Fix _init_rope compatibility with transformers >= 5.x rope_scaling standardization

## Problem

Starting from transformers 4.43+, `AutoConfig` automatically standardizes a missing or `None` `rope_scaling` to `{"rope_type": "default", "factor": 1.0}` during config initialization. MiniCPM-SALA's `_init_rope()` only handles `None`, `"linear"`, `"dynamic"`, and `"longrope"` — encountering `"default"` causes:

```
ValueError: Unknown RoPE scaling type default
```

This is triggered when loading MiniCPM-SALA via any library that depends on transformers >= 5.x (e.g. GPTQModel 7.0.0 pulls in transformers 5.8.0), even though the model does not use RoPE scaling (`config.json` has no `rope_scaling` field).

## Changes

Modified both `_init_rope()` methods in `modeling_minicpm_sala.py` (`MiniCPMInfLLMv2Attention` at line 879 and `MiniCPMSALAMLA` at line 2144):

- Treat `rope_scaling` of `None` or `scaling_type in (None, "default")` as no-scaling, falling through to the standard `MiniCPMRotaryEmbedding`.
- Support both legacy `"type"` key and new `"rope_type"` key for cross-version compatibility.
- Extract `rope_scaling` into a local variable to avoid repeated `self.config.rope_scaling` access.

## Impact

- **No behavioral change**: The model's `config.json` has no `rope_scaling` field; this fix only ensures the auto-filled `"default"` is correctly treated as no scaling.
- **Backward compatible**: Works with both transformers 4.x (`rope_scaling=None`) and 5.x (`rope_scaling={"rope_type": "default", ...}`).
- **Verified**: Tested with GPTQModel 7.0.0 + transformers 5.8.0 in the SOAR evaluation Docker image. Without this fix, `GPTQModel.load()` raises `ValueError`; with this fix, model loads successfully.

modeling_minicpm_sala.py

@@ -877,15 +877,23 @@ class MiniCPMAttention(nn.Module):
             )
 
     def _init_rope(self):
-        if self.config.rope_scaling is None:
+        # transformers>=4.43 standardizes rope_scaling: a missing/None
+        # rope_scaling is auto-filled to {"rope_type": "default", "factor": 1.0}
+        # at config-load time. Treat both the original None case and the
+        # standardized "default" as no scaling so loading does not raise on
+        # newer transformers releases.
+        rope_scaling = self.config.rope_scaling
+        scaling_type = None
+        if isinstance(rope_scaling, dict):
+            scaling_type = rope_scaling.get("type") or rope_scaling.get("rope_type")
+        if rope_scaling is None or scaling_type in (None, "default"):
             self.rotary_emb = MiniCPMRotaryEmbedding(
                 self.head_dim,
                 max_position_embeddings=self.max_position_embeddings,
                 base=self.rope_theta,
             )
         else:
-            scaling_type = self.config.rope_scaling["rope_type"]
-            scaling_factor = self.config.rope_scaling.get("factor", None)
+            scaling_factor = rope_scaling.get("factor", None)
             if scaling_type == "linear":
                 self.rotary_emb = MiniCPMLinearScalingRotaryEmbedding(
                     self.head_dim,
@@ -904,10 +912,10 @@ class MiniCPMAttention(nn.Module):
                 self.rotary_emb = MiniCPMLongRoPE(
                     self.head_dim,
                     max_position_embeddings=self.max_position_embeddings,
-                    short_factor=self.config.rope_scaling["short_factor"],
-                    long_factor=self.config.rope_scaling["long_factor"],
+                    short_factor=rope_scaling["short_factor"],
+                    long_factor=rope_scaling["long_factor"],
                     base=self.rope_theta,
-                    original_max_position_embeddings=self.config.rope_scaling[
+                    original_max_position_embeddings=rope_scaling[
                         "original_max_position_embeddings"
                     ],
                 )
@@ -2142,15 +2150,23 @@ class LightningAttention(nn.Module):
         self._init_rope()
 
     def _init_rope(self):
-        if self.config.rope_scaling is None:
+        # transformers>=4.43 standardizes rope_scaling: a missing/None
+        # rope_scaling is auto-filled to {"rope_type": "default", "factor": 1.0}
+        # at config-load time. Treat both the original None case and the
+        # standardized "default" as no scaling so loading does not raise on
+        # newer transformers releases.
+        rope_scaling = self.config.rope_scaling
+        scaling_type = None
+        if isinstance(rope_scaling, dict):
+            scaling_type = rope_scaling.get("type") or rope_scaling.get("rope_type")
+        if rope_scaling is None or scaling_type in (None, "default"):
             self.rotary_emb = MiniCPMRotaryEmbedding(
                 self.head_dim,
                 max_position_embeddings=self.config.max_position_embeddings,
                 base=self.config.rope_theta,
             )
         else:
-            scaling_type = self.config.rope_scaling["rope_type"]
-            scaling_factor = self.config.rope_scaling.get("factor", None)
+            scaling_factor = rope_scaling.get("factor", None)
             if scaling_type == "linear":
                 self.rotary_emb = MiniCPMLinearScalingRotaryEmbedding(
                     self.head_dim,
@@ -2169,10 +2185,10 @@ class LightningAttention(nn.Module):
                 self.rotary_emb = MiniCPMLongRoPE(
                     self.head_dim,
                     max_position_embeddings=self.config.max_position_embeddings,
-                    short_factor=self.config.rope_scaling["short_factor"],
-                    long_factor=self.config.rope_scaling["long_factor"],
+                    short_factor=rope_scaling["short_factor"],
+                    long_factor=rope_scaling["long_factor"],
                     base=self.config.rope_theta,
-                    original_max_position_embeddings=self.config.rope_scaling[
+                    original_max_position_embeddings=rope_scaling[
                         "original_max_position_embeddings"
                     ],
                 )
@@ -3274,4 +3290,4 @@ class MiniCPMSALAForSequenceClassification(MiniCPMSALAPreTrainedModel):
             past_key_values=transformer_outputs.past_key_values,
             hidden_states=transformer_outputs.hidden_states,
             attentions=transformer_outputs.attentions,
-        )
+        )