Commit 652d559ed for llama.cpp
commit 652d559edf278347160baeb91ab654ac57ac429e
Author: Aman Gupta <amangupta052@gmail.com>
Date: Fri Nov 20 17:19:14 2026 +0800
llama : move the graph node budget into the model classes (#32648)
graph_max_nodes() picked the nodes-per-tensor factor from lists of
architectures. Each model class now gives it through a new virtual,
llama_model::graph_nodes_per_tensor(), which defaults to 8. The
architectures that need 32 or 64 override it in models.h, and DFlash
keeps its hparams-based rule in dflash.cpp.
LoRA adapter nodes are now added for every architecture, not only for
the ones on the default factor.
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 7f1cc4c44..ea41283f9 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2911,42 +2911,9 @@ void llama_context::output_reorder() {
//
uint32_t llama_context::graph_max_nodes() const {
- uint32_t res;
- if (model.arch == LLM_ARCH_KIMI_K3 || model.arch == LLM_ARCH_KIMI_K31 || model.arch == LLM_ARCH_GLM5_NEXT ||
- model.arch == LLM_ARCH_HRM_TEXT) {
- res = std::max<uint32_t>(1024u, 64u * model.n_tensors());
- } else if (model.arch == LLM_ARCH_QWEN3NEXT ||
- model.arch == LLM_ARCH_KIMI_LINEAR ||
- model.arch == LLM_ARCH_METIS16 ||
- model.arch == LLM_ARCH_BAILINGMOE3 ||
- model.arch == LLM_ARCH_QWEN35 ||
- model.arch == LLM_ARCH_QWEN35MOE ||
- model.arch == LLM_ARCH_ZONOS3 ||
- model.arch == LLM_ARCH_CLEF ||
- model.arch == LLM_ARCH_CLOVER_MOE ||
- model.arch == LLM_ARCH_QWEN4EXP ||
- model.arch == LLM_ARCH_DEEPSEEK4 ||
- model.arch == LLM_ARCH_DEEPSEEK41 ||
- model.arch == LLM_ARCH_XING4_0 ||
- (model.arch == LLM_ARCH_DFLASH && model.hparams.dsv4_hc_mult > 0) ||
- model.arch == LLM_ARCH_NANBEIGE ||
- model.arch == LLM_ARCH_MINIMAX_01 ||
- model.arch == LLM_ARCH_MINIMAX_M3 ||
- model.arch == LLM_ARCH_HY_V4 ||
- model.arch == LLM_ARCH_INKLING2) {
- res = std::max<uint32_t>(1024u, 32u * model.n_tensors());
- } else if (model.arch == LLM_ARCH_DFLASH && model.hparams.dflash_selector_rank > 0) {
- // DFlash2's convolutions and selector are shape work rather than matmuls,
- // so they cost ~8.6 nodes per tensor against ~5.9 for a plain DFlash draft.
- // The selector lattice scales with the draft block size; 12u was calibrated
- // on block-8 drafters, so scale the budget for larger blocks (e.g. Ripple's 16).
- const uint32_t block_scale = std::max<uint32_t>(1u, model.hparams.dflash_block_size/8u);
- res = std::max<uint32_t>(1024u, 12u*block_scale*model.n_tensors());
- } else {
- res = std::max<uint32_t>(1024u, 8u*model.n_tensors());
- for (const auto & lora : model.loras) {
- res += lora->get_n_nodes();
- }
+ uint32_t res = std::max<uint32_t>(1024u, model.graph_nodes_per_tensor()*model.n_tensors());
+ for (const auto & lora : model.loras) {
+ res += lora->get_n_nodes();
}
// the MoE cache adds a lookup to each MoE layer, and a second expert path with a CPU share (up to ~16 tensors per layer),
diff --git a/src/llama-model.h b/src/llama-model.h
index 8d2f5629e..1cc5d743a 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -1052,6 +1052,9 @@ struct llama_model {
llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
+ // graph size budget: max graph nodes per model tensor
+ virtual uint32_t graph_nodes_per_tensor() const { return 8; }
+
ggml_cgraph * build_graph(const llm_graph_params & params) const;
virtual void load_stats (llama_model_loader & ml) = 0;
diff --git a/src/models/dflash.cpp b/src/models/dflash.cpp
index 000222f4c..98a8d1d9a 100644
--- a/src/models/dflash.cpp
+++ b/src/models/dflash.cpp
@@ -125,6 +125,20 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
}
}
+uint32_t llama_model_dflash::graph_nodes_per_tensor() const {
+ if (hparams.dsv4_hc_mult > 0) {
+ return 32;
+ }
+ if (hparams.dflash_selector_rank > 0) {
+ // DFlash2's convolutions and selector are shape work rather than matmuls,
+ // so they cost ~8.6 nodes per tensor against ~5.9 for a plain DFlash draft.
