Hardware-aware pareto-optimal search for compressed transformers
Bibliographic record
Abstract
BERT-based transformers are a type of large-scale deep neural networks (DNNs) that have been widely adopted to solve natural language understanding (NLU) tasks.Because of their huge architecture, it is a difficult task to deploy them on some resourceconstrained devices, including laptops and embedded systems.To solve this problem, recent research have focused on model compression to reduce model size without large drop in performance.Moreover, researchers also use hardware performance metrics, such as latency and throughput, to measure model inference performance on target devices and include the hardware feedback in the design phase of model compression techniques.In this thesis, we perform hardware-aware pareto-optimal search on compressed BERTbased transformers for different devices.First, we explore the relationship between the depth of BERT-based models and their model inference performance on CPU and GPU.We generate sub-networks with different depth using BERT base and DistilBERT.We measure their latency and throughput during model evaluation and analyze the results.Then, we extend our experiment to models with both adaptive depth and width and search for pareto-optimal model architecture on target devices.We define the search space to be task-specific sub-networks with different depth and width produced by DynaBERT.We evaluate the sub-networks' performance on the GLUE benchmark and measure their inference latency on CPUs and GPUs of three devices.We set the evaluation performance and inference latency as our search objectives.We search for pareto-optimal sub-networks for each GLUE task.In the end, we compare the results on the CPU and GPU of the same device and between different devices.We discover that there exists a linear relationship between inference latency and model depth.The throughput decreases as latency and model depth increases.We also notice i
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 | 0.001 | 0.004 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.004 | 0.000 |
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".