Always use native create_causal_mask; fix FA2 and save_pretrained on v5

- Drop the cache gate and `_create_causal_mask_fallback`. This file already needs transformers >= 4.54, where `masking_utils.create_causal_mask` exists. The gate pushed every cached call on 5.6-5.16 onto the fallback, which lost the SDPA `is_causal` fast path and the flex BlockMask, and gave FA3/FA4 a 4D mask. The signature differences (`input_embeds` -> `inputs_embeds`, removal of `cache_position`) are now resolved once from `inspect.signature`, and `position_ids` is passed as None because 3D M-RoPE coordinates are not packed-sequence positions.
- Declare `_supports_flash_attn = True` on `PaddleOCRPreTrainedModel`. v5 reads only this name, so `attn_implementation="flash_attention_2"` raised "PaddleOCRVisionModel does not support Flash Attention 2 yet". 4.x accepts both names.
- Use dict-style `_tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}`. v5 calls `.items()` on it, so the list form made `save_pretrained` raise `AttributeError: 'list' object has no attribute 'keys'`. 4.x only iterates it, which yields the same key. `tie_word_embeddings` is false in config.json, so lm_head stays untied.

Same change as PaddlePaddle/PaddleOCR-VL-1.6#15 (commits 91669010 and 0f4ece18); the modeling, configuration and config.json files of this PR are identical to that one before the change.

Tested on a tiny random model with transformers 4.57.6, 5.12.1 and 5.22.1, on CPU fp32 (eager, sdpa) and A100 bf16 (eager, sdpa, flash_attention_2): generation matches the previous PR head, cache == no cache, a left-padded row matches the same row unpadded, packed multi-image vision encoding matches per-image encoding, and save_pretrained + from_pretrained restores the rotary/position buffers, keeps lm_head tying as configured, and reproduces pre-save generation.

Co-Authored-By: Claude Code <noreply@anthropic.com>

modeling_paddleocr_vl.py

@@ -30,6 +30,7 @@ from transformers.cache_utils import (
     Cache,
     DynamicCache,
 )
+from transformers.masking_utils import create_causal_mask as _hf_create_causal_mask
 from transformers.generation import GenerationMixin
 from transformers.integrations import use_kernel_forward_from_hub
 from transformers.modeling_attn_mask_utils import AttentionMaskConverter
@@ -68,24 +69,16 @@ from .configuration_paddleocr_vl import PaddleOCRVisionConfig, PaddleOCRVLConfig
 
 
 # Resolve optional framework interfaces once, outside the forward path.
-try:
-    from transformers.masking_utils import create_causal_mask as _hf_create_causal_mask
-except ImportError:
-    _hf_create_causal_mask = None
-
 try:
     from transformers import initialization as _hf_initialization
 except ImportError:
     _hf_initialization = None
 
-try:
-    _HF_MASK_PARAMETERS = frozenset(inspect.signature(_hf_create_causal_mask).parameters)
-except (TypeError, ValueError):
-    _HF_MASK_PARAMETERS = frozenset()
-
-_HF_MASK_INPUT_NAME = next(
-    (name for name in ("inputs_embeds", "input_embeds") if name in _HF_MASK_PARAMETERS), None
-)
+# create_causal_mask exists in every release this file can import on (>= 4.54),
+# but its signature changed: `input_embeds` -> `inputs_embeds`, and
+# `cache_position` was later removed in favor of reading offsets from the cache.
+_HF_MASK_PARAMETERS = frozenset(inspect.signature(_hf_create_causal_mask).parameters)
+_HF_MASK_INPUT_NAME = "inputs_embeds" if "inputs_embeds" in _HF_MASK_PARAMETERS else "input_embeds"
 
 
 def _guard_initializer(function):
@@ -118,60 +111,18 @@ def _create_causal_mask(config, inputs_embeds, attention_mask, past_key_values,
             return additive.masked_fill(~mask, torch.finfo(inputs_embeds.dtype).min)
         return mask.to(dtype=inputs_embeds.dtype)
 
