Revert "[recipe] feat: enable THD packing by default for Qwen3.5-VL finetune" - #3894
Conversation
…inetune …" This reverts commit ae0ae7e.
Light ReviewClean revert of #3481 — flips pack_sequences_in_batch back to False and removes the associated comment in _qwen35_vl_apply_common. No logic issues. One observation: The revert PR body says only Reverts #3481 with no explanation of why. The original PR documented a hard dependency on Megatron-LM #2645 (GDN packed-sequence support). If the revert is because that MCore change has not landed in the pinned submodule yet, a one-line note in the PR description would help future readers understand the timeline (and know when it is safe to re-enable). Test coverage: The existing unit tests in tests/unit_tests/recipes/qwen_vl/test_qwen35_vl_recipes.py do not assert on pack_sequences_in_batch. This is fine for a revert to the prior default, but means neither the original enablement nor this revert is guarded by a unit test. Consider adding a lightweight assertion (e.g., assert cfg.dataset.pack_sequences_in_batch is False for a representative SFT config) so the next flip is intentional and CI-visible. Suggested test cases: No perf tests impacted. |
…inetune" (NVIDIA-NeMo#3894) Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Reverts #3481