[WIP] Add gemma4 drafter model support#2240
Draft
athitten wants to merge 2 commits into
Draft
Conversation
Signed-off-by: Abhishree <abhishreetm@gmail.com>
athitten
changed the title
Apply fixes for joint base + drafter training: * Drop ``use_cache=False`` override in ``composite.forward``. Without the ``DynamicCache``, HF's sliding-window mask path silently degrades (SDPA mask-skip can collapse sliding layers into plain causal attention), inflating the initial training loss. The YAML's ``text_config.use_cache: true`` now takes effect. * Change drafter label shift from ``k + 1`` to ``k``. The VLM collate pre-shifts labels by 1 so ``labels[t] == input_ids[t + 1]``; the prior ``k + 1`` shift was training the drafter to predict ``input_ids[t + 2]`` instead of ``input_ids[t + 1]``. * Add hard asserts: ``cp_size == 1`` and ``torch_dtype == bfloat16`` in ``Gemma4WithDrafter.from_pretrained``. * Add plan knobs: ``freeze_base_for_drafter``, ``share_embedding_with_base`` (one-shot init copy; FSDP2-safe), ``base_activation_checkpointing``. * Recipe: factor joint loss into ``FinetuneRecipeForVLM._maybe_add_drafter_loss``, gate log on ``is_remote_logging_step`` (was per-microbatch), and make validation drafter-aware so ``val_loss`` reflects drafter drift. * Remove dead ``from_pretrained`` override in drafter wrapper. * Drop redundant ``text_config.output_hidden_states`` from YAML; expand the ``use_cache: true`` comment to explain the real reason (sliding-window mask, not KV sharing). * Add ``test_post_collate_semantic_alignment`` that pins the label-shift convention so a future regression to ``k + 1`` fails loudly. Refine ``test_drafter_loss_reaches_drafter_params`` to reflect that ``post_projection`` only sees gradient in multi-step chains. Signed-off-by: Abhishree <abhishreetm@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do ?
Add a one line overview of what this PR aims to accomplish.
Changelog
Before your PR is "Ready for review"
Pre checks:
If you haven't finished some of the above items you can still open "Draft" PR.
Additional Information