LLM inference in C/C++
C++ ggml
mtmd: fix gemma 4 ffn act (#30882)
Xuan-Son Nguyen5ba3b9fd3da3
diff --git a/conversion/gemma.py b/conversion/gemma.pyindex 924d8e784a..ab2d90610d 100644--- a/conversion/gemma.py+++ b/conversion/gemma.py@@ -999,6 +999,8 @@ class Gemma4VisionAudioModel(MmprojModel):assert self.hparams_vision is not Noneself.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.GEMMA4V)self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))+ assert self.hparams_vision.get("hidden_activation", "gelu_pytorch_tanh") == "gelu_pytorch_tanh"+ self.gguf_writer.add_vision_use_gelu(True)# audio paramsif self.has_audio_encoder:diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cppindex 3a4dd682ab..41aba280ba 100644--- a/tools/mtmd/clip.cpp+++ b/tools/mtmd/clip.cpp@@ -1735,6 +1735,8 @@ struct clip_model_loader {case PROJECTOR_TYPE_GEMMA4UV:{hparams.rope_theta = 100.0f;+ hparams.ffn_op = FFN_GELU;+ log_ffn_op = "gelu";hparams.n_merge = 3; // pooling_kernel_sizehparams.image_resize_algo = RESIZE_ALGO_BICUBIC;get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);