feat: Add basic text generation support with native models, initially supporting Gemma3 (#12392)

comfy/text_encoders/llama.py

@@ -3,6 +3,8 @@ import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any, Tuple
 import math
+from tqdm import tqdm
+import comfy.utils
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
@@ -313,6 +315,13 @@ class Gemma3_4B_Config:
     final_norm: bool = True
     lm_head: bool = False
 
+GEMMA3_VISION_CONFIG = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
+
+@dataclass
+class Gemma3_4B_Vision_Config(Gemma3_4B_Config):
+    vision_config = GEMMA3_VISION_CONFIG
+    mm_tokens_per_image = 256
+
 @dataclass
 class Gemma3_12B_Config:
     vocab_size: int = 262208
@@ -336,7 +345,7 @@ class Gemma3_12B_Config:
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
     lm_head: bool = False
-    vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
+    vision_config = GEMMA3_VISION_CONFIG
     mm_tokens_per_image = 256
 
 class RMSNorm(nn.Module):
@@ -441,8 +450,10 @@ class Attention(nn.Module):
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
         past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        sliding_window: Optional[int] = None,
     ):
         batch_size, seq_length, _ = hidden_states.shape
+
         xq = self.q_proj(hidden_states)
         xk = self.k_proj(hidden_states)
         xv = self.v_proj(hidden_states)
@@ -477,6 +488,11 @@ class Attention(nn.Module):
             else:
                 present_key_value = (xk, xv, index + num_tokens)
 
+            if sliding_window is not None and xk.shape[2] > sliding_window:
+                xk = xk[:, :, -sliding_window:]
+                xv = xv[:, :, -sliding_window:]
+                attention_mask = attention_mask[..., -sliding_window:] if attention_mask is not None else None
+
         xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
         xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
 
@@ -559,10 +575,12 @@ class TransformerBlockGemma2(nn.Module):
         optimized_attention=None,
         past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
+        sliding_window = None
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
+                sliding_window = self.sliding_attention
                 if x.shape[1] > self.sliding_attention:
-                    sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype)
+                    sliding_mask = torch.full((x.shape[1], x.shape[1]), torch.finfo(x.dtype).min, device=x.device, dtype=x.dtype)
                     sliding_mask.tril_(diagonal=-self.sliding_attention)
                     if attention_mask is not None:
                         attention_mask = attention_mask + sliding_mask
@@ -581,6 +599,7 @@ class TransformerBlockGemma2(nn.Module):
             freqs_cis=freqs_cis,
             optimized_attention=optimized_attention,
             past_key_value=past_key_value,
+            sliding_window=sliding_window,
         )
 
         x = self.post_attention_layernorm(x)
@@ -765,6 +784,104 @@ class BaseLlama:
     def forward(self, input_ids, *args, **kwargs):
         return self.model(input_ids, *args, **kwargs)
 
