Skip to content

LLM inference in C/C++

C++ ggml

mtmd: fix gemma 4 audio rms norm eps (#23815)

Xuan-Son Nguyend6be3158e1fa
diff --git a/conversion/gemma.py b/conversion/gemma.py
index 1b427a30cd..76beedcf0d 100644
--- a/conversion/gemma.py
+++ b/conversion/gemma.py
@@ -786,14 +786,15 @@ class Gemma4VisionAudioModel(MmprojModel):
super().set_gguf_parameters()
# vision params
+ assert self.hparams_vision is not None
self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.GEMMA4V)
- self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+ self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
# audio params
- if self.hparams_audio:
- self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
- self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
- self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+ assert self.hparams_audio is not None
+ self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
+ self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
+ self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-6))
def is_audio_tensor(self, name: str) -> bool:
return "audio_tower" in name or "embed_audio" in name
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 5fd583d40b..bbcae7609b 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -1552,6 +1552,9 @@ struct clip_model_loader {
hparams.audio_n_fft = 512;
hparams.audio_window_len = 320; // 20ms frame (NOT 25ms/400)
hparams.audio_hop_len = 160;
+ // due to a mistake in the original conversion code, rms_norm_eps is set to a wrong value
+ // since all gemma4a models use 1e-6, we just hardcode it here to avoid re-conversion
+ hparams.eps = 1e-6f;
} break;
case PROJECTOR_TYPE_GRANITE_SPEECH:
{