wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-05-19 05:27:24 +08:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

Update gemma4.py

Browse Source
This commit is contained in:
kijai 2026-04-30 14:31:23 +03:00
parent ee728a795f
commit de5e490e42
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/text_encoders/gemma4.py
View File

@ -12,9 +12,9 @@ from comfy.text_encoders.llama import RMSNorm, MLP, BaseLlama, BaseGenerate, _ma
# Intentional minor divergences from transformers -reference implementation:
# Embedding sqrt(hidden_size) scale applied as a Python scalar (full precision) instead of dtype-matched buffer tensor.
# RMSNorm uses torch fused F.rms_norm
# Input image and audio resizing/resampling slightly different numerically
# - Embedding sqrt(hidden_size) scale applied as a Python scalar (full precision) instead of dtype-matched buffer tensor.
# - RMSNorm uses torch fused F.rms_norm, very slight numerical differences, but considerably faster
# - Input image and audio resizing/resampling slightly different numerically
GEMMA4_VISION_CONFIG = {"hidden_size": 768, "image_size": 896, "intermediate_size": 3072, "num_attention_heads": 12, "num_hidden_layers": 16, "patch_size": 16, "head_dim": 64, "rms_norm_eps": 1e-6, "position_embedding_size": 10240, "pooling_kernel_size": 3}