Skip to content

Add MPS support to the gemma node #345

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Zac (zboyles) wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Add MPS support to the gemma node #345

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7728d42
fix(gemma): add MPS support — fork_rng guard, dtype handling (float16…
zboyles Jun 7, 2026
90828c4
fix(gemma): log traceback when device detection fails (Copilot review)
zboyles Jun 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 13 additions & 2 deletions gemma_encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,7 +171,8 @@ def __init__(self, embedding_directory=None, tokenizer_data={}):
def ltxv_gemma_clip(encoder_path, ltxv_path, processor=None, dtype=None):
class _LTXVGemmaTextEncoderModel(LTXVGemmaTextEncoderModel):
def __init__(self, device="cpu", dtype=dtype, model_options={}):
dtype = torch.bfloat16 # TODO: make this configurable
if dtype is None:
dtype = torch.bfloat16

gemma_model = Gemma3ForConditionalGeneration.from_pretrained(
encoder_path,
Expand Down Expand Up @@ -264,6 +265,15 @@ def load_model(self, gemma_path: str, ltxv_path: str, max_length: int):
logger.warning(f"Could not load processor from {model_root}: {e}")

clip_dtype = torch.bfloat16
try:
# MPS does not have full support bfloat16, use float16 instead
if comfy.model_management.get_torch_device().type == "mps":
clip_dtype = torch.float16
except Exception as e:
logger.debug(
f"Could not detect device type for dtype selection: {e}", exc_info=True
)

ltxv_full_path = folder_paths.get_full_path("checkpoints", ltxv_path)
clip_target = comfy.supported_models_base.ClipTarget(
tokenizer=tokenizer_class,
Expand Down Expand Up @@ -428,9 +438,10 @@ def _enhance(
)
model_inputs = _pad_inputs_for_attention_alignment(model_inputs, pad_token_id)

devices = [model.device] if model.device.type == "cuda" else []
with (
torch.inference_mode(),
torch.random.fork_rng(devices=[model.device]),
torch.random.fork_rng(devices=devices),
torch.autocast(device_type=model.device.type, dtype=model.dtype),
):
torch.manual_seed(seed)
Expand Down