+static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx) {
 
 #ifdef USE_CUDA_GRAPH
-    static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
 
-    // Objects required for CUDA Graph
     if (cuda_ctx->cuda_graph == nullptr) {
         cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
     }
 
-    bool use_cuda_graph = true;
-    bool cuda_graph_update_required = false;
-
     if (cuda_ctx->cuda_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;
-#ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
-#endif
         }
     }
 
-    // Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
-    // or previous graph capture failure.
-    // Also disable for multi-gpu for now. TO DO investigate
-    if (disable_cuda_graphs_due_to_env
-        || cuda_ctx->cuda_graph->disable_due_to_gpu_arch
-        || cuda_ctx->cuda_graph->disable_due_to_too_many_updates
-        || cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
-        use_cuda_graph = false;
-    }
+    return cuda_ctx->cuda_graph->is_enabled();
+#else
+    GGML_UNUSED(cuda_ctx);
+    return false;
+#endif // USE_CUDA_GRAPH
+}
 
-    if (use_cuda_graph) {
-        cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
+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;
 
-        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph);
+    ggml_cuda_set_device(cuda_ctx->device);
 
-        // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
-        if (use_cuda_graph && cuda_graph_update_required) {
-            cuda_ctx->cuda_graph->number_consecutive_updates++;
-        } else {
-            cuda_ctx->cuda_graph->number_consecutive_updates = 0;
-        }
+    bool use_cuda_graph             = false;
+    bool cuda_graph_update_required = false;
 
-        if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
-            cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
-#ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
-#endif
-        }
+#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);
     }
+#endif // USE_CUDA_GRAPH
 
     if (use_cuda_graph && cuda_graph_update_required) {
         // Start CUDA graph capture
@@ -3144,18 +3771,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    if (!use_cuda_graph) {
-        cuda_ctx->cuda_graph->use_cpy_indirection = false;
-    }
-
-#else
-    bool use_cuda_graph = false;
-    bool cuda_graph_update_required = false;
-#endif // USE_CUDA_GRAPH
-
-    bool graph_evaluated_or_captured = false;
-
-    evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
+    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required);
 
