llama : increased max_nodes as large MoE models use massive amounts of nodes during warmup

sszymczy · sszymczy · commit c8bc6e4ff4b9 · 2025-02-01T12:43:14.000+01:00
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
@@ -3552,7 +3552,7 @@ size_t llama_model::size() const {
 }
 
 size_t llama_model::max_nodes() const {
-    return std::max<size_t>(8192, tensors_by_name.size()*5);
+    return std::max<size_t>(65536, tensors_by_name.size()*5);
 }
 
 size_t llama_model::n_devices() const {

Original file line number	Diff line number	Diff line change
`@@ -3552,7 +3552,7 @@ size_t llama_model::size() const {`
`3552`	`3552`	`}`
`3553`	`3553`
`3554`	`3554`	`size_t llama_model::max_nodes() const {`
`3555`		`- return std::max<size_t>(8192, tensors_by_name.size()*5);`
	`3555`	`+ return std::max<size_t>(65536, tensors_by_name.size()*5);`
`3556`	`3556`	`}`
`3557`	`3557`
`3558`	`3558`	`size_t llama_model::n_devices() const {`