diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 6cd39ac7d..6252147a7 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -3095,7 +3095,6 @@ llm_graph_params llama_context::graph_params( llm_graph_type gtype) const { return { /*.model =*/ &model, - /*.hparams =*/ model.hparams, /*.cparams =*/ cparams, /*.ubatch =*/ ubatch, /*.gtype =*/ gtype, @@ -3106,8 +3105,6 @@ llm_graph_params llama_context::graph_params( /*.mctx =*/ mctx, /*.cross =*/ &cross, /*.moe_cache =*/ moe_cache.get(), - /*.prec_policy =*/ &model.prec_policy, - /*.hdmd =*/ model.hdmd.rot.empty() ? nullptr : &model.hdmd, /*.layer_inp_dev =*/ layer_inp_dev, /*.layer_inp_src =*/ supports_embd_row() ? cparams.ctx_other->get_layer_inp_dev() : nullptr, /*.h_nextn_dev =*/ cparams.embeddings_nextn_dev ? h_nextn_dev : nullptr, diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 58c5b6c9c..6706e238e 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -1594,9 +1594,10 @@ void llm_graph_result::set_params(const llm_graph_params & params) { llm_graph_context::llm_graph_context(const llm_graph_params & params) : model (*params.model), - hparams (params.hparams), + hparams (model.hparams), cparams (params.cparams), ubatch (params.ubatch), + prec_policy (model.prec_policy), n_embd (hparams.n_embd), n_layer (hparams.n_layer()), n_layer_nextn (hparams.n_layer_nextn), @@ -1630,8 +1631,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) : mctx (params.mctx), cross (params.cross), moe_cache (params.moe_cache), - prec_policy (params.prec_policy), - hdmd (params.hdmd), + hdmd (model.hdmd.rot.empty() ? nullptr : &model.hdmd), layer_inp_dev (params.layer_inp_dev), layer_inp_src (params.layer_inp_src), h_nextn_dev (params.h_nextn_dev), @@ -1728,9 +1728,7 @@ ggml_tensor * llm_graph_context::build_lora_mm( ggml_tensor * cur_mm = build_hadamard_input(w, cur); ggml_tensor * res = ggml_mul_mat(ctx0, w, cur_mm); - if (prec_policy) { - prec_policy->apply(res); - } + prec_policy.apply(res); if (w->type == GGML_TYPE_NVFP4) { ggml_prec_set_acc(res, GGML_PREC_BF16); @@ -1812,9 +1810,7 @@ ggml_tensor * llm_graph_context::build_lora_mm_id( ggml_mul_mat_id(ctx0, w, cur_mm, ids) : ggml_mul_mat_id(ctx0, cached ? moe_cache->get_experts(w) : w, cur_mm, slots); - if (prec_policy) { - prec_policy->apply(res); - } + prec_policy.apply(res); if (w->type == GGML_TYPE_NVFP4) { ggml_prec_set_acc(res, GGML_PREC_BF16); diff --git a/src/llama-graph.h b/src/llama-graph.h index 862e64e05..06ccc4075 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -804,7 +804,6 @@ class llm_graph_result; struct llm_graph_params { const llama_model * model = nullptr; - llama_hparams hparams; llama_cparams cparams; llama_ubatch ubatch; // note: intentionally make a copy @@ -820,10 +819,6 @@ struct llm_graph_params { const llama_cross * cross; const llama_moe_cache * moe_cache; - const llama_prec_policy * prec_policy = nullptr; - - const llama_hadamard * hdmd = nullptr; - ggml_tensor * layer_inp_dev = nullptr; // own layer input device buffer (write) ggml_tensor * layer_inp_src = nullptr; // ctx_other's layer input device buffer (read) ggml_tensor * h_nextn_dev = nullptr; // own h_nextn device buffer (write) @@ -1056,6 +1051,8 @@ struct llm_graph_context { const llama_cparams & cparams; const llama_ubatch & ubatch; + const llama_prec_policy & prec_policy; + const int64_t n_embd; const int64_t n_layer; const int64_t n_layer_nextn; @@ -1096,8 +1093,6 @@ struct llm_graph_context { const llama_cross * cross; const llama_moe_cache * moe_cache; - const llama_prec_policy * prec_policy; - const llama_hadamard * hdmd; ggml_tensor * layer_inp_dev;