     return GGML_STATUS_SUCCESS;
 }
@@ -3185,6 +3801,243 @@ 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);
+
+    static bool enable_graph_optimization = [] {
+        const char * env     = getenv("GGML_CUDA_GRAPH_OPT");
+        return env != nullptr && atoi(env) == 1;
+    }();
+
+    if (!enable_graph_optimization) {
+        return;
+    }
+
+    ggml_cuda_stream_context & stream_context = cuda_ctx->stream_context();
+    stream_context.reset();
+
+    if (!use_cuda_graph || ggml_backend_cuda_get_device_count() != 1) {
+        return;
+    }
+
+    // number of out-degrees for a particular node
+    std::unordered_map<const ggml_tensor *, int> fan_out;
+    // reverse mapping of node to index in the cgraph
+    std::unordered_map<const ggml_tensor *, int> node_indices;
+
+    const auto & is_noop = [](const ggml_tensor * node) -> bool {
+        return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE ||
+               node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
+    };
+
+    const auto & depends_on = [](const ggml_tensor * dst, const ggml_tensor * src) -> bool {
+        for (uint32_t s = 0; s < GGML_MAX_SRC; ++s) {
+            if (dst->src[s] == src) {
+                return true;
+            }
+        }
+        // implicit dependency if they view the same tensor
+        const ggml_tensor * dst2 = dst->view_src ? dst->view_src : dst;
+        const ggml_tensor * src2 = src->view_src ? src->view_src : src;
+        if (dst2 == src2) {
+            return true;
+        }
+        return false;
+    };
+
+    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
+        const ggml_tensor * node = cgraph->nodes[node_idx];
+        node_indices[node]       = node_idx;
+
+        if (is_noop(node)) {
+            continue;
+        }
+        for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+            const ggml_tensor * src = cgraph->nodes[node_idx]->src[src_idx];
+            //TODO: check why nrows > 1 fails
+            if (node && !is_noop(node) && ggml_nrows(node) <= 1) {
+                fan_out[src] += 1;
+            }
+        }
+    }
+
+    // Target Q, K, V for concurrency
+    // this is a more general way to find nodes which can be candidates for concurrency (although it has not been tested for anything else):
+    // 1. find fan-out (fork) nodes where the same input is used at least N times (in QKV, it would be "attn-norm")
+    // 2. find the join node, where 2 or more of the outputs are required (in QKV, this would "KQ" or "flash-attn")
+    // 3. account for all branches from the fork to the join
+    // 4. To extend lifetimes of the tensors, we interleave the branches (see below for more details)
+    // 5. save the original cgraph and restore it in graph_compute, to enable fusion within streams
+    // See discussion: https://github.com/ggml-org/llama.cpp/pull/16991#issuecomment-3522620030
+
+    const int min_fan_out = 3;
+    const int max_fan_out = 3;
+
+    // store {fork_idx, join_idx}
+    std::vector<std::pair<int, int>> concurrent_node_ranges;
+
+    for (const auto & [root_node, count] : fan_out) {
+        if (count >= min_fan_out && count <= max_fan_out) {
+            const int root_node_idx = node_indices[root_node];
+
+            // only optimize for attn_norm
+            // TODO: make this more generic
+            if (!strstr(root_node->name, "attn_norm")) {
+                continue;
+            }
+
+            bool is_part_of_event = false;
+            for (const auto & [start, end] : concurrent_node_ranges) {
+                if (root_node_idx >= start && root_node_idx <= end) {
+                    is_part_of_event = true;
+                }
+            }
+
+            if (is_part_of_event) {
+                continue;
+            }
+
+            std::vector<std::vector<const ggml_tensor *>> nodes_per_branch;
+            for (int i = root_node_idx + 1; i < cgraph->n_nodes; ++i) {
+                const ggml_tensor * node = cgraph->nodes[i];
+                if (!is_noop(node) && depends_on(node, root_node)) {
+                    nodes_per_branch.push_back({ node });
+                }
+            }
+
+            GGML_ASSERT(nodes_per_branch.size() == (size_t) count);
+
+            //find the join point
+            const ggml_tensor * join_node = nullptr;
+
+            const auto & belongs_to_branch = [&](const ggml_tensor *                      node,
+                                                 const std::vector<const ggml_tensor *> & branch) -> bool {
+                for (const ggml_tensor * n : branch) {
+                    if (depends_on(node, n)) {
+                        return true;
+                    }
+                }
+                return false;
+            };
+
+            for (int i = root_node_idx + 1; i < cgraph->n_nodes; ++i) {
+                const ggml_tensor * curr_node = cgraph->nodes[i];
+
+                int num_joins = 0;
+                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
+                    if (belongs_to_branch(curr_node, nodes_per_branch[branch_idx])) {
+                        num_joins++;
+                    }
+                }
+
+                if (num_joins >= 2) {
+                    join_node = curr_node;
+                    break;
+                }
+
+                bool found_branch = false;
+                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
