From 9e1591780d735133caba2edb52f436cb4e8bc1f9 Mon Sep 17 00:00:00 2001 From: Bias92 Date: Wed, 4 Mar 2026 00:42:58 +0900 Subject: [PATCH] fix: correct EXAONE3 FFN_DOWN tensor mapping prefix MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The EXAONE3 FFN_DOWN mapping used prefix "model.layers.h.{bid}.mlp.c_proj" which is incorrect — EXAONE uses "transformer.h.{bid}.mlp.c_proj" prefix (matching gpt2/refact/qwen/jais). The correct mapping already exists on a different line but without the "exaone" comment tag. This fix: - Removes the dead/unreachable mapping with wrong prefix "model.layers.h." - Adds "exaone" tag to the existing correct mapping for documentation The wrong mapping was never hit at runtime because EXAONE weights use "transformer.h.{bid}.mlp.c_proj" which was already mapped, but the dead entry is misleading and could cause confusion. Signed-off-by: User Signed-off-by: Bias92 --- gguf-py/gguf/tensor_mapping.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index e575610900..6554a30ced 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -574,7 +574,7 @@ class TensorNameMap: # Feed-forward down MODEL_TENSOR.FFN_DOWN: ( "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox - "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais + "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais exaone "transformer.blocks.{bid}.ffn.down_proj", # mpt "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon "h.{bid}.mlp.dense_4h_to_h", # bloom @@ -599,7 +599,6 @@ class TensorNameMap: "model.layers.{bid}.residual_mlp.w2", # arctic "encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2 "encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm - "model.layers.h.{bid}.mlp.c_proj", # exaone "model.layers.{bid}.feed_forward.down_proj", # llama4 jamba granite-hybrid "transformer_encoder.{bid}.ffn.w3", # neobert "model.layers.{bid}.block_sparse_moe.down", # smallthinker