+class BaseGenerate:
+    def logits(self, x):
+        input = x[:, -1:]
+        if hasattr(self.model, "lm_head"):
+            module = self.model.lm_head
+        else:
+            module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
+
+    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=[], initial_tokens=[], execution_dtype=None, min_tokens=0):
+        device = embeds.device
+        model_config = self.model.config
+
+        if execution_dtype is None:
+            if comfy.model_management.should_use_bf16(device):
+                execution_dtype = torch.bfloat16
+            else:
+                execution_dtype = torch.float32
+        embeds = embeds.to(execution_dtype)
+
+        if embeds.ndim == 2:
+            embeds = embeds.unsqueeze(0)
+
+        past_key_values = [] #kv_cache init
+        max_cache_len = embeds.shape[1] + max_length
+        for x in range(model_config.num_hidden_layers):
+            past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype),
+                                    torch.empty([embeds.shape[0], model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype), 0))
+
+        generator = torch.Generator(device=device).manual_seed(seed) if do_sample else None
+
+        generated_token_ids = []
+        pbar = comfy.utils.ProgressBar(max_length)
+
+        # Generation loop
+        for step in tqdm(range(max_length), desc="Generating tokens"):
+            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values)
+            logits = self.logits(x)[:, -1]
+            next_token = self.sample_token(logits, temperature, top_k, top_p, min_p, repetition_penalty, initial_tokens + generated_token_ids, generator, do_sample=do_sample)
+            token_id = next_token[0].item()
+            generated_token_ids.append(token_id)
+
+            embeds = self.model.embed_tokens(next_token).to(execution_dtype)
+            pbar.update(1)
+
+            if token_id in stop_tokens:
+                break
+
+        return generated_token_ids
+
+    def sample_token(self, logits, temperature, top_k, top_p, min_p, repetition_penalty, token_history, generator, do_sample=True):
+
+        if not do_sample or temperature == 0.0:
+            return torch.argmax(logits, dim=-1, keepdim=True)
+
+        # Sampling mode
+        if repetition_penalty != 1.0:
+            for i in range(logits.shape[0]):
+                for token_id in set(token_history):
+                    logits[i, token_id] *= repetition_penalty if logits[i, token_id] < 0 else 1/repetition_penalty
+
+        if temperature != 1.0:
+            logits = logits / temperature
+
+        if top_k > 0:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+        if min_p > 0.0:
+            probs_before_filter = torch.nn.functional.softmax(logits, dim=-1)
+            top_probs, _ = probs_before_filter.max(dim=-1, keepdim=True)
+            min_threshold = min_p * top_probs
+            indices_to_remove = probs_before_filter < min_threshold
+            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+        if top_p < 1.0:
+            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+            cumulative_probs = torch.cumsum(torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1)
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 0] = False
+            indices_to_remove = torch.zeros_like(logits, dtype=torch.bool)
+            indices_to_remove.scatter_(1, sorted_indices, sorted_indices_to_remove)
+            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+        probs = torch.nn.functional.softmax(logits, dim=-1)
+
+        return torch.multinomial(probs, num_samples=1, generator=generator)
+
 class BaseQwen3:
     def logits(self, x):
         input = x[:, -1:]
@@ -871,7 +988,7 @@ class Ovis25_2B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
+class Qwen25_7BVLI(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen25_7BVLI_Config(**config_dict)
@@ -881,6 +998,9 @@ class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
         self.visual = qwen_vl.Qwen2VLVisionTransformer(hidden_size=1280, output_hidden_size=config.hidden_size, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+        # todo: should this be tied or not?
+        #self.lm_head = operations.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
+
     def preprocess_embed(self, embed, device):
         if embed["type"] == "image":
             image, grid = qwen_vl.process_qwen2vl_images(embed["data"])
@@ -923,7 +1043,7 @@ class Gemma2_2B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Gemma3_4B(BaseLlama, torch.nn.Module):
+class Gemma3_4B(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Gemma3_4B_Config(**config_dict)
@@ -932,7 +1052,25 @@ class Gemma3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Gemma3_12B(BaseLlama, torch.nn.Module):
+class Gemma3_4B_Vision(BaseLlama, BaseGenerate, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Gemma3_4B_Vision_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+        self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations)
+        self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations)
+        self.image_size = config.vision_config["image_size"]
+
+    def preprocess_embed(self, embed, device):
+        if embed["type"] == "image":
+            image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True)
+            return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None
+        return None, None
+
+class Gemma3_12B(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Gemma3_12B_Config(**config_dict)

comfy/text_encoders/lt.py

@@ -6,6 +6,7 @@ import comfy.text_encoders.genmo
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import torch
 import comfy.utils
+import math
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -22,40 +23,79 @@ def ltxv_te(*args, **kwargs):
     return comfy.text_encoders.genmo.mochi_te(*args, **kwargs)
 
 
-class Gemma3_12BTokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_left=True, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
-
+class Gemma3_Tokenizer():
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
 