+                    std::vector<const ggml_tensor *> & branch_vec = nodes_per_branch[branch_idx];
+                    if (belongs_to_branch(curr_node, branch_vec)) {
+                        //continue accumulating
+                        if (std::find(branch_vec.begin(), branch_vec.end(), curr_node) == branch_vec.end()) {
+                            branch_vec.push_back(curr_node);
+                        }
+                        found_branch = true;
+                    }
+                }
+
+                if (!found_branch && is_noop(curr_node)) {
+                    // we can put it in any branch because it will be ignored
+                    nodes_per_branch[0].push_back({ curr_node });
+                }
+            }
+
+            if (join_node) {
+                //Create ggml_cuda_concurrent_event
+                ggml_cuda_concurrent_event concurrent_event(nodes_per_branch.size());
+                concurrent_event.join_node = join_node;
+
+                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
+                    for (const ggml_tensor * n : nodes_per_branch[branch_idx]) {
+                        concurrent_event.stream_mapping[n] = branch_idx + 1;
+                    }
+                }
+
+                int fork_node_idx = node_indices[root_node];
+                int join_node_idx = node_indices[join_node];
+
+                int       current_branch_idx = 0;
+                int       current_node_idx   = fork_node_idx + 1;
+                const int n_branches         = nodes_per_branch.size();
+
+                int total_branch_nodes = 0;
+                for (std::vector<const ggml_tensor *> branch_nodes : nodes_per_branch) {
+                    total_branch_nodes += branch_nodes.size();
+                }
+
+                // there are other nodes in the middle which are unaccounted for
+                // usually (cpy) nodes, then ignore this fork
+                if (join_node_idx - fork_node_idx - 1 != total_branch_nodes) {
+                    GGML_LOG_DEBUG(
+                        "Skipping %s because the number of nodes in the middle is not equal to the total number of "
+                        "branch nodes %d != %d\n",
+                        root_node->name, join_node_idx - fork_node_idx - 1, total_branch_nodes);
+                    continue;
+                }
+
+                // Save the original order of nodes in this region before interleaving
+                // This is used later to restore grouping for fusion within streams
+                concurrent_event.original_order.reserve(total_branch_nodes);
+                for (int i = fork_node_idx + 1; i < join_node_idx; ++i) {
+                    concurrent_event.original_order.push_back(cgraph->nodes[i]);
+                }
+
+                std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> & concurrent_events = cuda_ctx->stream_context().concurrent_events;
+                GGML_ASSERT(concurrent_events.find(root_node) == concurrent_events.end());
+                concurrent_events.emplace(root_node, std::move(concurrent_event));
+                GGML_LOG_DEBUG("Adding stream at node %s %p\n", root_node->name, root_node);
+                concurrent_node_ranges.emplace_back(fork_node_idx, join_node_idx);
+
+                // interleave tensors to extend lifetimes so that ggml graph doesn't recycle them
+                // example transformation:
+                // [attn-norm, QMul, QNorm, QRope, KMul, KNorm, KRope, VMul, attn] ->
+                // [attn-norm, QMul, KMul, VMul, QNorm, VNorm, QRope, KRope, attn]
+                while (current_node_idx < join_node_idx) {
+                    std::vector<const ggml_tensor *> & branch_nodes = nodes_per_branch[current_branch_idx];
+
+                    bool has_node = false;
+                    for (std::vector<const ggml_tensor *> branch_node : nodes_per_branch) {
+                        has_node |= branch_node.size() > 0;
+                    }
+
+                    GGML_ASSERT(has_node);
+
+                    if (branch_nodes.empty()) {
+                        current_branch_idx = (current_branch_idx + 1) % n_branches;
+                        continue;
+                    }
+
+                    cgraph->nodes[current_node_idx] = const_cast<ggml_tensor *>(branch_nodes.front());
+                    current_node_idx++;
+                    branch_nodes.erase(branch_nodes.begin());
+
+                    // append all empty nodes
+                    while (!branch_nodes.empty() && is_noop(branch_nodes.front())) {
+                        cgraph->nodes[current_node_idx] = const_cast<ggml_tensor *>(branch_nodes.front());
+                        current_node_idx++;
+                        branch_nodes.erase(branch_nodes.begin());
+                    }
+
+                    current_branch_idx = (current_branch_idx + 1) % n_branches;
+                }
+            }
+        }
+    }
+}
+
 static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .get_name                = */ ggml_backend_cuda_get_name,
     /* .free                    = */ ggml_backend_cuda_free,
@@ -3199,7 +4052,7 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .event_record            = */ ggml_backend_cuda_event_record,
     /* .event_wait              = */ ggml_backend_cuda_event_wait,
-    /* .graph_optimize          = */ NULL,
+    /* .graph_optimize          = */ ggml_backend_cuda_graph_optimize,
 };
 
