Mist: Co-Optimize Memory Optimizations with Parallelism for Large Scale Distributed Training
Bibliographic record
Abstract
Various parallelism such as data, sharded data, tensor, and pipeline parallelism, and memory optimizations such as activation checkpointing and offloading have been proposed to work together to accelerate large scale distributed training. To find the best combination of these techniques, automatic distributed training strategy planners are proposed. However, existing systems cannot comprehensively co-optimize memory optimizations with parallelism, lack advanced overlap awareness, and do not consider inter-microbatch imbalance in pipeline parallelism, leading to sub-optimal performance. To address these shortcomings and the design challenge of the exploded search space, we propose Mist, a memory, overlap, and imbalance-aware automatic distributed training system. Mist comprises three key components: (1) an overlap-centric schedule template that orchestrates techniques in an overlapped manner and mitigates the tuning complexity, (2) an interference-aware symbolic analysis system that provides accurate and efficient predictions of the runtime and memory usage in the form of symbolic expressions, and (3) an imbalance-aware hierarchical auto-tuner that decouples the tuning into an inter-stage imbalance-aware Mixed Integer Linear Programming (MILP) problem and an intra-stage Constrained Optimization problem. Our evaluation results show that Mist achieves an average of 1.28× (up to 1.73×) and 1.27× (up to 2.04×) speedup compared to state-of-the-art manual implementation Megatron-LM and automatic method Aceso, respectively.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".