Summary

unsloth_compile_transformers() now compiles every transformers model_type that has a modeling_<x>.py file, on both transformers 4.57.6 and 5.x:

Production models (llama, qwen3, gemma3, mistral, mixtral, idefics{,2,3}, gpt_oss, deepseek_v3, ...) keep working unchanged on both versions.

Models unblocked

Nine narrowly-scoped fixes

1. create_new_function: empty new_source[0] IndexError. Guard with if new_source and new_source[0] == " ".

2. convert_attention_masks_to_bool: nested-paren regex mis-parse. The findall regex doesn't handle nested parens like return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min). Skip when the return body contains (.

3. eval(f"{loc}.{module}") AttributeError on dir()-derived names. Wrap 6 eval() sites with try/except AttributeError: continue (private names in __all__ like _BaseModelWithGenerate, or nn.Linear aliases that fail attribute lookup on some transformers versions).

4. Idefics-style loss = loss_fct(...) rewrite: substring + dedent double-corruption. The flat forward.replace(...) matched inside lm_loss = loss_fct(...) and inserted unindented lines that the downstream dedent then chopped 4 chars off. Replace with re.sub anchored to ^([ \t]+)loss = ...$ that captures and inherits the leading whitespace.

5. called_functions sig-rebuild fallback dropped trailing :. When inspect.signature repr disagrees with source quote style or fully-qualified annotation, the fallback consumed the colon. Restore : and \n so the body lands as a proper indented suite.

6. Decorator prepended without dedenting indented source. textwrap.dedent before adding the @torch.compile(...) decorator (xLSTM soft_cap is defined under an else: so inspect.getsource returns it indented 4 chars).

7. Topological ordering by full_source.find(name) used the wrong match. Bare-name find hits forward refs in docstrings / type annotations / Union[..., NAME] before the actual class NAME definition. Switch to find("class N(") / find("class N:") / find("def N(").

8. @auto_docstring (and 4 sibling decorators) used [^)]* to match the argument list, which stops at the first ) inside a string like "... (a log mel spectrogram), meaning ..." and leaves an unterminated string literal in the cache. Switch to a regex recursive group that respects nested parens AND nested string literals (single and triple-quoted).

9. Base-class import missing for classes used purely as bases. Sam3LiteTextLayerScaledResidual has only an inherited forward so it's filtered out of the per-module compile and out of functions, but class Sam3LiteTextRepMixer(Sam3LiteTextLayerScaledResidual) references it. Scan class N(B1, B2, ...) patterns in the emitted source and add bases that exist on the modeling file but aren't defined-in-cache to the auto-import list.

All fixes either widen try/except, anchor a substitution to respect indentation, switch to a strictly-more-correct sort key, or detect-and-import a previously-missed reference. None of them change semantics for working models.

Test plan

• Production-relevant models (llama, qwen3, qwen3_moe, qwen2, gemma{,2,3,3n}, mistral, mixtral, phi{,3,4}, starcoder2, codegen, bart, t5, whisper, idefics{,2,3}, vit, clip, siglip, dinov2, wav2vec2, llava{,_next,_onevision}, qwen2_vl, qwen2_5_vl, qwen2_5_omni, gpt_oss, deepseek_v3, dbrx, gpt_neo{,x}, olmo{,2}) -> 39/39 ok on both tf 4.57.6 and tf 5.8.0.
• Idefics family (the original target of the loss=loss_fct rewrite) -> 3/3 ok; synthetic equivalence test confirms the new re.sub produces a properly-indented expansion vs the old .replace()'s broken zero-indent emit.
• convert_attention_masks_to_bool -> simple-return rewrite still fires (Case A), tuple-of-bare-names still rewrites (Case C), nested-paren returns are correctly skipped (Case B).
• Full sweep on transformers 4.57.6: 359/359 with modeling files compile, 0 broken.
• Full sweep on transformers 5.8.0: 439/439 with modeling files compile, 0 broken.