[original4422]

Summary

 if trust_remote_code is None:
-    trust_remote_code = model_config.trust_remote_code
+    trust_remote_code = False

Require explicit opt-in for custom code in FlagAutoModel.from_finetuned, including Hub namespaces, local paths, and checkpoint-parent aliases that match a registry basename. Warn when a recognized model is configured for custom code, and document the compatibility change. Explicit True/False and model-class, pooling, and query-format selection are preserved.

Fixes #1579.

Evidence

• Before: new tests against the original public dispatcher: 12 failed, 15 passed; failures captured omitted trust becoming True at the selected model constructor.
• After: 32 passed (27 new dispatcher cases and 5 existing import tests), covering all seven current True registry entries, path aliases, explicit flags, and selection/forwarding controls.
• Tests intercept only the downstream model constructor. The dispatcher and registry use the original package; zero outbound network attempts, model loads, or remote model-code executions.
• Scoped Python compilation, package wheel build, and git diff --check pass. Reused the original declared editable-install dependencies with Python 3.13.14, Torch 2.15.0, and Transformers 5.22.1.
• Full model/GPU/provider integration tests and the full Sphinx site were not run.

Merge Danger

Door: two-way

Models requiring custom code now need trust_remote_code=True explicitly, after reviewing that code. This is an intentional compatibility change.

Blast Radius: embedder-loading

Prepared with AI assistance.