[bugfix]dynemb export supports input_tile=3

[eric-gecheng]

No description provided.

[github-actions]

Consider reusing is_input_tile_emb() from tzrec/acc/utils.py (already used in tzrec/utils/export_util.py) instead of re-deriving the INPUT_TILE=3 check inline. Keeps the env-var convention (input_tile[0] == "3") in one place and avoids drift if it changes.

[github-actions]

Minor defensive nit: if a single dotted entry ever contained more than one segment in _INPUT_TILE_USER_SEGMENTS, this loop would create N separate symlinks (each renaming a different segment). In practice the construction in EmbeddingGroupImpl / SequenceEmbeddingGroupImpl produces only one such segment per path, so this is theoretical — but adding a break after a successful (or skipped-because-exists) os.symlink would make the intent explicit and match the construction-side rename.

[github-actions]

Code review summary

Targeted bugfix for the INPUT_TILE=3 export path: the checkpoint only stores non-_user dynamicemb dirs, but the exported model expects *_user twins, so loading raised "can't find path to load". Symlinking the user-side paths on rank 0 (with a barrier) is a clean, minimal fix.

Noteworthy:

• Reuse is_input_tile_emb() instead of inlining the env check (see inline comment).
• Inner-loop break would make the single-segment intent explicit (see inline comment).
• Test coverage gap: _link_dynamicemb_input_tile_user_paths is purely filesystem ops and trivially unit-testable, but no unit test exists. The existing restore_model tests in tzrec/utils/checkpoint_util_test.py don't exercise the has_dynamicemb branch at all, and the INPUT_TILE=3 integration tests in tzrec/tests/rank_integration_test.py never combine with dynamicemb. A small unit test covering: matching segment → symlink created, non-matching entries skipped, idempotency under existing symlink, and OSError swallow path would lock in the contract.

Verified accurate:

• Code references in the docstring/comments match EmbeddingGroupImpl.ebc_user/mc_ebc_user and SequenceEmbeddingGroupImpl.ec_dict_user/mc_ec_dict_user.
• No user-facing docs claim "dynemb does not support INPUT_TILE=3", so no doc updates needed.
• Rank-0 + dist.barrier() pattern is consistent with existing checkpoint-restore code; performance impact is negligible (cold-start, O(few dozen entries)).