+    def tokenize_with_weights(self, text, return_word_ids=False, image=None, llama_template=None, skip_template=True, **kwargs):
+        self.llama_template = "<start_of_turn>system\nYou are a helpful assistant.<end_of_turn>\n<start_of_turn>user\n{}<end_of_turn>\n<start_of_turn>model\n"
+        self.llama_template_images = "<start_of_turn>system\nYou are a helpful assistant.<end_of_turn>\n<start_of_turn>user\n\n<image_soft_token>{}<end_of_turn>\n\n<start_of_turn>model\n"
+
+        if image is None:
+            images = []
+        else:
+            samples = image.movedim(-1, 1)
+            total = int(896 * 896)
+
+            scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+            width = round(samples.shape[3] * scale_by)
+            height = round(samples.shape[2] * scale_by)
+
+            s = comfy.utils.common_upscale(samples, width, height, "area", "disabled").movedim(1, -1)
+            images = [s[:, :, :, :3]]
+
+        if text.startswith('<start_of_turn>'):
+            skip_template = True
+
+        if skip_template:
+            llama_text = text
+        else:
+            if llama_template is None:
+                if len(images) > 0:
+                    llama_text = self.llama_template_images.format(text)
+                else:
+                    llama_text = self.llama_template.format(text)
+            else:
+                llama_text = llama_template.format(text)
+
+        text_tokens = super().tokenize_with_weights(llama_text, return_word_ids)
+
+        if len(images) > 0:
+            embed_count = 0
+            for r in text_tokens:
+                for i, token in enumerate(r):
+                    if token[0] == 262144 and embed_count < len(images):
+                        r[i] = ({"type": "image", "data": images[embed_count]},) + token[1:]
+                        embed_count += 1
+        return text_tokens
+
+class Gemma3_12BTokenizer(Gemma3_Tokenizer, sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        special_tokens = {"<image_soft_token>": 262144, "<end_of_turn>": 106}
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_left=True, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False, "special_tokens": special_tokens}, tokenizer_data=tokenizer_data)
+
+
 class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer)
 
+
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
         llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
         if llama_quantization_metadata is not None:
             model_options = model_options.copy()
             model_options["quantization_metadata"] = llama_quantization_metadata
-
+        self.dtypes = set()
+        self.dtypes.add(dtype)
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs):
-        text = llama_template.format(text)
-        text_tokens = super().tokenize_with_weights(text, return_word_ids)
-        embed_count = 0
-        for k in text_tokens:
-            tt = text_tokens[k]
-            for r in tt:
-                for i in range(len(r)):
-                    if r[i][0] == 262144:
-                        if image_embeds is not None and embed_count < image_embeds.shape[0]:
-                            r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:]
-                            embed_count += 1
-        return text_tokens
+    def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        tokens_only = [[t[0] for t in b] for b in tokens]
+        embeds, _, _, embeds_info = self.process_tokens(tokens_only, self.execution_device)
+        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106])  # 106 is <end_of_turn>
 
 class LTXAVTEModel(torch.nn.Module):
     def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
@@ -112,6 +152,9 @@ class LTXAVTEModel(torch.nn.Module):
 
         return out.to(out_device), pooled
 
+    def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        return self.gemma3_12b.generate(tokens["gemma3_12b"], do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed)
+
     def load_sd(self, sd):
         if "model.layers.47.self_attn.q_norm.weight" in sd:
             return self.gemma3_12b.load_sd(sd)
@@ -152,3 +195,14 @@ def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
     return LTXAVTEModel_
+
+def gemma3_te(dtype_llama=None, llama_quantization_metadata=None):
+    class Gemma3_12BModel_(Gemma3_12BModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return Gemma3_12BModel_

comfy/text_encoders/lumina2.py

@@ -1,23 +1,23 @@
 from comfy import sd1_clip
 from .spiece_tokenizer import SPieceTokenizer
 import comfy.text_encoders.llama
-
+from comfy.text_encoders.lt import Gemma3_Tokenizer
+import comfy.utils
 
 class Gemma2BTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2304, embedding_key='gemma2_2b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        special_tokens = {"<end_of_turn>": 107}
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2304, embedding_key='gemma2_2b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False, "special_tokens": special_tokens}, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
 
