| """Vision patch embedding for Qwen VL models. |
| |
| Flattens 3D video/image patches via Conv3d weight reshaped into a linear projection. |
| |
| Unified across Qwen2-VL and Qwen3-VL: |
| - bias: Qwen2-VL uses bias=False, Qwen3-VL uses bias=True. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import torch |
| import torch.nn as nn |
|
|
| from ..L1.conv3d import Conv3d |
| from ..L1.linear import Matmul |
|
|
|
|
| class VisionPatchEmbed(nn.Module): |
| def __init__(self, patch_size: int, temporal_patch_size: int, |
| in_channels: int, embed_dim: int, bias: bool = False): |
| super().__init__() |
| self.embed_dim = embed_dim |
| self.input_size = in_channels * temporal_patch_size * patch_size * patch_size |
| kernel = (temporal_patch_size, patch_size, patch_size) |
| self.proj = Conv3d(in_channels, embed_dim, kernel, bias=bias) |
| self.linear = Matmul() |
|
|
| def forward(self, x: torch.Tensor) -> torch.Tensor: |
| x = x.view(x.shape[0], self.input_size) |
| return self.linear( |
| x, |
| self.proj.weight.view(self.embed_dim, self.input_size), |
| self.proj.bias, |
| ) |
|
|
