fix: work around transformers lazy_load_kernel offline regression

[hallerite]

Summary

• Work around a regression in transformers >= 5.5 that breaks Mamba-based models (NemotronH, Zamba2, Jamba, etc.) when HF_HUB_OFFLINE=1 is set — as our multi-node SLURM template does
• lazy_load_kernel() resolves kernel versions via HfApi().list_repo_refs(), which raises OfflineModeIsEnabled even when kernels are cached locally
• The compat patch catches hub errors and loads kernels from the local HF Hub cache ($HF_HOME/hub/models--kernels-community--*/)

The regression was introduced across three transformers PRs:

• PR #41577 (Oct 2025): lazy_load_kernel introduced
• PR #43955 (Feb 2026): version=1 integer versioning bypasses snapshot_download cache fallback
• PR #44176 (Feb 2026): all Mamba models switched from direct imports to lazy_load_kernel

No upstream issue exists yet — will file one on huggingface/transformers.

🤖 Generated with Claude Code

---

Note

Medium Risk
Monkeypatches transformers runtime behavior in a model-kernel loading path; if upstream behavior changes, this could mask real hub-kernel issues or alter which kernels get used in offline/online environments.

Overview
Extends prime_rl._compat from a single ring-flash-attn workaround into a small set of documented early-import compatibility shims.

Adds a new patch for transformers.integrations.hub_kernels.lazy_load_kernel on transformers >= 5.5 to gracefully handle HF_HUB_OFFLINE=1 by catching OfflineModeIsEnabled and returning None (so Mamba-based models can skip hub kernels and proceed with the project’s direct kernel path).

^{Reviewed by Cursor Bugbot for commit d3d91a6. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a125615. Configure here.}

[cursor]

Narrow except ImportError won't catch AttributeError from missing attributes

Medium Severity

The outer except ImportError on line 110 only catches ImportError, but lines 46–47 access _hub_kernels.lazy_load_kernel and _hub_kernels._KERNEL_MODULE_MAPPING, which raise AttributeError if those names don't exist. Since transformers.integrations.hub_kernels existed before lazy_load_kernel was added, there are valid transformers versions where the module imports successfully but the attribute access crashes the app at startup. Catching (ImportError, AttributeError) would match the graceful-degradation intent.

Additional Locations (1)

• src/prime_rl/_compat.py#L45-L47

 

^{Reviewed by Cursor Bugbot for commit a125615. Configure here.}

[hallerite]

that's the point, though