Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
Increase performance for Gemma3n models on NVGPUs by enabling CUDA Graph execution (#11525)
Oliver Simonsea85e27bbd76
diff --git a/llama/patches/0019-metal-add-mean-kernel-14267.patch b/llama/patches/0019-metal-add-mean-kernel-14267.patchindex a52f0fdfe..e65aeb7b4 100644--- a/llama/patches/0019-metal-add-mean-kernel-14267.patch+++ b/llama/patches/0019-metal-add-mean-kernel-14267.patch@@ -16,7 +16,7 @@ ggml-ci2 files changed, 67 insertions(+), 14 deletions(-)diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m-index ee4f2dcb..f20f5615 100644+index a9eeebc6..110c9ece 100644--- a/ggml/src/ggml-metal/ggml-metal.m+++ b/ggml/src/ggml-metal/ggml-metal.m@@ -489,6 +489,7 @@ enum ggml_metal_kernel_type {diff --git a/llama/patches/0020-CUDA-add-mean-operation-14313.patch b/llama/patches/0020-CUDA-add-mean-operation-14313.patchindex efcb1e8bc..2f4e37949 100644--- a/llama/patches/0020-CUDA-add-mean-operation-14313.patch+++ b/llama/patches/0020-CUDA-add-mean-operation-14313.patch@@ -52,7 +52,7 @@ index 64fb4ff4..5b9a0fe3 100644static __device__ __forceinline__ float warp_reduce_max(float x) {#pragma unrolldiff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu-index 4c829153..9e64e5ae 100644+index d6960174..2b9fabf4 100644--- a/ggml/src/ggml-cuda/ggml-cuda.cu+++ b/ggml/src/ggml-cuda/ggml-cuda.cu@@ -35,6 +35,7 @@diff --git a/llama/patches/0021-Enable-CUDA-Graphs-for-gemma3n.patch b/llama/patches/0021-Enable-CUDA-Graphs-for-gemma3n.patchnew file mode 100644index 000000000..b9dd6cdc6--- /dev/null+++ b/llama/patches/0021-Enable-CUDA-Graphs-for-gemma3n.patch@@ -0,0 +1,50 @@+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001+From: Oliver Simons <[email protected]>+Date: Tue, 22 Jul 2025 11:02:28 +0200+Subject: [PATCH] Enable CUDA Graphs for gemma3n.++Similar to+https://github.com/ggml-org/llama.cpp/pull/14741,+though ollama has a slightly different model graph+than llama.cpp which requires different workaround+checks.+---+ ggml/src/ggml-cuda/ggml-cuda.cu | 16 ++++++++++++----+ 1 file changed, 12 insertions(+), 4 deletions(-)++diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu+index 2b9fabf4..28ccf4be 100644+--- a/ggml/src/ggml-cuda/ggml-cuda.cu++++ b/ggml/src/ggml-cuda/ggml-cuda.cu+@@ -2474,6 +2474,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud+ // 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_src1_name = " (reshaped)";++ const std::string gemma3n_node_name = "node_";+++ for (int i = 0; i < cgraph->n_nodes; i++) {+ ggml_tensor * node = cgraph->nodes[i];++@@ -2495,12 +2498,17 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud+ #endif+ }++- if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {+- // disable CUDA graphs for batch size > 1 for now.+- // Changes in batch size or context size can cause changes to the grid size of some kernels.++ // workarounds to exclude Gemma3n's `project_per_layer_input` operation from the batch-size heuristic, specific to ollama's implementation of gemma3n++ // number of layers is different for per_layer_proj between gemma3n:2b and gemma3n:4b, which is why we don't check that value here++ if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && !(node->ne[0] == 256++ && node->ne[2] == 1++ && node->ne[3] == 1++ && node->src[0] ? std::string(node->src[0]->name).find(gemma3n_node_name) != std::string::npos : false++ && node->src[1] ? node->src[1]->name == gemma3n_per_layer_proj_src1_name : false)) {++ // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.+ 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_INFO("%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]);+ #endif+ }+diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu b/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cuindex 2b9fabf4f..28ccf4bef 100644--- a/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu+++ b/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu@@ -2474,6 +2474,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud// Loop over nodes in GGML graph to obtain info needed for CUDA graphcuda_ctx->cuda_graph->cpy_dest_ptrs.clear();+ const std::string gemma3n_per_layer_proj_src1_name = " (reshaped)";+ const std::string gemma3n_node_name = "node_";+for (int i = 0; i < cgraph->n_nodes; i++) {ggml_tensor * node = cgraph->nodes[i];@@ -2495,12 +2498,17 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud#endif}- if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {- // disable CUDA graphs for batch size > 1 for now.- // Changes in batch size or context size can cause changes to the grid size of some kernels.+ // workarounds to exclude Gemma3n's `project_per_layer_input` operation from the batch-size heuristic, specific to ollama's implementation of gemma3n+ // number of layers is different for per_layer_proj between gemma3n:2b and gemma3n:4b, which is why we don't check that value here+ if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && !(node->ne[0] == 256+ && node->ne[2] == 1+ && node->ne[3] == 1+ && node->src[0] ? std::string(node->src[0]->name).find(gemma3n_node_name) != std::string::npos : false+ && node->src[1] ? node->src[1]->name == gemma3n_per_layer_proj_src1_name : false)) {+ // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.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_INFO("%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]);#endif}diff --git a/model/models/gemma3n/model_text.go b/model/models/gemma3n/model_text.goindex 715b8a0ea..b75a2abb3 100644--- a/model/models/gemma3n/model_text.go+++ b/model/models/gemma3n/model_text.go@@ -203,10 +203,9 @@ func (a AltUp) Predict(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptionscoefficients := a.PredictionCoefficient.Forward(ctx, modalities)coefficients = coefficients.Reshape(ctx, opts.altupInputs, opts.altupInputs, coefficients.Dim(1), coefficients.Dim(2))- hiddenStates = hiddenStates.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)- predictions := coefficients.Mulmat(ctx, hiddenStates)- predictions = predictions.Add(ctx, hiddenStates)- return predictions.Permute(ctx, 2, 0, 1, 3).Contiguous(ctx)+ predictions := coefficients.Mulmat(ctx, hiddenStates.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx))+ predictions = predictions.Permute(ctx, 2, 0, 1, 3).Contiguous(ctx)+ return predictions.Add(ctx, hiddenStates)}func (a AltUp) Correct(ctx ml.Context, predictions, activated, one ml.Tensor, opts *TextOptions) ml.Tensor {