Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Gext_Pt1-596M-F16.gguf filter=lfs diff=lfs merge=lfs -text
+Gext_Pt1-596M-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

Gext_Pt1-596M-F16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47fcfa69c3dc0481d80da53fb5644b12e7678e8bb8c2d49303f7222ecde61247
+size 1198181952

Gext_Pt1-596M-Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:458a00f8db4925da41b4dfe1d7157a568c54823b61a06e195b0f874cc7621765
+size 396704320

added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

config.json

@@ -0,0 +1,58 @@
+{
+  "architectures": [
+    "GexTQwenForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_gex.GexTConfig",
+    "AutoModel": "modeling_gex.GexTQwenForCausalLM",
+    "AutoModelForCausalLM": "modeling_gex.GexTQwenForCausalLM"
+  },
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "gext",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936,
+  "vision_config": {
+    "hidden_size": 768,
+    "in_chans": 3,
+    "intermediate_size": 2073,
+    "depth":12,
+    "fullatt_block_indexes": [
+      2,
+      5,
+      8,
+      11
+    ],
+    "window_size": 14,
+    "model_type": "gotvary",
+    "out_hidden_size": 5120,
+    "spatial_patch_size": 14,
+    "tokens_per_second": 2,
+    "torch_dtype": "bfloat16",
+    "image_size": 1024,
+    "patch_size": 16,
+    "num_attention_heads": 12,
+    "n_head": 12,
+    "num_hidden_layers": 12
+  }
+}

configuration_gex.py

@@ -0,0 +1,7 @@
+from transformers import Qwen2Config,Qwen3Config
+
+class GexConfig(Qwen2Config):
+    model_type = "gex"
+
+class GexTConfig(Qwen3Config):
+    model_type = "gext"

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "transformers_version": "4.51.3"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:515cd3cc86762a87b22ea6d54ad1ea542e9bfaaeede52139305620c60e17951d
+size 1389802360

