wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-07-21 23:41:28 +08:00
Code Issues Actions 10 Packages Projects Releases Wiki Activity

feat: Gemma4 text generation support (CORE-30) (#13376)

Browse Source 
* initial gemma4 support

* parity with reference implementation

outputs can 100% match transformers with same sdpa flags, checkpoint this and then optimize

* Cleanup, video fixes

* cleanup, enable fused rms norm by default

* update comment

* Cleanup

* Update sd.py

* Various fixes

* Add fp8 scaled embedding support

* small fixes

* Translate think tokens

* Fix image encoder attention mask type

So it works with basic attention

* Handle thinking tokens different only for Gemma4

* Code cleanup

* Update nodes_textgen.py

* Use embed scale class instead of buffer

Slight difference to HF, but technically more accurate and simpler code

* Default to fused rms_norm

* Update gemma4.py
This commit is contained in:
Jukka Seppänen
2026-05-02 22:46:15 -04:00
committed by GitHub
parent f756d801a1
commit be95871adc
11 changed files with 1453 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files
comfy_extras/nodes_textgen.py
+8 -5
View File
@@ -32,6 +32,8 @@ class TextGenerate(io.ComfyNode):
io.Clip.Input("clip"),
io.String.Input("prompt", multiline=True, dynamic_prompts=True, default=""),
io.Image.Input("image", optional=True),
io.Image.Input("video", optional=True, tooltip="Video frames as image batch. Assumed to be 24 FPS; subsampled to 1 FPS internally."),
io.Audio.Input("audio", optional=True),
io.Int.Input("max_length", default=256, min=1, max=2048),
io.DynamicCombo.Input("sampling_mode", options=sampling_options, display_name="Sampling Mode"),
io.Boolean.Input("thinking", optional=True, default=False, tooltip="Operate in thinking mode if the model supports it."),
@@ -43,9 +45,9 @@ class TextGenerate(io.ComfyNode):
)
@classmethod
def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True, video=None, audio=None) -> io.NodeOutput:
tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking)
tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking, video=video, audio=audio)
# Get sampling parameters from dynamic combo
do_sample = sampling_mode.get("sampling_mode") == "on"
@@ -70,7 +72,8 @@ class TextGenerate(io.ComfyNode):
seed=seed
)
generated_text = clip.decode(generated_ids, skip_special_tokens=True)
generated_text = clip.decode(generated_ids)
return io.NodeOutput(generated_text)
@@ -161,12 +164,12 @@ class TextGenerateLTX2Prompt(TextGenerate):
)
@classmethod
def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True, video=None, audio=None) -> io.NodeOutput:
if image is None:
formatted_prompt = f"<start_of_turn>system\n{LTX2_T2V_SYSTEM_PROMPT.strip()}<end_of_turn>\n<start_of_turn>user\nUser Raw Input Prompt: {prompt}.<end_of_turn>\n<start_of_turn>model\n"
else:
formatted_prompt = f"<start_of_turn>system\n{LTX2_I2V_SYSTEM_PROMPT.strip()}<end_of_turn>\n<start_of_turn>user\n\n<image_soft_token>\n\nUser Raw Input Prompt: {prompt}.<end_of_turn>\n<start_of_turn>model\n"
return super().execute(clip, formatted_prompt, max_length, sampling_mode, image, thinking, use_default_template)
return super().execute(clip, formatted_prompt, max_length, sampling_mode, image=image, thinking=thinking, use_default_template=use_default_template, video=video, audio=audio)
class TextgenExtension(ComfyExtension):