Fix meta-device crash on newer transformers; document required transformers==4.46.3

modeling_marketing_classifier.py: build the encoder via AutoModel.from_config() instead of AutoModel.from_pretrained() inside __init__. The nested from_pretrained call conflicted with the meta-device/low_cpu_mem_usage context that the outer PreTrainedModel.from_pretrained() wraps __init__ in on transformers>=5, causing 'You are using from_pretrained with a meta device context manager' errors (reported from a fresh Colab install). from_config is meta-safe and behaves identically here since the real encoder weights are always supplied by this class's own checkpoint, not downloaded separately -- verified byte-identical output vs the original MarketingClassifier class before and after this change.

Separately, transformers>=5 is not usable with this model regardless of the above fix: the frozen encoder's own custom code (Snowflake/snowflake-arctic-embed-m-v2.0's modeling_hf_alibaba_nlp_gte.py, not ours) hits an IndexError in its RoPE position_ids handling under transformers 5.x. README now states the transformers==4.46.3 requirement explicitly in How to Use.

README.md

@@ -144,6 +144,8 @@ Only BMse is recommended for production. Others are retained for reproducibility
 
 ## How to Use
 
+All usage paths below require `transformers` in the 4.46.x line (the frozen encoder's own custom code is not compatible with `transformers` 5.x): `pip install "transformers==4.46.3"`.
+
 ### Pipeline (recommended)
 
 ```python

modeling_marketing_classifier.py

@@ -23,9 +23,15 @@ class MarketingClassifierForRegression(PreTrainedModel):
 
         encoder_config = AutoConfig.from_pretrained(config.backbone, trust_remote_code=True)
         encoder_config.use_memory_efficient_attention = False
-        self.encoder = AutoModel.from_pretrained(
-            config.backbone,
-            config=encoder_config,
+        # Build the encoder architecture only (random weights, meta-device safe) --
+        # do NOT call AutoModel.from_pretrained() here. This class's own weights
+        # (loaded by the outer from_pretrained() call, which already contains the
+        # full encoder + head state dict) are what actually populate `self.encoder`.
+        # A nested from_pretrained() call here conflicts with the meta-device /
+        # low_cpu_mem_usage context that from_pretrained wraps __init__ in on
+        # newer transformers versions.
+        self.encoder = AutoModel.from_config(
+            encoder_config,
             add_pooling_layer=False,
             trust_remote_code=True,
             attn_implementation="sdpa",