Skip to content

Latest commit

 

History

History
178 lines (122 loc) · 7.49 KB

README.md

File metadata and controls

178 lines (122 loc) · 7.49 KB
Raw

Repo Level Order Ray Transform

Please see the set of transform project conventions for details on general project conventions, transform configuration, testing and IDE set up.

Summary

This transform does repository level packing of data and arranging them to prioritise semantic dependancies. This was done to prepare long context data for Scaling Granite Code Models to 128K Context . Quoting the paper.

To create long-context data, we develop a new approach that packs files from the same repository together, arranging them to prioritize semantic dependencies. We identify these dependencies by analyzing file imports and create a directed acyclic graph, where each file is a node and edges represent API imports between files. After breaking any cycles in the graph, we perform a topological sort to establish an ordering of files based on their semantic dependencies. We then organize the files in a repository by placing documentation and build files first, followed by the ordered set of files with semantic dependencies, and finally the remaining non-connected files. These non-connected files are arranged according to their folder structure, using a depth-first search to traverse the repository. Finally, we determine the dominant programming language of a repository based on file extensions and presence of build files, to organise repo-ordered files by programming languages

This transform can group the data by repo_name and apply additional transformations like( sorting or output_by_language or combining rows) on the grouped data. This transform requires the input data to have at least the following columns:

  • repo name: Name of the repo, it is used for grouping in this transform.

  • title : Which is usually file path.

  • language: Programming language of content

The input data for this transform should be in parquet format. The input data is expected to have code data arranged in rows such that each row represents a file. The required columns in the input data shoud correspond to a) repository name b) file path c) content. This transform supports searching the repo across the whole dataset and writing the files of a repo as a single parquet file.

The transform gives the option to write the repo to file in the following ways.

a) sort the repo content by file path and write a parquet with multiple rows

b) sort the repo content by file path and write a parquet with a single combined row.

c) sort the repo content in semantic ordering and write the parquet with multiple rows.

d) sort the repo content in semantic ordering and write the parquet with a single combined row.

Additionally this transform can grooup the repos in the folders named after the most dominant language in the repo. For more information on this transform, please refer to here.

Configuration and command line Options

This transform is dependant on ray runtime.

Transform Configuration.

  • For output: either the output is directly a file or if dominant language flag is enabled, it should output it in folder of that langauge.
  • Enable sorting: if sorting is enabled, we should be able to choose one of the available ones (SORT_BY_PATH, SORT_SEMANTIC, SORT_SEMANTIC_NORMALISED)
    • SORT_BY_PATH: Normal ascending sort by filename
    • SORT_SEMANTIC: Uses semantic analysis to sort the files within a repo.
    • SORT_SEMANTIC_NORMALISED: Normalises the title to remove https:// prefix, if present, and then runs SORT_SEMANTIC
  • Enable superrows: of writing superrows is enabled, then it combines rows of a repo to a single row otherwise writes normal.

Limitation of transform. This expects a flat input folder structure with parquet files.

Running

Launched Command Line Options

Ray runtime options of ray launcher are available.

Running the samples

To run the samples, use the following make targets

  • run-local-sample - runs src/repo_level_order_local_ray.py

These targets will activate the virtual environment and set up any configuration needed. Use the -n option of make to see the detail of what is done to run the sample.

For example,

make run-local-sample
...

Then

ls output

To see results of the transform.

Using the CLI

  1. Simplest Usage:

Running on local computer setup with data available on local filesystem. Sorting the data at repo level using default algorithm. (SORT_BY_PATH). In this configuration, the workers run locally, store is local. This is a recommended method if number of available cpus is less.

export INPUT_FOLDER="input_data_folder/"
export OUTPUT_FOLDER="output_data_folder/"


local_conf="{'input_folder' : '$INPUT_FOLDER', 'output_folder' : '$OUTPUT_FOLDER'  }"
rm -rf /tmp/mystore # remove if it exists

python src/repo_level_order_transform_ray.py \
       --run_locally True \
       --data_local_config "$local_conf" \
       --repo_lvl_store_type  local  \
       --repo_lvl_store_backend_dir '/tmp/mystore' \
       --repo_lvl_sorting_enabled true,
       --repo_lvl_sorting_algo SORT_SEMANTIC

NOTE: Make sure you use an empty folder as repo_lvl_store_backend_dir.

  1. Running on Cluster

This configuration can be used when data is on S3 like COS storage and computation is done on a ray cluster.

Recommended repo_lvl_store_type for cluster is 'ray'. We need to dedicate some ray actors specially for ray store

for example, if we have a ray cluster with 100 actors available. We can dedicate some for the store. Let us randomly use 35 for computation and 10 for store and let remaining free.

NOTE: The transform reads parquet files in two stages and each stage has its own actor pool for computations. The actor pool to read parquet data in the first stage is managed by the library and the actor pool used in the second stage is managed by the ray runtime of the transform. The number of actors in each pool is same and is configured using --runtime_num_workers. So we have to carefully choose number of actors required based on the number resources available. The number of workers/actors should be less than 35% of total resources available.

We need to add the following cli args:

--runtime_num_workers 35 --repo_lvl_store_type "ray" --repo_lvl_store_ray_cpus 0.2 --repo_lvl_store_ray_nworkers 10

When we want the output with the following configuration enabled:

NOTE: store_backend=s3/local are persistent stores and retain the mappings stores in them, they need to be cleaned/deleted after use. It is recommented to use a different location for store if data is different. They are added to aid in large data processing in multiple stages. store_backend=ray, is not presistent and ephimeral and does not retain any data. It is the simplest to use if resources are available.

export INPUT_FOLDER="input_cos_bucket"
export OUTPUT_FOLDER="output_cos_bucket"

s3_kreds="{ ... }" # in the deafult way used for all transforms. 
s3_conf="{'input_folder' : '$INPUT_FOLDER', 'output_folder' : '$OUTPUT_FOLDER'  }"

python src/repo_level_order_transform_ray.py \
       --run_locally True \
       --data_s3_cred "$s3_kreds" \
       --data_s3_config "$s3_conf" \
       --repo_lvl_store_type  ray  \
       --repo_lvl_combine_rows True\
       --repo_lvl_sorting_enabled True\
       --repo_lvl_store_ray_cpus 0.2 \
       --repo_lvl_store_ray_nworkers 1 \
       --repo_lvl_sorting_algo SORT_SEMANTIC \
       --repo_lvl_output_by_langs True