+ // The selector lattice scales with the draft block size; 12u was calibrated
+ // on block-8 drafters, so scale the budget for larger blocks (e.g. Ripple's 16).
+ return 12u*std::max<uint32_t>(1u, hparams.dflash_block_size/8u);
+ }
+ return llama_model_base::graph_nodes_per_tensor();
+}
+
void llama_model_dflash::load_tensors_layers_base() {
LLAMA_LOAD_LOCALS;
diff --git a/src/models/models.h b/src/models/models.h
index e0a4db5f7..979ceffc2 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -442,6 +442,7 @@ struct llama_model_nanbeige : public llama_model_base {
llama_model_nanbeige(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
int n_loops = 1;
int n_layer_phys = 0;
@@ -801,6 +802,7 @@ struct llama_model_zonos3 : public llama_model_base {
llama_model_zonos3(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -1282,6 +1284,7 @@ struct llama_model_deepseek4 : public llama_model_base {
llama_model_deepseek4(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_graph_context {
struct no_build {};
@@ -1439,6 +1442,7 @@ struct llama_model_xing4_0 : public llama_model_base {
llama_model_xing4_0(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_graph_context {
graph(const llama_model & model, const llm_graph_params & params);
@@ -1571,6 +1575,7 @@ struct llama_model_dflash : public llama_model_base {
llama_model_dflash(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override;
enum variant_type {
VARIANT_DFLASH,
@@ -1833,6 +1838,7 @@ struct llama_model_metis16 : public llama_model_base {
llama_model_metis16(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
uint32_t n_pass = 1;
uint32_t n_stream = 1;
@@ -2134,6 +2140,7 @@ struct llama_model_hrm_text : public llama_model_base {
llama_model_hrm_text(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 64; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -2176,6 +2183,7 @@ struct llama_model_bailingmoe3 : public llama_model_base {
llama_model_bailingmoe3(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -2346,6 +2354,7 @@ struct llama_model_hy_v4 : public llama_model_base {
llama_model_hy_v4(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -2475,6 +2484,7 @@ struct llama_model_inkling2 : public llama_model_base {
llama_model_inkling2(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -2599,6 +2609,7 @@ struct llama_model_minimax_01 : public llama_model_base {
llama_model_minimax_01(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -2629,6 +2640,7 @@ struct llama_model_minimax_m3 : public llama_model_base {
llama_model_minimax_m3(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
msa_params msa_p;
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -2673,6 +2685,7 @@ struct llama_model_qwen3next : public llama_model_base {
llama_model_qwen3next(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -2715,6 +2728,7 @@ struct llama_model_qwen35 : public llama_model_base {
llama_model_qwen35(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -2851,6 +2865,7 @@ struct llama_model_qwen4exp : public llama_model_base {
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -2965,6 +2980,7 @@ struct llama_model_qwen35moe : public llama_model_base {
llama_model_qwen35moe(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -3007,6 +3023,7 @@ struct llama_model_clover_moe : public llama_model_base {
llama_model_clover_moe(const llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -3055,6 +3072,7 @@ struct llama_model_kimi_k3 : public llama_model_base {
llama_model_kimi_k3(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 64; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -3095,6 +3113,7 @@ struct llama_model_kimi_k31 : public llama_model_base {
llama_model_kimi_k31(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 64; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);
@@ -3141,6 +3160,7 @@ struct llama_model_glm5_next : public llama_model_base {
llama_model_glm5_next(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 64; }
// k-pool indexer inputs on top of the generic hybrid input
class llm_graph_input_kpool;
@@ -3225,6 +3245,7 @@ struct llama_model_kimi_linear : public llama_model_base {
llama_model_kimi_linear(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ uint32_t graph_nodes_per_tensor() const override { return 32; }
struct graph : public llm_build_delta_net_base {
graph(const llm_graph_params & params);