Llama-3 experiments#6
Merged
Merged
Conversation
- experiments/lm: full torchtitan-based training pipeline (MaxPTrainer, maxp_converter, maxp_llama3 scale configs, SLURM launch sweep, debug runner) - Steps auto-computed as 20× non-embed params via meta-device param count - Replace torchtitan submodule with torchtitan>=0.2.2 pip dependency; remove all sys.path hacks from scripts - Fix DTensor/Tensor mix in alignment computation (FSDP2 compat) - Method rename: mp-full→mup-full, mp-noalign→mup-no - docs/experiments.md: consolidated experiment plan - pyproject.toml: add torchtitan, trim unused dev deps Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- README: add installation with cu128/cu130/cpu extras, LLM experiments section - pyproject: cu128/cu130 extras for CUDA index selection, remove torchaudio - torchtitan pinned to nightly 1a2fef04 (2026-04-21) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Switch fineweb loaders to streaming=True for multi-worker support - Use fineweb-edu base dataset across all scales for fair comparison - Make --batch-size global (divided by WORLD_SIZE for local batch size) - Add --num-workers (default 8) and --prefetch-factor (default 4) args - Set OMP_NUM_THREADS=16, HF_DATASETS_OFFLINE=1 in SLURM template - Remove SCALE_DATASETS dict, default dataset to fineweb-edu Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Re-enable async checkpointing via --checkpoint-interval. Resync optimizer param refs by FQN after no_grad capture forwards so FSDP2 reshard-induced Parameter swaps don't break get_optimizer_state_dict. Add use_training_activations flag to MaxPConverter config.
Skip runs with existing checkpoints by default; --resume bypasses the skip to resubmit interrupted training.
Force reshard_after_forward=always to prevent no_grad forwards from leaving layers unsharded; remove now-unnecessary _capture_and_resync. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.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.
No description provided.