modeling_gex.py

@@ -0,0 +1,763 @@
+from typing import List, Optional, Tuple, Union
+from functools import partial
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from torch.nn.attention import SDPBackend, sdpa_kernel
+
+from torchvision import transforms
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+
+from torchvision.transforms.functional import InterpolationMode
+from transformers import (
+    Qwen2Config,
+    Qwen2Model,
+    Qwen2ForCausalLM,
+    Qwen3ForCausalLM,
+    Qwen3Model,
+    Qwen3Config,
+)
+
+try:
+    from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+    from liger_kernel.transformers import LigerLayerNorm
+    from liger_kernel.transformers.layer_norm import LigerLayerNormFunction
+
+    def liger_layer_norm(input, normalized_shape, weight, bias, eps):
+        return LigerLayerNormFunction.apply(input, weight, bias, eps)
+
+    use_liger = True
+except ImportError:
+    use_liger = False
+
+
+from .configuration_gex import GexConfig, GexTConfig
+
+
+LayerNorm = (
+    partial(LigerLayerNorm, bias=True) if use_liger else partial(nn.LayerNorm, eps=1e-6)
+)
+layer_norm = liger_layer_norm if use_liger else torch.nn.functional.layer_norm
+
+BOS_TOEKN_IDS: int = 151652
+EOS_TOEKN_IDS: int = 151643
+IMG_PAD_IDS: int = 151655
+IMG_END_IDS: int = 25
+
+
+@torch.no_grad
+def process_batch_labels(labels, pad_token_id=EOS_TOEKN_IDS):
+    # 创建 mask：标记所有 pad_token_id 的位置
+    pad_mask = labels == pad_token_id
+
+    # 找到每个样本第一个 pad_token_id 的位置
+    first_pad_pos = pad_mask.int().argmax(dim=1, keepdim=True)  # shape: (bsz,)
+    first_pad_pos[first_pad_pos == 0] = 256
+
+    # 生成要替换为 -100 的位置 mask
+    replace_mask = torch.arange(labels.size(1), device=labels.device) > first_pad_pos
+
+    # 执行替换（保留第一个 pad_token_id）
+    labels[replace_mask] = -100
+
+    return labels
+
+
+class GexImageEvalProcessor:
+    def __init__(self, image_size=1024, mean=None, std=None):
+        if mean is None:
+            mean = (0.48145466, 0.4578275, 0.40821073)
+        if std is None:
+            std = (0.26862954, 0.26130258, 0.27577711)
+
+        self.normalize = transforms.Normalize(mean, std)
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(
+                    (image_size, image_size), interpolation=InterpolationMode.BICUBIC
+                ),
+                transforms.ToTensor(),
+                self.normalize,
+            ]
+        )
+
+    def __call__(self, item):
+        return self.transform(item)
+
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.num_channels = num_channels
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.permute(0, 2, 3, 1)
+        return layer_norm(
+            x,
+            normalized_shape=(self.num_channels,),
+            weight=self.weight,
+            bias=self.bias,
+            eps=self.eps,
+        ).permute(0, 3, 1, 2)
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = 64
+        self.scale = 64**-0.5
+        self.seq_len = input_size[0] * input_size[1]
+        self.input_size = input_size
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=True)
+        self.proj = nn.Linear(dim, dim)
+
+        # self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, self.head_dim))
+        # self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, self.head_dim))
+        self.rel_pos_h = nn.Parameter(
+            torch.zeros(input_size[0], input_size[0], self.head_dim)
+        )
+        self.rel_pos_w = nn.Parameter(
+            torch.zeros(input_size[1], input_size[1], self.head_dim)
+        )
+
+    def init_rel_pos(self):
+        q_size, k_size = self.input_size
+        q_coords = torch.arange(q_size)[:, None]
+
+        k_coords = torch.arange(k_size)[None, :]
+        relative_coords = (q_coords - k_coords) + (k_size - 1)
+
+        self.rel_pos_h = nn.Parameter(self.rel_pos_h.data[relative_coords.long()])
+        self.rel_pos_w = nn.Parameter(self.rel_pos_w.data[relative_coords.long()])
+
+    def get_attn_bias(self, q: torch.Tensor):
+        q = q.view(-1, *self.input_size, 64)
+
+        rel_h = torch.einsum("bhwc,hkc->bhwk", q, self.rel_pos_h)
+        rel_w = torch.einsum("bhwc,wkc->bhwk", q, self.rel_pos_w)
+
+        return (rel_h.unsqueeze(-1) + rel_w.unsqueeze(-2)).reshape(
+            -1, self.num_heads, self.seq_len, self.seq_len
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        qkv = torch.split(
+            self.qkv(x).view(-1, self.seq_len, 3 * 768),
+            768,
+            dim=2,
+        )
+
+        q, k, v = (
+            i.unflatten(-1, (self.num_heads, -1)).transpose(1, 2).contiguous()
+            for i in qkv
+        )
+
+        attn_bias = self.get_attn_bias(q)
+        with sdpa_kernel(
+            [
+                SDPBackend.FLASH_ATTENTION,
