correct comment

Author: ngxson

convert_lora_to_gguf.py

@@ -433,8 +433,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
                     assert isinstance(dest_data, LoraTorchTensor)
                     lora_a, lora_b = dest_data.get_lora_A_B()
 
-                    # token_embd A and B are already transposed by mergekit-extract-lora
-                    # we transpose A back again because it is used by llm_build_inp_embd()
+                    # note: mergekit-extract-lora flip and transpose A and B
+                    # here we only need to transpose token_embd.lora_a, see llm_build_inp_embd()
                     if "token_embd.weight" in dest_name:
                         lora_a = lora_a.T
 

src/llama-adapter.cpp

@@ -270,7 +270,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
         struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
         // validate tensor shape
         if (is_token_embd) {
-            // expect B to be transposed, see llm_build_inp_embd()
+            // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
             if (model_tensor->ne[0] != w.b->ne[1] || model_tensor->ne[1] != w.a->ne[1]) {
                 throw std::runtime_error("tensor '" + name + "' has incorrect shape");
             }