modeling_ax_ve.py

# Copyright 2026 A.X K2 team and HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from typing import Callable, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F

from transformers.activations import ACT2FN
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.utils import auto_docstring
from .configuration_ax_ve import AXVEConfig, AXVEVisionConfig

NORM2FN = {
    'layer_norm': nn.LayerNorm,
}

def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb_vision(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
    orig_q_dtype = q.dtype
    orig_k_dtype = k.dtype
    q, k = q.float(), k.float()
    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    q_embed = q_embed.to(orig_q_dtype)
    k_embed = k_embed.to(orig_k_dtype)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class AXVEPatchEmbed(nn.Module):
    def __init__(
        self,
        config,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size = config.patch_size
        self.in_channels = in_channels = 3
        self.embed_dim = embed_dim = config.hidden_size

        kernel_size = [patch_size, patch_size]
        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=True)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states

class AXVEVisionRotaryEmbedding(nn.Module):
    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs.float()


class AXVEVisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.linear_fc1 = nn.Linear(self.hidden_size, self.intermediate_size, bias=True)
        self.linear_fc2 = nn.Linear(self.intermediate_size, self.hidden_size, bias=True)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_state):
        return self.linear_fc2(self.act_fn(self.linear_fc1(hidden_state)))


class AXVEVisionAttention(nn.Module):
    def __init__(self, config: AXVEVisionConfig) -> None:
        super().__init__()
        self.dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.dim // self.num_heads
        self.num_key_value_groups = 1  # needed for eager attention
        self.q_proj = nn.Linear(self.dim, self.dim, bias=True)
        self.k_proj = nn.Linear(self.dim, self.dim, bias=True)
        self.v_proj = nn.Linear(self.dim, self.dim, bias=True)
        self.o_proj = nn.Linear(self.dim, self.dim)
        self.scaling = self.head_dim**-0.5
        self.config = config
        self.attention_dropout = 0.0
        self.is_causal = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple] = None,
        **kwargs,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        query_states = self.q_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        key_states = self.k_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        value_states = self.v_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)

        query_states = query_states.transpose(0, 1).unsqueeze(0)
        key_states = key_states.transpose(0, 1).unsqueeze(0)
        value_states = value_states.transpose(0, 1).unsqueeze(0)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        if self.config._attn_implementation == "flash_attention_2":
            # Flash Attention: Use cu_seqlens for variable length attention
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
            attn_output, _ = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask=None,
                scaling=self.scaling,
                dropout=0.0 if not self.training else self.attention_dropout,
                cu_seq_lens_q=cu_seqlens,
                cu_seq_lens_k=cu_seqlens,
                max_length_q=max_seqlen,
                max_length_k=max_seqlen,
                is_causal=False,
                **kwargs,
            )
        else:
            # Other implementations: Process each chunk separately
            lengths = cu_seqlens[1:] - cu_seqlens[:-1]
            splits = [
                torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
            ]

            attn_outputs = [
                attention_interface(
                    self,
                    q,
                    k,
                    v,
                    attention_mask=None,
                    scaling=self.scaling,
                    dropout=0.0 if not self.training else self.attention_dropout,
                    is_causal=False,
                    **kwargs,
                )[0]
                for q, k, v in zip(*splits)
            ]
            attn_output = torch.cat(attn_outputs, dim=1)

        attn_output = attn_output.reshape(seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output


class AXVEVisionBlock(GradientCheckpointingLayer):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = nn.LayerNorm(config.hidden_size, eps=1e-6)
        self.norm2 = nn.LayerNorm(config.hidden_size, eps=1e-6)
        self.attn = AXVEVisionAttention(config=config)
        self.mlp = AXVEVisionMLP(config=config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple] = None,
        **kwargs,
    ) -> torch.Tensor:
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states),
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


@auto_docstring
class AXVEPreTrainedModel(PreTrainedModel):
    config: AXVEConfig
    base_model_prefix = "model"
    input_modalities = ("image",)
    supports_gradient_checkpointing = True
    _no_split_modules = ["AXVEVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": AXVEVisionBlock,
        "attentions": AXVEVisionAttention,
    }


class AXVEVisionModel(AXVEPreTrainedModel):
    config: AXVEVisionConfig
    _no_split_modules = ["AXVEVisionBlock"]
    input_modalities = ("image",)
    _can_record_outputs = {
        "hidden_states": AXVEVisionBlock,
        "attentions": AXVEVisionAttention,
    }

    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)
        self.embed_dim = config.hidden_size
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = config.patch_size
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size
        

        self.patch_embed = AXVEPatchEmbed(
            config=config,
        )

        self.pos_embed = nn.Embedding(config.num_position_embeddings, config.hidden_size)
        self.num_grid_per_side = int(config.num_position_embeddings**0.5)

        head_dim = config.hidden_size // config.num_attention_heads
        self.rotary_pos_emb = AXVEVisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList([AXVEVisionBlock(config) for _ in range(config.num_hidden_layers)])                
        self.final_layernorm = NORM2FN['layer_norm'](self.embed_dim, eps=config.layer_norm_eps)

        self.residual_visual_indexes = config.residual_visual_indexes
        self.gradient_checkpointing = False

        self.post_init()

    def rot_pos_emb(self, grid_hw: torch.Tensor) -> torch.Tensor:
        merge_size = self.spatial_merge_size
        grid_hw_list = grid_hw.tolist()

        max_hw = max(max(h, w) for h, w in grid_hw_list)
        freq_table = self.rotary_pos_emb(max_hw)  # (max_hw, dim // 2)
        device = freq_table.device

        total_tokens = sum(h * w for h, w in grid_hw_list)
        pos_ids = torch.empty((total_tokens, 2), dtype=torch.long, device=device)

        offset = 0
        for height, width in grid_hw_list:
            merged_h, merged_w = height // merge_size, width // merge_size

            block_rows = torch.arange(merged_h, device=device)  # block row indices
            block_cols = torch.arange(merged_w, device=device)  # block col indices
            intra_row = torch.arange(merge_size, device=device)  # intra-block row offsets
            intra_col = torch.arange(merge_size, device=device)  # intra-block col offsets

