[Bagel]Add image edit (#588)

Signed-off-by: princepride <wangzhipeng628@gmail.com>

vllm_omni/diffusion/models/bagel/bagel_transformer.py

@@ -29,6 +29,43 @@ from vllm.vllm_flash_attn import flash_attn_varlen_func
 
 from vllm_omni.diffusion.layers.rope import RotaryEmbedding
 
+
+def patchify(imgs, p):
+    """
+    imgs: (N, 3, H, W) or (3, H, W)
+    x: (N, L, patch_size**2 *3) or (L, patch_size**2 *3)
+    """
+    is_batch = imgs.ndim == 4
+    if not is_batch:
+        imgs = imgs.unsqueeze(0)
+
+    # n: batch, c: channel, h: grid_h, p: patch_h, w: grid_w, q: patch_w
+    x = imgs.reshape(imgs.shape[0], 3, imgs.shape[2] // p, p, imgs.shape[3] // p, p)
+    # Permute to (n, grid_h, grid_w, c, patch_h, patch_w) to match Conv2d (c, h, w) flattening
+    x = torch.einsum("nchpwq->nhwcpq", x)
+    x = x.reshape(imgs.shape[0], -1, 3 * p**2)
+
+    if not is_batch:
+        x = x.squeeze(0)
+    return x
+
+
+class MLPconnector(nn.Module):
+    def __init__(self, input_dim, output_dim, activation="gelu_pytorch_tanh"):
+        super().__init__()
+        self.fc1 = nn.Linear(input_dim, output_dim)
+        if activation == "gelu":
+            self.act = nn.GELU()
+        elif activation == "gelu_pytorch_tanh":
+            self.act = nn.GELU(approximate="tanh")
+        else:
+            self.act = nn.ReLU()
+        self.fc2 = nn.Linear(output_dim, output_dim)
+
+    def forward(self, x):
+        return self.fc2(self.act(self.fc1(x)))
+
+
 torch._dynamo.config.cache_size_limit = 512
 torch._dynamo.config.accumulated_cache_size_limit = 4096
 flex_attention = torch.compile(flex_attention)
@@ -600,18 +637,30 @@ def get_flattened_position_ids_extrapolate(img_h, img_w, patch_size, max_num_pat
 class BagelConfig(PretrainedConfig):
     def __init__(
         self,
+        visual_gen=True,
+        visual_und=True,
         llm_config=None,
+        vit_config=None,
         vae_config=None,
         latent_patch_size=2,
         max_latent_size=32,
+        vit_max_num_patch_per_side=70,
+        connector_act="gelu_pytorch_tanh",
+        interpolate_pos=False,
         timestep_shift=1.0,
         **kwargs,
     ):
         super().__init__(**kwargs)
+        self.visual_gen = visual_gen
+        self.visual_und = visual_und
         self.llm_config = llm_config
+        self.vit_config = vit_config
         self.vae_config = vae_config
         self.latent_patch_size = latent_patch_size
         self.max_latent_size = max_latent_size
+        self.vit_max_num_patch_per_side = vit_max_num_patch_per_side
+        self.connector_act = connector_act
+        self.interpolate_pos = interpolate_pos
         self.timestep_shift = timestep_shift
 
 
@@ -619,23 +668,32 @@ class Bagel(torch.nn.Module):
     config_class = BagelConfig
     base_model_prefix = "bagel"
 
-    def __init__(self, language_model, config: BagelConfig):
+    def __init__(self, language_model, vit_model, config: BagelConfig):
         super().__init__()
         self.language_model = language_model
         self.hidden_size = config.llm_config.hidden_size
         self.use_moe = "Mo" in config.llm_config.layer_module
         self.num_heads = config.llm_config.num_attention_heads
 
