comfyanonymous/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-26 19:19:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement gemma 3 as a text encoder. (#10241)

Browse Source 
Not useful yet.
This commit is contained in:
comfyanonymous
2025-10-06 19:08:08 -07:00
committed by GitHub
parent 8c19910427
commit 8aea746212
4 changed files with 142 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
comfy/sd.py
View File

@@ -890,6 +890,7 @@ class TEModel(Enum):
QWEN25_3B = 10
QWEN25_7B = 11
BYT5_SMALL_GLYPH = 12
GEMMA_3_4B = 13
def detect_te_model(sd):
if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -912,6 +913,8 @@ def detect_te_model(sd):
return TEModel.BYT5_SMALL_GLYPH
return TEModel.T5_BASE
if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
if 'model.layers.0.self_attn.q_norm.weight' in sd:
return TEModel.GEMMA_3_4B
return TEModel.GEMMA_2_2B
if 'model.layers.0.self_attn.k_proj.bias' in sd:
weight = sd['model.layers.0.self_attn.k_proj.bias']
@@ -1016,6 +1019,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
elif te_model == TEModel.GEMMA_3_4B:
clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data), model_type="gemma3_4b")
clip_target.tokenizer = comfy.text_encoders.lumina2.NTokenizer
tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
elif te_model == TEModel.LLAMA3_8:
clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)