+                SDPBackend.CUDNN_ATTENTION,
+                SDPBackend.EFFICIENT_ATTENTION,
+            ],
+            set_priority=True,
+        ):
+            attn_output = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=attn_bias, is_causal=False
+            )
+        attn_output = attn_output.transpose(1, 2).flatten(-2)
+
+        x = self.proj(attn_output)
+
+        return x.view(-1, *self.input_size, 768)
+
+
+class MLP(nn.Module):
+    def __init__(
+        self,
+    ):
+        super().__init__()
+        self.lin1 = nn.Linear(768, 4 * 768)
+        self.lin2 = nn.Linear(4 * 768, 768)
+        self.act = nn.GELU()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lin2(self.act(self.lin1(x)))
+
+
+class Block(nn.Module):
+    def __init__(self, idx: int, window_size: int = 14):
+        super().__init__()
+
+        self.idx = idx
+        self.window_size = window_size
+
+        self.norm1 = LayerNorm(768)
+
+        self.attn = Attention(
+            dim=768,
+            num_heads=12,
+            input_size=(64, 64) if window_size == 0 else (14, 14),
+        )
+
+        self.norm2 = LayerNorm(768)
+        self.mlp = MLP()
+
+    @staticmethod
+    def window_partition(x: torch.Tensor) -> torch.Tensor:
+        x = F.pad(x, (0, 0, 0, 6, 0, 6))
+        x = (
+            x.view(-1, 5, 14, 5, 14, 768)
+            .permute(0, 1, 3, 2, 4, 5)
+            .contiguous()
+            .view(-1, 14, 14, 768)
+        )
+        return x
+
+    @staticmethod
+    def window_unpartition(x: torch.Tensor) -> torch.Tensor:
+        x = (
+            x.view(-1, 5, 5, 14, 14, 768)
+            .permute(0, 1, 3, 2, 4, 5)
+            .contiguous()
+            .view(-1, 70, 70, 768)
+        )
+        return x[:, :64, :64, :].contiguous()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        if self.window_size > 0:
+            x = self.window_partition(x)
+
+        x = self.attn(x)
+
+        if self.window_size > 0:
+            x = self.window_unpartition(x)
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class GexVit(nn.Module):
+    def __init__(self, global_attn_indexes=[2, 5, 8, 11], **kwargs):
+        super().__init__()
+        self.global_attn_indexes = global_attn_indexes
+        self.patch_embed = PatchEmbed()
+
+        self.pos_embed = nn.Parameter(torch.zeros(1, 64, 64, 768))
+
+        self.blocks = nn.ModuleList(
+            [
+                Block(idx=i, window_size=14 if i not in global_attn_indexes else 0)
+                for i in range(12)
+            ]
+        )
+
+        self.neck = nn.ModuleList(
+            [
+                nn.Conv2d(
+                    768,
+                    256,
+                    kernel_size=1,
+                    bias=False,
+                ),
+                LayerNorm2d(256),
+                nn.Conv2d(
+                    256,
+                    256,
+                    kernel_size=3,
+                    padding=1,
+                    bias=False,
+                ),
+                LayerNorm2d(256),
+            ]
+        )
+
+        self.net_2 = nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1, bias=False)
+        self.net_3 = nn.Conv2d(
+            512, 1024, kernel_size=3, stride=2, padding=1, bias=False
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        x = x + self.pos_embed
+
+        for blk in self.blocks:
+            x = blk(x)
+
+        x = x.permute(0, 3, 1, 2)
+
+        for m in self.neck:
+            x = m(x)
+
+        x = self.net_2(x)
+        x = self.net_3(x)
+
+        return x
+
+
+class GexQwenModel(Qwen2Model):
+    config_class = GexConfig
+    _auto_class = "AutoModel"
+
+    def __init__(self, config: Qwen2Config):
+        super().__init__(config)
+        self.vit = GexVit()
+        self.vit_proj = nn.Linear(1024, 1024)
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if inputs_embeds is None and input_ids is not None:
+            inputs_embeds = self.embed_tokens(input_ids)
+            assert images is not None
+            # img_pos = input_ids == IMG_PAD_IDS
+            # if torch.any(img_pos):
+            vit_feature = self.vit(images).flatten(2).permute(0, 2, 1)
+            vit_feature = self.vit_proj(vit_feature)
+            # img_ids = img_pos.nonzero().squeeze_()
+            # inputs_embeds[img_ids[:, 0], img_ids[:, 1]] = vit_feature.view(-1,1024)
+            inputs_embeds[:, 1:257, :] = vit_feature
+        with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
+            return super().forward(
+                input_ids=None,
+                attention_mask=attention_mask,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds,
+                use_cache=use_cache,