-        self.latent_patch_size = config.latent_patch_size
-        self.timestep_shift = config.timestep_shift
-        self.latent_downsample = config.vae_config.downsample * config.latent_patch_size
-        self.max_latent_size = config.max_latent_size
-        self.latent_channel = config.vae_config.z_channels
-        self.patch_latent_dim = self.latent_patch_size**2 * self.latent_channel
-        self.time_embedder = TimestepEmbedder(self.hidden_size)
-        self.vae2llm = nn.Linear(self.patch_latent_dim, self.hidden_size)
-        self.llm2vae = nn.Linear(self.hidden_size, self.patch_latent_dim)
-        self.latent_pos_embed = PositionEmbedding(self.max_latent_size, self.hidden_size)
+        if config.visual_gen:
+            self.latent_patch_size = config.latent_patch_size
+            self.timestep_shift = config.timestep_shift
+            self.latent_downsample = config.vae_config.downsample * config.latent_patch_size
+            self.max_latent_size = config.max_latent_size
+            self.latent_channel = config.vae_config.z_channels
+            self.patch_latent_dim = self.latent_patch_size**2 * self.latent_channel
+            self.time_embedder = TimestepEmbedder(self.hidden_size)
+            self.vae2llm = nn.Linear(self.patch_latent_dim, self.hidden_size)
+            self.llm2vae = nn.Linear(self.hidden_size, self.patch_latent_dim)
+            self.latent_pos_embed = PositionEmbedding(self.max_latent_size, self.hidden_size)
+
+        if config.visual_und:
+            self.vit_model = vit_model
+            self.vit_patch_size = config.vit_config.patch_size
+            self.vit_max_num_patch_per_side = config.vit_max_num_patch_per_side
+            self.vit_hidden_size = config.vit_config.hidden_size
+            self.connector = MLPconnector(self.vit_hidden_size, self.hidden_size, config.connector_act)
+            self.vit_pos_embed = PositionEmbedding(self.vit_max_num_patch_per_side, self.hidden_size)
 
         self.get_flattened_position_ids = get_flattened_position_ids_extrapolate
 
@@ -643,8 +701,9 @@ class Bagel(torch.nn.Module):
         self._init_weights()
 
     def _init_weights(self):
-        nn.init.constant_(self.llm2vae.weight, 0)
-        nn.init.constant_(self.llm2vae.bias, 0)
+        if self.config.visual_gen:
+            nn.init.constant_(self.llm2vae.weight, 0)
+            nn.init.constant_(self.llm2vae.bias, 0)
 
     def prepare_prompts(self, curr_kvlens, curr_rope, prompts, tokenizer, new_token_ids):
         packed_text_ids = list()
@@ -713,6 +772,261 @@ class Bagel(torch.nn.Module):
 
