apunkt/doc-to-lora
1
0
Fork 0
mirror of https://github.com/SakanaAI/doc-to-lora.git synced 2026-07-26 17:11:02 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

position_ids fix

Browse source
This commit is contained in:
51616 2025-07-03 11:56:03 +00:00
parent fdeaf47353
commit b64abaceda
1 changed files with 7 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

13
src/ctx_to_lora/modeling/idefics2.py
View file

@ -779,7 +779,7 @@ class Idefics2PerceiverResampler(Idefics2PreTrainedModel):
) )
new_position_ids[new_cur_idx : new_cur_idx + n_new_ids] = new_ids new_position_ids[new_cur_idx : new_cur_idx + n_new_ids] = new_ids
cur_idx = idx cur_idx = idx
new_cur_idx = cur_idx + self.n_latents * (i + 1) new_cur_idx = cur_idx + self.n_latents * (bsz - 1)
new_position_ids[new_cur_idx:] = last_pos_ids new_position_ids[new_cur_idx:] = last_pos_ids
position_ids = new_position_ids position_ids = new_position_ids
@ -836,13 +836,14 @@ class Idefics2PerceiverResampler(Idefics2PreTrainedModel):
compressed_context = layer_outputs[0] compressed_context = layer_outputs[0]
position_ids = torch.arange( if self.n_self_attn_layers:
self.n_latents, device=position_ids.device, dtype=torch.int32 position_ids = torch.arange(
).repeat(1, bsz) self.n_latents, device=position_ids.device, dtype=torch.int32
).repeat(1, bsz)
for processor in self.layers[self.n_cross_attn_layers :]:
cu_seq_lens_k = cu_seq_lens_q cu_seq_lens_k = cu_seq_lens_q
max_length_k = max_length_q max_length_k = max_length_q
for processor in self.layers[self.n_cross_attn_layers :]:
# print(f"Using self-attention for resampler") # print(f"Using self-attention for resampler")
# print(f"updated position_ids: {position_ids}") # print(f"updated position_ids: {position_ids}")
# print(f"updated cu_seq_lens_k: {cu_seq_lens_k}") # print(f"updated cu_seq_lens_k: {cu_seq_lens_k}")