model: fix step3.5 n_rot (ggml-org#20318)

ngxson · web-flow · commit 0842b9b465f7 · 2026-03-09T23:42:24.000+01:00
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
@@ -7348,7 +7348,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     // ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer.
                     uint32_t n_rot_max = 0;
                     for (int i = 0; i < n_layer; ++i) {
-                        n_rot_max = std::max(n_rot_max, hparams.n_rot());
+                        n_rot_max = std::max(n_rot_max, hparams.n_rot(i));
                     }
                     if (n_rot_max == 0) {
                         n_rot_max = n_rot;

Original file line number	Diff line number	Diff line change
`@@ -7348,7 +7348,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {`
`7348`	`7348`	`// ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer.`
`7349`	`7349`	`uint32_t n_rot_max = 0;`
`7350`	`7350`	`for (int i = 0; i < n_layer; ++i) {`
`7351`		`- n_rot_max = std::max(n_rot_max, hparams.n_rot());`
	`7351`	`+ n_rot_max = std::max(n_rot_max, hparams.n_rot(i));`
`7352`	`7352`	`}`
`7353`	`7353`	`if (n_rot_max == 0) {`
`7354`	`7354`	`n_rot_max = n_rot;`