         return past_key_values
 
+    def prepare_vae_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids, timestep=0):
+        patchified_vae_latent_shapes, packed_vae_position_ids = list(), list()
+        packed_vae_token_indexes = list()
+        packed_text_ids, packed_text_indexes = list(), list()
+        packed_seqlens, packed_position_ids, packed_indexes = list(), list(), list()
+        packed_key_value_indexes = list()
+
+        _curr = curr = 0
+        vae_image_tensors = list()
+        newlens, new_rope = list(), list()
+        for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            packed_text_ids.append(new_token_ids["start_of_image"])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            image_tensor = transforms(image)
+            vae_image_tensors.append(image_tensor)
+            vae_position_ids = self.get_flattened_position_ids(
+                image_tensor.size(1),
+                image_tensor.size(2),
+                self.latent_downsample,
+                max_num_patches_per_side=self.max_latent_size,
+            )
+            packed_vae_position_ids.append(vae_position_ids)
+            H, W = image_tensor.shape[1:]
+            h = H // self.latent_downsample
+            w = W // self.latent_downsample
+            patchified_vae_latent_shapes.append((h, w))
+
+            num_img_tokens = w * h
+            packed_vae_token_indexes.extend(range(_curr, _curr + num_img_tokens))
+            packed_indexes.extend(range(curr, curr + num_img_tokens))
+            curr += num_img_tokens
+            _curr += num_img_tokens
+
+            packed_text_ids.append(new_token_ids["end_of_image"])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            packed_position_ids.extend([curr_position_id] * (num_img_tokens + 2))
+            packed_seqlens.append(num_img_tokens + 2)
+            newlens.append(curr_kvlen + num_img_tokens + 2)
+            new_rope.append(curr_position_id + 1)
+
+        image_sizes = [item.shape for item in vae_image_tensors]
+        max_image_size = [max(item) for item in list(zip(*image_sizes))]
+        padded_images = torch.zeros(size=(len(vae_image_tensors), *max_image_size))
+        for i, image_tensor in enumerate(vae_image_tensors):
+            padded_images[i, :, : image_tensor.shape[1], : image_tensor.shape[2]] = image_tensor
+
+        generation_input = {
+            "padded_images": padded_images,
+            "patchified_vae_latent_shapes": patchified_vae_latent_shapes,
+            "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0),
+            "packed_timesteps": torch.tensor([timestep]),
+            "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long),
+            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
+            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
+            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
+            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
+            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+        }
+
+        return generation_input, newlens, new_rope
+
+    @torch.no_grad
+    def forward_cache_update_vae(
+        self,
+        vae_model,
+        past_key_values: NaiveCache,
+        padded_images: torch.Tensor,
+        patchified_vae_latent_shapes: list,
+        packed_vae_position_ids: torch.LongTensor,
+        packed_timesteps: torch.Tensor,
+        packed_vae_token_indexes: torch.LongTensor,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_position_ids: torch.LongTensor,
+        packed_seqlens: torch.IntTensor,
+        packed_indexes: torch.LongTensor,
+        key_values_lens: torch.IntTensor,
+        packed_key_value_indexes: torch.Tensor,
+    ):
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
+        packed_sequence[packed_text_indexes] = packed_text_embedding
+
+        padded_latent = vae_model.encode(padded_images)
+
+        p = self.latent_patch_size
+        packed_latent = list()
+        for latent, (h, w) in zip(padded_latent, patchified_vae_latent_shapes):
+            latent = latent[:, : h * p, : w * p].reshape(self.latent_channel, h, p, w, p)
+            latent = torch.einsum("chpwq->hwpqc", latent).reshape(-1, p * p * self.latent_channel)
+            packed_latent.append(latent)
+        packed_latent = torch.cat(packed_latent, dim=0)
+        packed_pos_embed = self.latent_pos_embed(packed_vae_position_ids)
+        packed_timestep_embeds = self.time_embedder(packed_timesteps)
+        packed_latent = self.vae2llm(packed_latent) + packed_timestep_embeds + packed_pos_embed
+        if packed_latent.dtype != packed_sequence.dtype:
+            packed_latent = packed_latent.to(packed_sequence.dtype)
+        packed_sequence[packed_vae_token_indexes] = packed_latent
+
+        extra_inputs = {}
+        if self.use_moe:
+            extra_inputs = {
+                "mode": "gen",
+                "packed_vae_token_indexes": packed_vae_token_indexes,
+                "packed_text_indexes": packed_text_indexes,
+            }
+
+        output = self.language_model.forward(
+            packed_query_sequence=packed_sequence,
+            query_lens=packed_seqlens,
+            packed_query_position_ids=packed_position_ids,
+            packed_query_indexes=packed_indexes,
+            past_key_values=past_key_values,
+            key_values_lens=key_values_lens,
