tckm.yan/aha
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update interpolate

Browse Source
This commit is contained in:
jhqxxx
2026-03-01 12:44:14 +08:00
parent 6a73ee04c3
commit a61b899741
15 changed files with 598 additions and 486 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/paddleocr_vl/model.rs
+4 -3
View File
@@ -16,9 +16,10 @@ use crate::{
},
},
position_embed::rope::{Qwen2_5VLTextRotaryEmbedding, Qwen2_5VisionRotaryEmbedding},
utils::interpolate::interpolate_bilinear,
utils::tensor_utils::{
get_vision_next_indices, interpolate_bilinear, masked_scatter_dim0, nonzero_index,
prepare_causal_attention_mask, zero_index,
get_vision_next_indices, masked_scatter_dim0, nonzero_index, prepare_causal_attention_mask,
zero_index,
},
};
@@ -143,7 +144,7 @@ impl SiglipVisionEmbeddings {
.reshape((1, sqrt_num_positions, sqrt_num_positions, self.embed_dim))?
.permute((0, 3, 1, 2))?;
let patch_pos_embed =
interpolate_bilinear(&patch_pos_embed, (new_height, new_width), Some(false))?;
interpolate_bilinear(&patch_pos_embed, (new_height, new_width), Some(false), None)?;
let patch_pos_embed =
patch_pos_embed
.permute((0, 2, 3, 1))?