increase graph_max_nodes for finetune

fix regression during finetune on Llama-3.2-1B-F32:
GGML_ASSERT(cgraph->n_nodes < cgraph->size) failed

git bisect applying the most recent finetune (SGD) change
showed that
d498af3 Georgi Gerganov 2025-07-18 14:31:15 +0300 graph : avoid huge
warm-up graphs for MoE models (ggml-org#14753)

which greatly decreased graph_max_nodes
has been responsible for finetune failing on reasonably sized models
for the past two months. partially reverting the decrease (maybe
larger models still fail)

note: env LLAMA_SET_ROWS=0 is needed also or else:
GML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE) failed
(the node->op in question is indeed a rows op)
unfortunately a git revert on:

8a4280c Georgi Gerganov 2025-08-28 12:27:02 +0300 kv-cache : remove
LLAMA_SET_ROWS checks (ggml-org#15505)

is not straightforward, so this branch is behind that.

Author: graehl

src/llama-context.cpp

@@ -1338,7 +1338,7 @@ void llama_context::output_reorder() {
 //
 
 uint32_t llama_context::graph_max_nodes() const {
-    return std::max<uint32_t>(1024u, 8u*model.n_tensors());
+    return std::max<uint32_t>(4096u, 8u*model.n_tensors());
 }
 
 llm_graph_result * llama_context::get_gf_res_reserve() const {