+            packed_key_value_indexes=packed_key_value_indexes,
+            update_past_key_values=True,
+            is_causal=False,
+            **extra_inputs,
+        )
+        past_key_values = output.past_key_values
+
+        return past_key_values
+
+    def prepare_vit_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids):
+        packed_vit_token_indexes = list()
+        vit_token_seqlens, packed_vit_tokens, packed_vit_position_ids = list(), list(), list()
+        packed_text_ids, packed_text_indexes = list(), list()
+        packed_seqlens, packed_position_ids, packed_indexes = list(), list(), list()
+        packed_key_value_indexes = list()
+
+        _curr = curr = 0
+        newlens, new_rope = list(), list()
+        for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            packed_text_ids.append(new_token_ids["start_of_image"])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            image_tensor = transforms(image)
+            vit_position_ids = self.get_flattened_position_ids(
+                image_tensor.size(1),
+                image_tensor.size(2),
+                self.vit_patch_size,
+                max_num_patches_per_side=self.vit_max_num_patch_per_side,
+            )
+            vit_tokens = patchify(image_tensor, self.vit_patch_size)
+            packed_vit_tokens.append(vit_tokens)
+            num_img_tokens = vit_tokens.shape[0]
+            packed_vit_position_ids.append(vit_position_ids)
+            vit_token_seqlens.append(num_img_tokens)
+            packed_vit_token_indexes.extend(range(_curr, _curr + num_img_tokens))
+            packed_indexes.extend(range(curr, curr + num_img_tokens))
+            curr += num_img_tokens
+            _curr += num_img_tokens
+
+            packed_text_ids.append(new_token_ids["end_of_image"])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            packed_position_ids.extend([curr_position_id] * (num_img_tokens + 2))
+            packed_seqlens.append(num_img_tokens + 2)
+            newlens.append(curr_kvlen + num_img_tokens + 2)
+            new_rope.append(curr_position_id + 1)
+
+        generation_input = {
+            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
+            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
+            "vit_token_seqlens": torch.tensor(vit_token_seqlens, dtype=torch.int),
+            "packed_vit_tokens": torch.cat(packed_vit_tokens, dim=0),
+            "packed_vit_position_ids": torch.cat(packed_vit_position_ids, dim=0),
+            "packed_vit_token_indexes": torch.tensor(packed_vit_token_indexes, dtype=torch.long),
+            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
+            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
+            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+        }
+
+        return generation_input, newlens, new_rope
+
+    @torch.no_grad
+    def forward_cache_update_vit(
+        self,
+        past_key_values: NaiveCache,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_vit_tokens: torch.Tensor,
+        packed_vit_token_indexes: torch.LongTensor,
+        packed_vit_position_ids: torch.LongTensor,
+        vit_token_seqlens: torch.IntTensor,
+        packed_position_ids: torch.LongTensor,
+        packed_seqlens: torch.IntTensor,
+        packed_indexes: torch.LongTensor,
+        packed_key_value_indexes: torch.LongTensor,
+        key_values_lens: torch.IntTensor,
+    ):
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
+        packed_sequence[packed_text_indexes] = packed_text_embedding
+
+        cu_seqlens = torch.nn.functional.pad(torch.cumsum(vit_token_seqlens, dim=0), (1, 0))
+        cu_seqlens = cu_seqlens.to(torch.int32)
+        max_seqlen = torch.max(vit_token_seqlens).item()
+        packed_vit_token_embed = self.vit_model(
+            packed_pixel_values=packed_vit_tokens,
+            packed_flattened_position_ids=packed_vit_position_ids,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+        )
+        packed_vit_token_embed = self.connector(packed_vit_token_embed)
+        pos_emb = self.vit_pos_embed(packed_vit_position_ids)
+        packed_vit_token_embed = packed_vit_token_embed + pos_emb
+        if packed_vit_token_embed.dtype != packed_sequence.dtype:
+            packed_vit_token_embed = packed_vit_token_embed.to(packed_sequence.dtype)
+        packed_sequence[packed_vit_token_indexes] = packed_vit_token_embed
+
+        extra_inputs = {}
+        if self.use_moe:
+            extra_inputs = {"mode": "und"}
+
+        output = self.language_model.forward(
+            packed_query_sequence=packed_sequence,
+            query_lens=packed_seqlens,
+            packed_query_position_ids=packed_position_ids,
+            packed_query_indexes=packed_indexes,
+            past_key_values=past_key_values,
+            packed_key_value_indexes=packed_key_value_indexes,
+            key_values_lens=key_values_lens,
+            update_past_key_values=True,
+            is_causal=False,
+            **extra_inputs,
+        )
+        past_key_values = output.past_key_values
+
+        return past_key_values
+
     def prepare_input(self, curr_kvlens, curr_rope, image_sizes, new_token_ids=None):
         packed_text_ids, packed_text_indexes = list(), list()
         packed_vae_position_ids, packed_vae_token_indexes, packed_init_noises = list(), list(), list()

