Flux 2 (#10879)

comfy/text_encoders/flux.py

@@ -1,10 +1,13 @@
 from comfy import sd1_clip
 import comfy.text_encoders.t5
 import comfy.text_encoders.sd3_clip
+import comfy.text_encoders.llama
 import comfy.model_management
-from transformers import T5TokenizerFast
+from transformers import T5TokenizerFast, LlamaTokenizerFast
 import torch
 import os
+import json
+import base64
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -68,3 +71,105 @@ def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None):
                 model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
             super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_
+
+def load_mistral_tokenizer(data):
+    if torch.is_tensor(data):
+        data = data.numpy().tobytes()
+
+    try:
+        from transformers.integrations.mistral import MistralConverter
+    except ModuleNotFoundError:
+        from transformers.models.pixtral.convert_pixtral_weights_to_hf import MistralConverter
+
+    mistral_vocab = json.loads(data)
+
+    special_tokens = {}
+    vocab = {}
+
+    max_vocab = mistral_vocab["config"]["default_vocab_size"]
+
+    for w in mistral_vocab["vocab"]:
+        r = w["rank"]
+        if r >= max_vocab:
+            continue
+
+        vocab[base64.b64decode(w["token_bytes"])] = r
+
+    for w in mistral_vocab["special_tokens"]:
+        if "token_bytes" in w:
+            special_tokens[base64.b64decode(w["token_bytes"])] = w["rank"]
+        else:
+            special_tokens[w["token_str"]] = w["rank"]
+
+    all_special = []
+    for v in special_tokens:
+        all_special.append(v)
+
+    special_tokens.update(vocab)
+    vocab = special_tokens
+    return {"tokenizer_object": MistralConverter(vocab=vocab, additional_special_tokens=all_special).converted(), "legacy": False}
+
+class MistralTokenizerClass:
+    @staticmethod
+    def from_pretrained(path, **kwargs):
+        return LlamaTokenizerFast(**kwargs)
+
+class Mistral3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.tekken_data = tokenizer_data.get("tekken_model", None)
+        super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"tekken_model": self.tekken_data}
+
+class Flux2Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="mistral3_24b", tokenizer=Mistral3Tokenizer)
+        self.llama_template = '[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]{}[/INST]'
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        return tokens
+
+class Mistral3_24BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        textmodel_json_config = {}
+        num_layers = model_options.get("num_layers", None)
+        if num_layers is not None:
+            textmodel_json_config["num_hidden_layers"] = num_layers
+            if num_layers < 40:
+                textmodel_json_config["final_norm"] = False
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 1, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Mistral3Small24B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Flux2TEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, name="mistral3_24b", clip_model=Mistral3_24BModel):
+        super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        out, pooled, extra = super().encode_token_weights(token_weight_pairs)
+
+        out = torch.stack((out[:, 10], out[:, 20], out[:, 30]), dim=1)
+        out = out.movedim(1, 2)
+        out = out.reshape(out.shape[0], out.shape[1], -1)
+        return out, pooled, extra
+
+def flux2_te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None, pruned=False):
+    class Flux2TEModel_(Flux2TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            if llama_quantization_metadata is not None:
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            if pruned:
+                model_options = model_options.copy()
+                model_options["num_layers"] = 30
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return Flux2TEModel_

comfy/text_encoders/llama.py

@@ -34,6 +34,28 @@ class Llama2Config:
     rope_scale = None
     final_norm: bool = True
 
+@dataclass
+class Mistral3Small24BConfig:
+    vocab_size: int = 131072
+    hidden_size: int = 5120
+    intermediate_size: int = 32768
+    num_hidden_layers: int = 40
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 8192
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 1000000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = None
+    k_norm = None
+    rope_scale = None
+    final_norm: bool = True
+
 @dataclass
 class Qwen25_3BConfig:
     vocab_size: int = 151936
@@ -465,6 +487,15 @@ class Llama2(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Mistral3Small24B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Mistral3Small24BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen25_3B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()