-    cache_supported = past_key_values is None or (
-        hasattr(past_key_values, "get_mask_sizes")
-        and ("cache_position" in _HF_MASK_PARAMETERS or hasattr(past_key_values, "get_query_offset"))
-    )
-    if _hf_create_causal_mask is not None and _HF_MASK_INPUT_NAME is not None and cache_supported:
-        mask_kwargs = {
-            "config": config,
-            _HF_MASK_INPUT_NAME: inputs_embeds,
-            "attention_mask": attention_mask,
-            "past_key_values": past_key_values,
-        }
-        if "cache_position" in _HF_MASK_PARAMETERS:
-            mask_kwargs["cache_position"] = cache_position
-        if "position_ids" in _HF_MASK_PARAMETERS:
-            # 3D M-RoPE coordinates are not packed text-sequence positions.
-            mask_kwargs["position_ids"] = None
-        return _hf_create_causal_mask(**mask_kwargs)
-
-    return _create_causal_mask_fallback(
-        config, inputs_embeds, attention_mask, past_key_values, cache_position
-    )
-
-
-def _create_causal_mask_fallback(config, inputs_embeds, attention_mask, past_key_values, cache_position):
-    # Keep mask semantics local: Transformers renamed inputs and removed
-    # cache_position across releases. M-RoPE coordinates are handled separately.
-    if attention_mask is not None and attention_mask.ndim == 4:
-        mask = attention_mask.to(device=inputs_embeds.device)
-        if mask.dtype == torch.bool:
-            additive = torch.zeros(mask.shape, device=mask.device, dtype=inputs_embeds.dtype)
-            return additive.masked_fill(~mask, torch.finfo(inputs_embeds.dtype).min)
-        return mask.to(dtype=inputs_embeds.dtype)
-    if config._attn_implementation == "flash_attention_2":
-        return attention_mask if attention_mask is not None and (attention_mask == 0).any() else None
-    batch_size, query_length = inputs_embeds.shape[:2]
-    past_length = past_key_values.get_seq_length() if past_key_values is not None else 0
-    key_length = past_length + query_length
-    if attention_mask is not None:
-        key_length = max(key_length, attention_mask.shape[-1])
-    if past_key_values is not None and getattr(past_key_values, "is_compileable", False):
-        maximum = past_key_values.get_max_cache_shape()
-        if isinstance(maximum, int):
-            key_length = max(key_length, maximum)
-    query_positions = cache_position[-query_length:].to(inputs_embeds.device)
-    key_positions = torch.arange(key_length, device=inputs_embeds.device)
-    allowed = key_positions[None, :] <= query_positions[:, None]
-    allowed = allowed[None, None, :, :].expand(batch_size, 1, query_length, key_length)
-    if attention_mask is not None:
-        padding = attention_mask[:, :key_length].to(device=inputs_embeds.device, dtype=torch.bool)
-        if padding.shape[-1] < key_length:
-            padding = F.pad(padding, (0, key_length - padding.shape[-1]), value=False)
-        allowed = allowed & padding[:, None, None, :]
-    mask = torch.zeros(allowed.shape, dtype=inputs_embeds.dtype, device=inputs_embeds.device)
-    return mask.masked_fill(~allowed, torch.finfo(inputs_embeds.dtype).min)
+    mask_kwargs = {
+        "config": config,
+        _HF_MASK_INPUT_NAME: inputs_embeds,
+        "attention_mask": attention_mask,
+        "past_key_values": past_key_values,
+    }
+    if "cache_position" in _HF_MASK_PARAMETERS:
+        mask_kwargs["cache_position"] = cache_position
+    if "position_ids" in _HF_MASK_PARAMETERS:
+        # 3D M-RoPE coordinates are not packed text-sequence positions.
+        mask_kwargs["position_ids"] = None
+    return _hf_create_causal_mask(**mask_kwargs)
 
 
 def _init_derived_buffers(module):
@@ -779,7 +730,7 @@ class Ernie4_5Model(Ernie4_5PreTrainedModel):
 
 
 class Ernie4_5ForCausalLM(Ernie4_5PreTrainedModel, GenerationMixin):
-    _tied_weights_keys = ["lm_head.weight"]
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
     _tp_plan = {"lm_head": "colwise_rep"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
@@ -1429,7 +1380,8 @@ class PaddleOCRPreTrainedModel(PreTrainedModel):
         "PaddleOCRVisionEmbeddings",
         "PaddleOCRMultiheadAttentionPoolingHead",
     ]
-    _supports_flash_attn_2 = True
+    # v5 only reads `_supports_flash_attn`; 4.x accepts both names.
+    _supports_flash_attn = True
     _supports_sdpa = True
 
     def _init_weights(self, module):
@@ -2033,7 +1985,7 @@ class PaddleOCRVLCausalLMOutputWithPast(ModelOutput):
 
 
 class PaddleOCRVLForConditionalGeneration(Ernie4_5PreTrainedModel, GenerationMixin):
-    _tied_weights_keys = ["lm_head.weight"]
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
     config_class = PaddleOCRVLConfig
     _no_split_modules = ["Ernie4_5_DecoderLayer", "PaddleOCREncoderLayer"]