vllm_omni/diffusion/models/bagel/pipeline_bagel.py

@@ -12,10 +12,12 @@ from collections.abc import Iterable
 from dataclasses import dataclass
 from math import isqrt
 
+import numpy as np
 import torch
+import torch.nn.functional as F
 from PIL import Image
 from torch import nn
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, SiglipImageProcessor, SiglipVisionConfig, SiglipVisionModel
 from vllm.logger import init_logger
 from vllm.model_executor.models.utils import AutoWeightsLoader
 
@@ -89,6 +91,12 @@ class _VaeCfg:
     downsample: int = 8
 
 
+@dataclass
+class _VitCfg:
+    patch_size: int = 14
+    hidden_size: int = 1152
+
+
 def default_ae_params() -> AutoEncoderParams:
     return AutoEncoderParams(
         resolution=256,
@@ -104,6 +112,37 @@ def default_ae_params() -> AutoEncoderParams:
     )
 
 
+class SiglipNaViTWrapper(nn.Module):
+    def __init__(self, vision_model):
+        super().__init__()
+        # If input is SiglipVisionModel, unwrap it to get SiglipVisionTransformer
+        if hasattr(vision_model, "vision_model"):
+            self.vision_model = vision_model.vision_model
+        else:
+            self.vision_model = vision_model
+
+        # Configure weights for linear equivalent of patch embedding
+        self.patch_embed_weight = self.vision_model.embeddings.patch_embedding.weight
+        self.patch_embed_bias = self.vision_model.embeddings.patch_embedding.bias
+
+    def forward(self, packed_pixel_values, packed_flattened_position_ids, cu_seqlens, max_seqlen):
+        w = self.patch_embed_weight.view(self.patch_embed_weight.shape[0], -1)
+        x = F.linear(packed_pixel_values, w, self.patch_embed_bias)
+        pos = self.vision_model.embeddings.position_embedding(packed_flattened_position_ids)
+        x = x + pos
+        hidden_states = x.unsqueeze(0)
+        seq_len = x.shape[0]
+        mask = torch.full((1, 1, seq_len, seq_len), torch.finfo(x.dtype).min, device=x.device, dtype=x.dtype)
+        cu_seqlens_list = cu_seqlens.tolist()
+        for i in range(len(cu_seqlens_list) - 1):
+            start = cu_seqlens_list[i]
+            end = cu_seqlens_list[i + 1]
+            mask[..., start:end, start:end] = 0.0
+
+        outputs = self.vision_model.encoder(inputs_embeds=hidden_states, attention_mask=mask)
+        return outputs.last_hidden_state.squeeze(0)
+
+
 class BagelPipeline(nn.Module):
     """Bagel generation pipeline (MoT) packaged for vllm-omni diffusion engine.
 
@@ -150,12 +189,26 @@ class BagelPipeline(nn.Module):
             local_files_only=True,
             trust_remote_code=True,
         )
+
+        # Try finding vision_config or interpolate from top-level config
+        vit_cfg_dict = bagel_cfg.get("vit_config") or {}
+        vit_cfg = _VitCfg(
+            patch_size=int(vit_cfg_dict.get("patch_size", 14)),
+            hidden_size=int(vit_cfg_dict.get("hidden_size", 1152)),
+        )
+        vit_config_path = os.path.join(model_path, "vit_config.json")
+        vit_conf = SiglipVisionConfig.from_json_file(vit_config_path)
+        self.vit_model = SiglipVisionModel(vit_conf)
+        self.image_processor = SiglipImageProcessor.from_pretrained(model_path, local_files_only=True)
+
+        if self.vit_model:
+            self.vit_model = SiglipNaViTWrapper(self.vit_model)
+            vit_cfg.hidden_size = self.vit_model.vision_model.config.hidden_size
+            vit_cfg.patch_size = self.vit_model.vision_model.config.patch_size
+
         self.tokenizer, self.new_token_ids, _ = add_special_tokens(self.tokenizer)
 
-        try:
-            tok_len = len(self.tokenizer)
-        except Exception:  # pragma: no cover - very old tokenizers
-            tok_len = getattr(self.tokenizer, "vocab_size", llm_config.vocab_size)
+        tok_len = len(self.tokenizer)
         required_max_id = max(int(v) for v in self.new_token_ids.values())
         llm_config.vocab_size = max(
             int(getattr(llm_config, "vocab_size", tok_len)),
@@ -169,9 +222,14 @@ class BagelPipeline(nn.Module):
 
         self.bagel = Bagel(
             language_model=self.language_model,
+            vit_model=self.vit_model,
             config=BagelConfig(
                 llm_config=llm_config,
                 vae_config=vae_cfg,
+                vit_config=vit_cfg,
+                vit_max_num_patch_per_side=int(bagel_cfg.get("vit_max_num_patch_per_side", 70)),
+                connector_act=str(bagel_cfg.get("connector_act", "gelu_pytorch_tanh")),
+                interpolate_pos=bool(bagel_cfg.get("interpolate_pos", False)),
                 latent_patch_size=int(bagel_cfg.get("latent_patch_size", 2)),
                 max_latent_size=int(bagel_cfg.get("max_latent_size", 32)),
                 timestep_shift=float(bagel_cfg.get("timestep_shift", 1.0)),
@@ -248,6 +306,69 @@ class BagelPipeline(nn.Module):
         injected_kv = getattr(req, "past_key_values", None)
         injected_metadata = getattr(req, "kv_metadata", None)
 
