Skip to content

brownkat6/reasoning-scheduling

Repository files navigation

Predictive Scheduling for Efficient Inference-Time Reasoning in Large Language Models

Brief current file directory

gen_datasets.py - pulls data from huggingface and saves it in the Dynasor data directory. Run once and never again for each of the datasets we want to add.

mlp_datagen.py - generates X, y for our prediction task. E.g. y is the early stopping probabilities of generating the correct answer if we terminate after W, 2W, .... tokens. X is the predictors we want to use, e.g. the hidden states and potentially other data. This script generates X and y for given datasets and saves it to netscratch.

generate_data.sh - sbatch script that batches out the data generation jobs with mlp_datagen.py`

mlp_train.py loads in data, trains in MLP, and saves the MLP

mlp.py - defines the MLP class

Dynasor/benchmark/TokenDeprivation/run_adaptive.py - script that loads in MLP predictions and/or oracle data and records the tokens per accuracy given adaptive greedy strategy of allocating more size/reasoning budget to queries that will benefit from it more

run.sh - runs benchmark tokens vs accuracy experiment for non-adaptive allocation strategy

run_adaptive.sh - runs benchmark tokens vs accuracy experiment for adaptive allocation strategy using MLP predictions

run_adaptive_oracle.sh - runs benchmark tokens vs accuracy experiment for adaptive allocation strategy using oracle ground truth for early stopping correctness probabilities (using the same data used to train the MLP)

vis_adaptive.py - creates visualization for tokens vs accuracy experiment for non-adaptive vs adaptive vs oracle

NOTE: most of dynasor vllm server scheduling layer is in Dynasor.dynasor.cli/*. Read and understand this. We will likely want to adapt/reuse large portions of this code later to write our own scheduling layer.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published