 static ggml_guid_t ggml_backend_cuda_guid() {
@@ -3270,6 +4123,7 @@ struct ggml_backend_cuda_device_context {
     std::string name;
     std::string description;
     std::string pci_bus_id;
+    int op_offload_min_batch_size;
 };
 
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -3282,10 +4136,110 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
+#if defined(__linux__)
+// Helper function to get available memory from /proc/meminfo for UMA systems
+static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_kb, long * free_swap_kb) {
+    FILE * meminfo_file = nullptr;
+    // 2KB buffer for reading /proc/meminfo since it does not report size info, should be enough
+    const size_t BUFFER_SIZE = 2048;
+    auto file_buffer = std::make_unique<char[]>(BUFFER_SIZE);
+    size_t bytes_read = 0;
+    long huge_tlb_total_pages = -1;
+    long huge_tlb_free_pages = -1;
+    long huge_tlb_page_size = -1;
+
+    if (available_memory_kb == nullptr || free_swap_kb == nullptr) {
+        return false;
+    }
+
+    meminfo_file = fopen("/proc/meminfo", "r");
+    if (meminfo_file == nullptr) {
+        GGML_LOG_ERROR("%s: failed to open /proc/meminfo\n", __func__);
+        return false;
+    }
+
+    // Read file into buffer
+    bytes_read = fread(file_buffer.get(), 1, BUFFER_SIZE - 1, meminfo_file);
+    fclose(meminfo_file);
+
+    if (bytes_read == 0) {
+        GGML_LOG_ERROR("%s: failed to read from /proc/meminfo\n", __func__);
+        return false;
+    }
+    file_buffer[bytes_read] = '\0';
+
+    *available_memory_kb = -1;
+    *free_swap_kb = -1;
+
+    // Parse the file buffer line by line
+    char * line = file_buffer.get();
+    char * line_next;
+    while (line < file_buffer.get() + bytes_read) {
+        // Find the end of the current line
+        line_next = strchr(line, '\n');
+        if (line_next != nullptr) {
+            *line_next = '\0';
+            line_next++;
+        } else {
+            line_next = file_buffer.get() + bytes_read;
+        }
+
+        long value;
+        if (sscanf(line, "MemAvailable: %ld kB", &value) == 1) {
+            *available_memory_kb = value;
+        } else if (sscanf(line, "SwapFree: %ld kB", &value) == 1) {
+            *free_swap_kb = value;
+        } else if (sscanf(line, "HugePages_Total: %ld", &value) == 1) {
+            huge_tlb_total_pages = value;
+        } else if (sscanf(line, "HugePages_Free: %ld", &value) == 1) {
+            huge_tlb_free_pages = value;
+        } else if (sscanf(line, "Hugepagesize: %ld kB", &value) == 1) {
+            huge_tlb_page_size = value;
+        }
+
+        line = line_next;
+    }
+
+    if (huge_tlb_total_pages != 0 && huge_tlb_total_pages != -1) {
+        *available_memory_kb = huge_tlb_free_pages * huge_tlb_page_size;
+
+        // Hugetlbfs pages are not swappable.
+        *free_swap_kb = 0;
+    }
+
+    GGML_LOG_DEBUG("%s: final available_memory_kb: %ld\n", __func__, *available_memory_kb);
+    return true;
+}
+#endif // defined(__linux__)
+
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     ggml_cuda_set_device(ctx->device);
     CUDA_CHECK(cudaMemGetInfo(free, total));
+
+// ref: https://github.com/ggml-org/llama.cpp/pull/17368
+#if defined(__linux__)
+    // Check if this is a UMA (Unified Memory Architecture) system
+    cudaDeviceProp prop;
+    CUDA_CHECK(cudaGetDeviceProperties(&prop, ctx->device));
+
+    // Check if UMA is explicitly enabled via environment variable
+    bool uma_env = getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr;
+    bool is_uma = prop.integrated > 0 || uma_env;
+
+    if (is_uma) {
+        // For UMA systems (like DGX Spark), use system memory info
+        long available_memory_kb = 0;
+        long free_swap_kb = 0;
+
+        if (ggml_backend_cuda_get_available_uma_memory(&available_memory_kb, &free_swap_kb) && available_memory_kb > 0) {
+            *free = (size_t)available_memory_kb * 1024;
+        } else {
+            GGML_LOG_ERROR("%s: /proc/meminfo reading failed, using cudaMemGetInfo\n", __func__);
+        }
+    }
+#endif // defined(__linux__)
+
 }
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
@@ -3373,7 +4327,14 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_EXPM1:
+                case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_XIELU:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3488,6 +4449,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                        op->src[0]->type == GGML_TYPE_F32 &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
+        case GGML_OP_SET:
+            {
+                const ggml_type t = op->type;
+                return (t == GGML_TYPE_F32 || t == GGML_TYPE_I32) &&
+                    t == op->src[0]->type &&
+                    t == op->src[1]->type;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3536,6 +4504,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_I32) {
+                    return true;
+                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }
@@ -3642,12 +4613,17 @@ 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_SUM:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM:
+            return ggml_is_contiguous_rows(op->src[0]);
+        case GGML_OP_TOP_K:
         case GGML_OP_ARGSORT:
-            // TODO: Support arbitrary column width
+#ifndef GGML_CUDA_USE_CUB
             return op->src[0]->ne[0] <= 1024;
+#else
+            return true;
+#endif
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
@@ -3668,7 +4644,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
+        case GGML_OP_FILL:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
+        case GGML_OP_DIAG:
+        case GGML_OP_SOLVE_TRI:
             return true;
+
         default:
             return false;
     }
@@ -3696,11 +4678,9 @@ static int64_t get_op_batch_size(const ggml_tensor * op) {
 }
 
 static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-
-    return get_op_batch_size(op) >= min_batch_size;
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
 
-    GGML_UNUSED(dev);
+    return get_op_batch_size(op) >= dev_ctx->op_offload_min_batch_size;
 }
 
 static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
@@ -3811,6 +4791,16 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
         features.push_back({ "FA_ALL_QUANTS", "1" });
     #endif
 
+    {
+        const auto & info = ggml_cuda_info();
+        for (int id = 0; id < info.device_count; ++id) {
+            if (blackwell_mma_available(info.devices[id].cc)) {
+                features.push_back({ "BLACKWELL_NATIVE_FP4", "1"});
+                break;
+            }
+        }
+    }
+
     #undef _STRINGIFY
     #undef STRINGIFY
 
@@ -3858,13 +4848,13 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
             ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
+            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             for (int i = 0; i < ggml_cuda_info().device_count; i++) {
                 ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
                 dev_ctx->device = i;
                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
 
-                ggml_cuda_set_device(i);
                 cudaDeviceProp prop;
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;
@@ -3872,6 +4862,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 char pci_bus_id[16] = {};
                 snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
                 dev_ctx->pci_bus_id = pci_bus_id;
+                dev_ctx->op_offload_min_batch_size = min_batch_size;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface   = */ ggml_backend_cuda_device_interface,