+        # Image input handling
+        image_input = getattr(req, "pil_image", None)
+        if image_input and not isinstance(image_input, list):
+            image_input = [image_input]
+
+        if image_input:
+            # If we have an image, we prefill with it
+            if self.image_processor and self.vae:
+
+                def vit_transforms(img):
+                    # SigLIP processor returns dict with pixel_values; we want the tensor
+                    return self.image_processor(images=img, return_tensors="pt").pixel_values[0]
+
+                def vae_transforms(img):
+                    if img.mode != "RGB":
+                        img = img.convert("RGB")
+                    # Convert to [-1, 1] tensor (H, W, C) -> (C, H, W)
+                    arr = torch.from_numpy(np.array(img)).float() / 127.5 - 1.0
+                    return arr.permute(2, 0, 1)
+
+                # 1. Update VAE
+                gen_input_vae, newlens_vae, new_rope_vae = self.bagel.prepare_vae_images(
+                    curr_kvlens=gen_context["kv_lens"],
+                    curr_rope=gen_context["ropes"],
+                    images=image_input,
+                    transforms=vae_transforms,
+                    new_token_ids=self.new_token_ids,
+                )
+
+                for k, v in gen_input_vae.items():
+                    if torch.is_tensor(v):
+                        gen_input_vae[k] = v.to(self.device)
+
+                # VAE needs bfloat16 to match model strings usually, specifically encode
+                with torch.autocast(device_type="cuda", enabled=self.device.type == "cuda", dtype=torch.bfloat16):
+                    gen_context["past_key_values"] = self.bagel.forward_cache_update_vae(
+                        self.vae, gen_context["past_key_values"], **gen_input_vae
+                    )
+                gen_context["kv_lens"] = newlens_vae
+                gen_context["ropes"] = new_rope_vae
+
+                # 2. Update ViT
+                gen_input_img, newlens_img, new_rope_img = self.bagel.prepare_vit_images(
+                    curr_kvlens=gen_context["kv_lens"],
+                    curr_rope=gen_context["ropes"],
+                    images=image_input,
+                    transforms=vit_transforms,
+                    new_token_ids=self.new_token_ids,
+                )
+
+                for k, v in gen_input_img.items():
+                    if torch.is_tensor(v):
+                        gen_input_img[k] = v.to(self.device)
+
+                with torch.autocast(device_type="cuda", enabled=self.device.type == "cuda", dtype=torch.bfloat16):
+                    gen_context["past_key_values"] = self.bagel.forward_cache_update_vit(
+                        gen_context["past_key_values"], **gen_input_img
+                    )
+                gen_context["kv_lens"] = newlens_img
+                gen_context["ropes"] = new_rope_img
+            else:
+                logger.warning("Image provided but no image processor available.")
+
         if injected_kv is not None and injected_metadata is not None:
             logger.info("Using injected KV Cache from remote transfer")
 
@@ -409,6 +530,19 @@ class BagelPipeline(nn.Module):
                     yield "bagel." + n
                     break
 
+            # Map connector and vit_pos_embed to `bagel.*`
+            for pfx in ("connector.", "vit_pos_embed."):
+                if n.startswith(pfx):
+                    yield "bagel." + n
+                    break
+
+            if n.startswith("vit_model."):
+                yield "bagel." + n  # matches self.bagel.vit_model
+            elif n.startswith("vision_model."):
+                yield "bagel.vit_model." + n
+            elif n.startswith("model.vision_model."):
+                yield "bagel.vit_model." + n[len("model.") :]
+
         def _filtered_weights():
             total = 0
             kept = 0
@@ -439,6 +573,17 @@ class BagelPipeline(nn.Module):
                                     shapes[cand] = (npos, hdim)
                                     picked = cand
                                     break
+                            # Handle flattened patch embedding for SigLIP
+                            if cand.endswith("embeddings.patch_embedding.weight") and tensor.ndim == 2:
+                                # Checkpoint has (Hidden, C*P*P), model expects (Hidden, C, P, P)
+                                if shapes.get(cand) is not None:
+                                    target_shape = shapes[cand]
+                                    if tensor.numel() == torch.prod(torch.tensor(target_shape)):
+                                        # Reshape tensor to match target
+                                        tensor = tensor.view(target_shape)
+                                        picked = cand
+                                        break
+
                             shape_mismatch += 1
                             # Keep this quiet; shape mismatches are expected for ignored modules.
                 if picked is not None: