chore: clean up pre-5.3 transformers remnants

- input_embeds= → inputs_embeds= in create_causal_mask call (deprecated in 5.6)
- Remove LossKwargs try/except shim, use TransformersKwargs directly
- Simplify tokenizer import to top-level `from transformers import PreTrainedTokenizer`
- torch_dtype= → dtype= in inference/generate.py
- Delete commented-out dead imports in imageprocessor files
- Shorten outdated v4.54 deprecation warning in vision attention

imageprocessor_openpangu_vl.py

@@ -37,10 +37,6 @@ from transformers.image_utils import (
 from torchvision.transforms.v2 import functional as F
 import torch
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
-# from transformers.image_processing_utils_fast import (
-#     group_images_by_shape,
-#     reorder_images,
-# )
 
 
 def rescale(image, scale):

inference/generate.py

@@ -17,7 +17,7 @@ key_mapping = {
 model =  AutoModelForCausalLM.from_pretrained(
         model_path,
         trust_remote_code=True,
-        torch_dtype='auto',
+        dtype='auto',
         key_mapping=key_mapping).eval().cuda()
 
 conversation = [

inference/requirements.txt

@@ -1,4 +1,4 @@
-torch==2.5.1
-torch_npu==2.5.1
-transformers==4.53.2
+torch==2.7.1
+torch_npu==2.7.1
+transformers==5.3.0
 qwen_vl_utils==0.0.14

inference/vllm_ascend/pangu_infer/models/vllm_ascend/imageprocessor_openpangu_vl.py

@@ -37,10 +37,6 @@ from transformers.image_utils import (
 from torchvision.transforms.v2 import functional as F
 import torch
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
-# from transformers.image_processing_utils_fast import (
-#     group_images_by_shape,
-#     reorder_images,
-# )
 
 
 def rescale(image, scale):

modeling_openpangu_embedded.py

@@ -56,12 +56,7 @@ from transformers.processing_utils import Unpack
 from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
 from transformers.configuration_utils import PretrainedConfig
 
-try:
-    from transformers.utils import LossKwargs
-except ImportError:
-    # Transformers 5.3 exposes the generic kwargs TypedDict but no longer exports
-    # the loss-specific alias used by upstream OpenPangu code.
-    LossKwargs = TransformersKwargs
+LossKwargs = TransformersKwargs
 
 
 def _compute_default_rope_parameters(config, device=None, seq_len=None, layer_type=None):
@@ -599,7 +594,7 @@ class PanguEmbeddedModel(PanguEmbeddedPreTrainedModel):
 
         causal_mask = create_causal_mask(
             config=self.config,
-            input_embeds=inputs_embeds,
+            inputs_embeds=inputs_embeds,
             attention_mask=attention_mask,
             cache_position=cache_position,
             past_key_values=past_key_values

modeling_openpangu_vl.py

@@ -42,12 +42,7 @@ from transformers.processing_utils import Unpack
 from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
 from transformers.utils.import_utils import is_torchdynamo_compiling
 
-try:
-    from transformers.utils import LossKwargs
-except ImportError:
-    # Transformers 5.3 exposes the generic kwargs TypedDict but no longer exports
-    # the loss-specific alias used by upstream OpenPangu code.
-    LossKwargs = TransformersKwargs
+LossKwargs = TransformersKwargs
 
 
 def _compute_default_rope_parameters(config, device=None, seq_len=None, layer_type=None):
@@ -245,10 +240,7 @@ class OpenPanguVLVisionAttention(nn.Module):
         )
         if position_embeddings is None:
             logger.warning_once(
-                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
-                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
-                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
-                "removed and `position_embeddings` will be mandatory."
+                "`rotary_pos_emb` is deprecated; pass `position_embeddings` (cos, sin) instead."
             )
             emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
             cos = emb.cos()

tokenization_openpangu.py

@@ -26,10 +26,7 @@ from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
-try:
-    from transformers.tokenization_python import PreTrainedTokenizer  # type: ignore[reportMissingImports]
-except ImportError:
-    from transformers.tokenization_utils import PreTrainedTokenizer  # type: ignore[reportMissingImports]
+from transformers import PreTrainedTokenizer
 from transformers.utils import logging