ComfyUI/comfy/text_encoders/flux.py

from comfy import sd1_clip
import comfy.text_encoders.t5
import comfy.model_management
from transformers import T5TokenizerFast
import torch
import os

class T5XXLModel(sd1_clip.SDClipModel):
    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)

class T5XXLTokenizer(sd1_clip.SDTokenizer):
    def __init__(self, embedding_directory=None, tokenizer_data={}):
        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)


class FluxTokenizer:
    def __init__(self, embedding_directory=None, tokenizer_data={}):
        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory)
        self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory)

    def tokenize_with_weights(self, text:str, return_word_ids=False):
        out = {}
        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids)
        return out

    def untokenize(self, token_weight_pair):
        return self.clip_l.untokenize(token_weight_pair)

    def state_dict(self):
        return {}


class FluxClipModel(torch.nn.Module):
    def __init__(self, dtype_t5=None, device="cpu", dtype=None):
        super().__init__()
        dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False)
        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
        self.dtypes = set([dtype, dtype_t5])

    def set_clip_options(self, options):
        self.clip_l.set_clip_options(options)
        self.t5xxl.set_clip_options(options)

    def reset_clip_options(self):
        self.clip_l.reset_clip_options()
        self.t5xxl.reset_clip_options()

    def encode_token_weights(self, token_weight_pairs):
        token_weight_pairs_l = token_weight_pairs["l"]
        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]

        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
        l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
        return t5_out, l_pooled

    def load_sd(self, sd):
        if "text_model.encoder.layers.1.mlp.fc1.weight" in sd:
            return self.clip_l.load_sd(sd)
        else:
            return self.t5xxl.load_sd(sd)

def flux_clip(dtype_t5=None):
    class FluxClipModel_(FluxClipModel):
        def __init__(self, device="cpu", dtype=None):
            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype)
    return FluxClipModel_