+                position_ids=position_ids,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                **kwargs,
+            )
+
+
+class GexQwenForCausalLM(Qwen2ForCausalLM):
+    config_class = GexConfig
+    # supports_gradient_checkpointing = True
+    _auto_class = "AutoModelForCausalLM"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = GexQwenModel(config)
+
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        self.image_preprocess = GexImageEvalProcessor()
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        images: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if labels is not None and input_ids is None:
+            input_ids: torch.Tensor = labels
+            shifted_input_ids = input_ids.new_zeros(
+                (input_ids.shape[0], input_ids.shape[1] + 256), device=input_ids.device
+            )
+            shifted_input_ids[:, 257:].copy_(input_ids[:, :-1])
+            decoder_start_token_id = BOS_TOEKN_IDS
+            shifted_input_ids[:, 0] = decoder_start_token_id
+            shifted_input_ids[:, 1:257] = IMG_PAD_IDS
+            input_ids = shifted_input_ids
+            imgs_pad: torch.Tenosr = torch.full(
+                (1, 256), IMG_PAD_IDS, device=self.device, dtype=torch.long
+            )
+            labels = torch.cat(
+                [
+                    imgs_pad.expand(labels.shape[0], -1),
+                    process_batch_labels(labels),
+                ],
+                dim=-1,
+            )
+            # labels = process_batch_labels(labels)
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            images=images,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = (
+            slice(-logits_to_keep, None)
+            if isinstance(logits_to_keep, int)
+            else logits_to_keep
+        )
+
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        # if (past_key_values is None or len(past_key_values) <= 0):
+        #     logits = self.lm_head(hidden_states[:, 256:, :])
+        #     # if labels is not None:
+        #     #     lb = labels[:,256:].contiguous()
+        #     #     del labels
+        #     #     labels = lb
+        # else:
+        #     slice_indices = (
+        #         slice(-logits_to_keep, None)
+        #         if isinstance(logits_to_keep, int)
+        #         else logits_to_keep
+        #     )
+        #     logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=None,
+                shift_labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def generate(self, *args, images, **kwargs):
+        pad = torch.tensor(
+            [[BOS_TOEKN_IDS] + [IMG_PAD_IDS] * 256],
+            dtype=torch.long,
+            device=self.device,
+        )
+        if (input_ids := kwargs.pop("input_ids", None)) is not None:
+            input_ids = torch.cat(
+                [pad.expand(input_ids.shape[0], -1), input_ids], dim=-1
+            )
+        else:
+            input_ids = pad.expand(images.shape[0], -1)
+
+        res = super().generate(
+            *args,
+            input_ids=input_ids,
+            images=images,
+            max_length=kwargs.pop("max_length", 10) + 257,
+            **kwargs,
+        )
+        return res
+
+
+class GexTQwenModel(Qwen3Model):
+    config_class = GexTConfig
+    _auto_class = "AutoModel"
+
+    def __init__(self, config: Qwen3Config):
+        super().__init__(config)
+        self.vit = GexVit()
+        self.vit_proj = nn.Linear(1024, 1024)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if inputs_embeds is None and input_ids is not None:
+            inputs_embeds = self.embed_tokens(input_ids)
+            assert images is not None
+            # img_pos = input_ids == IMG_PAD_IDS
+            # if torch.any(img_pos):
+            vit_feature = self.vit(images).flatten(2).permute(0, 2, 1)
+            vit_feature = self.vit_proj(vit_feature)
+            # img_ids = img_pos.nonzero().squeeze_()
+            # inputs_embeds[img_ids[:, 0], img_ids[:, 1]] = vit_feature.view(-1,1024)
+            inputs_embeds[:, 1:257, :] = vit_feature
+        with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
+            return super().forward(
+                input_ids=None,
+                attention_mask=attention_mask,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds,
+                use_cache=use_cache,
+                position_ids=position_ids,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                cache_position=cache_position,
+                **flash_attn_kwargs,
+            )
+
+
+class GexTQwenForCausalLM(Qwen3ForCausalLM):
+    config_class = GexTConfig
+    # supports_gradient_checkpointing = True
+    _auto_class = "AutoModelForCausalLM"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = GexTQwenModel(config)
+
