Fix for LTX-2.3 in ComfyUI: "slice indices must be integers" in lt.py line 168

[caoddgao]

Documenting a local fix for an error I encountered while running the LTX-2.3 image-to-video workflow in ComfyUI.

Error
TypeError: slice indices must be integers or None or have an __index__ method

Trace location

File "C:\Software\ComfyUI\comfy\text_encoders\lt.py", line 168, in encode_token_weights
out = out[:, :, -torch.sum(extra["attention_mask"]).item():]

Relevant file location

1. file: ComfyUI/comfy/text_encoders/lt.py
2. function: encode_token_weights(self, token_weight_pairs)
3. line: around 168

Problematic line
out = out[:, :, -torch.sum(extra["attention_mask"]).item():]

Cause
The slice expression is using:
torch.sum(extra["attention_mask"]).item()
directly as a Python slice index. If this resolves to a non-integer numeric value, Python raises:
TypeError: slice indices must be integers or None or have an __index__ method

Local patch
Replace the single slice line with:

if "attention_mask" in extra and extra["attention_mask"] is not None:
    valid_tokens = int(torch.sum(extra["attention_mask"]).item())
    valid_tokens = max(0, valid_tokens)

    if valid_tokens > 0:
        out = out[:, :, -valid_tokens:]
    else:
        out = out[:, :, :0]

Placement
Inside:
def encode_token_weights(self, token_weight_pairs):
insert the patch immediately below:
out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
and above:
out_device = out.device

Minimal interpretation
The fix is simply to make the slice index integer-safe before applying it.
This may help anyone debugging this exact traceback in the LTX text encoder path.