            # Compute full-resolution positions
            row_idx = block_rows[:, None, None, None] * merge_size + intra_row[None, None, :, None]
            col_idx = block_cols[None, :, None, None] * merge_size + intra_col[None, None, None, :]

            row_idx = row_idx.expand(merged_h, merged_w, merge_size, merge_size).reshape(-1)
            col_idx = col_idx.expand(merged_h, merged_w, merge_size, merge_size).reshape(-1)

            coords = torch.stack((row_idx, col_idx), dim=-1)


            num_tokens = coords.shape[0]
            pos_ids[offset : offset + num_tokens] = coords
            offset += num_tokens

        embeddings = freq_table[pos_ids]  # lookup rotary embeddings
        embeddings = embeddings.flatten(1)
        return embeddings

    def fast_pos_embed_interpolate(self, grid_hw):
        grid_hw_list = grid_hw.tolist()
        grid_hs = [row[0] for row in grid_hw_list]
        grid_ws = [row[1] for row in grid_hw_list]
        device = self.pos_embed.weight.device

        idx_list = [[] for _ in range(4)]
        weight_list = [[] for _ in range(4)]

        for h, w in grid_hw_list:
            h_idxs = torch.linspace(0, self.num_grid_per_side - 1, h)
            w_idxs = torch.linspace(0, self.num_grid_per_side - 1, w)

            h_idxs_floor = h_idxs.int()
            w_idxs_floor = w_idxs.int()
            h_idxs_ceil = (h_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)
            w_idxs_ceil = (w_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)

            dh = h_idxs - h_idxs_floor
            dw = w_idxs - w_idxs_floor

            base_h = h_idxs_floor * self.num_grid_per_side
            base_h_ceil = h_idxs_ceil * self.num_grid_per_side

            indices = [
                (base_h[None].T + w_idxs_floor[None]).flatten(),
                (base_h[None].T + w_idxs_ceil[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_floor[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_ceil[None]).flatten(),
            ]

            weights = [
                ((1 - dh)[None].T * (1 - dw)[None]).flatten(),
                ((1 - dh)[None].T * dw[None]).flatten(),
                (dh[None].T * (1 - dw)[None]).flatten(),
                (dh[None].T * dw[None]).flatten(),
            ]

            for i in range(4):
                idx_list[i].extend(indices[i].tolist())
                weight_list[i].extend(weights[i].tolist())

        idx_tensor = torch.tensor(idx_list, dtype=torch.long, device=device)
        weight_tensor = torch.tensor(weight_list, dtype=self.pos_embed.weight.dtype, device=device)
        pos_embeds = self.pos_embed(idx_tensor).to(device) * weight_tensor[:, :, None]
        patch_pos_embeds = pos_embeds[0] + pos_embeds[1] + pos_embeds[2] + pos_embeds[3]

        patch_pos_embeds = patch_pos_embeds.split([h * w for h, w in zip(grid_hs, grid_ws)])

        patch_pos_embeds_permute = []
        merge_size = self.config.spatial_merge_size
        for pos_embed, h, w in zip(patch_pos_embeds, grid_hs, grid_ws):
            pos_embed = (
                pos_embed.view(1, h // merge_size, merge_size, w // merge_size, merge_size, -1)
                .permute(0, 1, 3, 2, 4, 5)
                .flatten(0, 4)
            )
            patch_pos_embeds_permute.append(pos_embed)
        patch_pos_embeds = torch.cat(patch_pos_embeds_permute)
        return patch_pos_embeds

    def forward(
        self, hidden_states: torch.Tensor, grid_hw: torch.Tensor, **kwargs
    ) -> tuple:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
                The final hidden states of the model.
            grid_hw (`torch.Tensor` of shape `(num_images, 2)`):
                The temporal, height and width of feature shape of each image in LLM.

        Returns:
            `torch.Tensor`: hidden_states.
        """
        hidden_states = self.patch_embed(hidden_states)

        pos_embeds = self.fast_pos_embed_interpolate(grid_hw)
        hidden_states = hidden_states + pos_embeds

        rotary_pos_emb = self.rot_pos_emb(grid_hw)

        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = (grid_hw[:, 0] * grid_hw[:, 1]).cumsum(
            dim=0,
            # FA2: cu_seqlens must be int32 / onnx.export: same dtype as grid_hw
            dtype=grid_hw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

        residual_feature_lists = []
        for layer_num, blk in enumerate(self.blocks):
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens,
                position_embeddings=position_embeddings,
                **kwargs,
            )
            if layer_num in self.residual_visual_indexes:
                residual_feature = hidden_states
                residual_feature_lists.append(residual_feature)

        ### add residual feature lists
        for residual_feature in residual_feature_lists:
            hidden_states += residual_feature

        hidden_states = self.final_layernorm(hidden_states)

        return hidden_states

__all__ = [
    "AXVEVisionModel",
    "AXVEPreTrainedModel",
]