fit : count nextn (MTP) blocks in n_gpu_layers so front layers stay on GPU (ggml-org#26177)

John-194 · web-flow · commit 0324696b8e5f · 2026-07-27T16:21:37.000+03:00
diff --git a/common/fit.cpp b/common/fit.cpp
@@ -136,7 +136,7 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
         devs.push_back(llama_model_get_device(model, i));
     }
 
-    hp_ngl         = llama_model_n_layer(model);
+    hp_ngl         = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);
     hp_n_ctx_train = llama_model_n_ctx_train(model);
     hp_n_expert    = llama_model_n_expert(model);
 

Original file line number	Diff line number	Diff line change
`@@ -136,7 +136,7 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(`
`136`	`136`	`devs.push_back(llama_model_get_device(model, i));`
`137`	`137`	`}`
`138`	`138`
`139`		`- hp_ngl = llama_model_n_layer(model);`
	`139`	`+ hp_ngl = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);`
`140`	`140`	`hp_n_ctx_train = llama_model_n_ctx_train(model);`
`141`	`141`	`hp_n_expert = llama_model_n_expert(model);`
`142`	`142`