-class Gemma3_4BTokenizer(sd1_clip.SDTokenizer):
+class Gemma3_4BTokenizer(Gemma3_Tokenizer, sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, disable_weights=True, tokenizer_data=tokenizer_data)
-
-    def state_dict(self):
-        return {"spiece_model": self.tokenizer.serialize_model()}
+        special_tokens = {"<image_soft_token>": 262144, "<end_of_turn>": 106}
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False, "special_tokens": special_tokens}, disable_weights=True, tokenizer_data=tokenizer_data)
 
 class LuminaTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -31,6 +31,9 @@ class Gemma2_2BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma2_2B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
+    def generate(self, embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        return super().generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[107])
+
 class Gemma3_4BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
         llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
@@ -40,6 +43,23 @@ class Gemma3_4BModel(sd1_clip.SDClipModel):
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
+    def generate(self, embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        return super().generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106])
+
+class Gemma3_4B_Vision_Model(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B_Vision, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+    def process_tokens(self, tokens, device):
+        embeds, _, _, embeds_info = super().process_tokens(tokens, device)
+        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        return embeds
+
 class LuminaModel(sd1_clip.SD1ClipModel):
     def __init__(self, device="cpu", dtype=None, model_options={}, name="gemma2_2b", clip_model=Gemma2_2BModel):
         super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
@@ -50,6 +70,8 @@ def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b
         model = Gemma2_2BModel
     elif model_type == "gemma3_4b":
         model = Gemma3_4BModel
+    elif model_type == "gemma3_4b_vision":
+        model = Gemma3_4B_Vision_Model
 
     class LuminaTEModel_(LuminaModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):

comfy/text_encoders/spiece_tokenizer.py

@@ -6,9 +6,10 @@ class SPieceTokenizer:
     def from_pretrained(path, **kwargs):
         return SPieceTokenizer(path, **kwargs)
 
-    def __init__(self, tokenizer_path, add_bos=False, add_eos=True):
+    def __init__(self, tokenizer_path, add_bos=False, add_eos=True, special_tokens=None):
         self.add_bos = add_bos
         self.add_eos = add_eos
+        self.special_tokens = special_tokens
         import sentencepiece
         if torch.is_tensor(tokenizer_path):
             tokenizer_path = tokenizer_path.numpy().tobytes()
@@ -27,8 +28,32 @@ class SPieceTokenizer:
         return out
 
     def __call__(self, string):
+        if self.special_tokens is not None:
+            import re
+            special_tokens_pattern = '|'.join(re.escape(token) for token in self.special_tokens.keys())
+            if special_tokens_pattern and re.search(special_tokens_pattern, string):
+                parts = re.split(f'({special_tokens_pattern})', string)
+                result = []
+                for part in parts:
+                    if not part:
+                        continue
+                    if part in self.special_tokens:
+                        result.append(self.special_tokens[part])
+                    else:
+                        encoded = self.tokenizer.encode(part, add_bos=False, add_eos=False)
+                        result.extend(encoded)
+                return {"input_ids": result}
+
         out = self.tokenizer.encode(string)
         return {"input_ids": out}
 
+    def decode(self, token_ids, skip_special_tokens=False):
+
+        if skip_special_tokens and self.special_tokens:
+            special_token_ids = set(self.special_tokens.values())
+            token_ids = [tid for tid in token_ids if tid not in special_token_ids]
+
+        return self.tokenizer.decode(token_ids)
+
     def serialize_model(self):
         return torch.ByteTensor(list(self.tokenizer.serialized_model_proto()))