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        self.image_preprocess = GexImageEvalProcessor()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        images: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+
+        if labels is not None and input_ids is None:
+            input_ids: torch.Tensor = labels
+            shifted_input_ids = input_ids.new_zeros(
+                (input_ids.shape[0], input_ids.shape[1] + 257), device=input_ids.device
+            )
+            shifted_input_ids[:, 258:].copy_(input_ids[:, :-1])
+            decoder_start_token_id = BOS_TOEKN_IDS
+            shifted_input_ids[:, 0] = decoder_start_token_id
+            shifted_input_ids[:, 257] = IMG_END_IDS
+            shifted_input_ids[:, 1:257] = IMG_PAD_IDS
+            input_ids = shifted_input_ids
+            imgs_pad: torch.Tenosr = torch.full(  # type: ignore
+                (1, 257), IMG_PAD_IDS, device=self.device, dtype=torch.long
+            )
+            imgs_pad[:, -1] = IMG_END_IDS
+            labels = torch.cat(
+                [
+                    imgs_pad.expand(labels.shape[0], -1),
+                    process_batch_labels(labels),
+                ],
+                dim=-1,
+            )  # type: ignore
+            # labels = process_batch_labels(labels)
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            cache_position=cache_position,
+            images=images,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+
+        # if (past_key_values is None or len(past_key_values) <= 0):
+        #     logits = self.lm_head(hidden_states[:, 256:, :])
+        #     # if labels is not None:
+        #     #     lb = labels[:,256:].contiguous()
+        #     #     del labels
+        #     #     labels = lb
+        # else:
+        #     slice_indices = (
+        #         slice(-logits_to_keep, None)
+        #         if isinstance(logits_to_keep, int)
+        #         else logits_to_keep
+        #     )
+        #     logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            if self.training and use_liger:
+                loss = LigerForCausalLMLoss(
+                    hidden_states=hidden_states,
+                    lm_head_weight=self.lm_head.weight,
+                    labels=None,
+                    shift_labels=labels,
+                    hidden_size=self.config.hidden_size,
+                    **kwargs,
+                )
+                logits = None
+
+            else:
+                slice_indices = (
+                    slice(-logits_to_keep, None)
+                    if isinstance(logits_to_keep, int)
+                    else logits_to_keep
+                )
+
+                logits = self.lm_head(hidden_states[:, slice_indices, :])
+                loss = self.loss_function(
+                    logits=logits,
+                    labels=None,
+                    shift_labels=labels,
+                    vocab_size=self.config.vocab_size,
+                    **kwargs,
+                )
+        else:
+            slice_indices = (
+                slice(-logits_to_keep, None)
+                if isinstance(logits_to_keep, int)
+                else logits_to_keep
+            )
+
+            logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def generate(self, *args, images, **kwargs):
+        pad = torch.tensor(
+            [[BOS_TOEKN_IDS] + [IMG_PAD_IDS] * 256 + [IMG_END_IDS]],
+            dtype=torch.long,
+            device=self.device,
+        )
+        if (input_ids := kwargs.pop("input_ids", None)) is not None:
+            input_ids = torch.cat(
+                [pad.expand(input_ids.shape[0], -1), input_ids], dim=-1
+            )
+        else:
+            input_ids = pad.expand(images.shape[0], -1)
+
+        res = super().generate(
+            *args,
+            input_ids=input_ids,
+            images=images,
+            max_length=kwargs.pop("max_length", 25) + 258,
+            **kwargs,
+        )
+        return res

preprocessor_config.json

@@ -0,0 +1,14 @@
+{
+    "image_size": 1024,
+    "image_mean": [
+        0.48145466,
+        0.4578275,
+        0.40821073
+    ],
+    "image_std": [
+        0.26862954,
+        0.26130258,
+        0.27577711
+    ],
+    "image_processor_type": "GexImageProcessor"
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654

tokenizer_config.json

@@ -0,0 +1,240 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set content = message.content %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in message.content %}\n                {%- set content = message.content.split('</think>')[-1].lstrip('\\n') %}\n                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- endif %}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}