Probably irrelevant, cleanup

This commit is contained in:
kijai
2025-03-08 12:11:04 +02:00
parent 2559003f40
commit 4bc21fdece
+15 -21
View File
@@ -410,28 +410,22 @@ class TextEncoder(nn.Module):
else: else:
last_hidden_state = outputs[self.output_key] last_hidden_state = outputs[self.output_key]
if prompt_template is not None: if prompt_template is not None:
if data_type == 'I2V_image': crop_start = prompt_template.get("crop_start", -1)
crop_start = prompt_template.get("crop_start", -1) text_crop_start = crop_start - 1 + prompt_template.get("image_emb_len", 576)
crop_end = prompt_template.get('assistant_emb_start', -1) image_crop_start = prompt_template.get("image_emb_start", 5)
elif data_type == 'I2V_video': image_crop_end = prompt_template.get('image_emb_end', 581)
crop_start = prompt_template.get("crop_start", -1) batch_indices, last_double_return_token_indices = torch.where(
text_crop_start = crop_start - 1 + prompt_template.get("image_emb_len", 576) batch_encoding["input_ids"] == prompt_template.get('double_return_token_id', 271))
image_crop_start = prompt_template.get("image_emb_start", 5) last_double_return_token_indices = last_double_return_token_indices.reshape(
image_crop_end = prompt_template.get('image_emb_end', 581) batch_encoding["input_ids"].shape[0], -1)[:, -1]
batch_indices, last_double_return_token_indices = torch.where( batch_indices = batch_indices.reshape(batch_encoding["input_ids"].shape[0], -1)[:, -1]
batch_encoding["input_ids"] == prompt_template.get('double_return_token_id', 271)) assistant_crop_start = last_double_return_token_indices - 1 + prompt_template.get(
last_double_return_token_indices = last_double_return_token_indices.reshape( "image_emb_len", 576) - 4
batch_encoding["input_ids"].shape[0], -1)[:, -1] assistant_crop_end = last_double_return_token_indices - 1 + prompt_template.get(
batch_indices = batch_indices.reshape(batch_encoding["input_ids"].shape[0], -1)[:, -1] "image_emb_len", 576)
assistant_crop_start = last_double_return_token_indices - 1 + prompt_template.get(
"image_emb_len", 576) - 4
assistant_crop_end = last_double_return_token_indices - 1 + prompt_template.get(
"image_emb_len", 576)
attention_mask_assistant_crop_start = last_double_return_token_indices - 4 attention_mask_assistant_crop_start = last_double_return_token_indices - 4
attention_mask_assistant_crop_end = last_double_return_token_indices attention_mask_assistant_crop_end = last_double_return_token_indices
else:
raise ValueError(f"Unsupported data type: {data_type}")
text_last_hidden_state = [] text_last_hidden_state = []
text_attention_mask = [] text_attention_mask = []