training-data-foundation / cards /gpu-kernel-trees.md
Battam's picture
Release release-2026-11-17.3, made by the table store from release-2026-11-17.2 (part 4)
eb68e83 verified
|
Raw History Blame Contribute Delete
3.57 kB

GitHub source trees of open-source GPU kernel languages and kernel libraries (Triton, FlashAttention, Composable Kernel, TileLang and CCCL) (gpu-kernel-trees)

Kind: corpus (raw). One whole upstream that a collector keeps reading in bulk, counted in rows: each row points to one file or record.

Every file of the source trees of open-source GPU kernel languages and kernel libraries on GitHub (Triton, FlashAttention, Composable Kernel, TileLang and CCCL (Thrust, CUB and libcu++)), each at the head commit of its default branch when it was read. One pointer per file, to its raw.githubusercontent.com address pinned to that commit, with size and git blob hash from the git tree. The content is not copied and stays on GitHub.

Summary from sample files: not judged yet. The text above is the source's own description.

Domain and form

  • Domain: D4.3 (GPU kernels and operators).
  • Form: source-code. Source code: repositories, programs with sample inputs and outputs, programming exercises, small example projects and proof libraries.

Unit and scale

  • One row points at: file.
  • Rows: 28,667.
  • Size: Recorded size 458 MB; deduplicated size 458 MB; size check: 100.0% match (20 of the 20 rows read compared).
  • Fetch check: all 20 sampled rows read fetched and matched their recorded size and hash (read 2026-11-14).
  • Batches: 2 (all, inc). Collection: complete.

File types and sizes

Type Rows Size
cpp 8,406 56.0 MB
h 4,442 45.5 MB
py 3,598 67.1 MB
cu 3,196 20.4 MB
hpp 2,464 51.0 MB
cuh 1,532 35.3 MB
txt 882 2.56 MB
rst 656 5.85 MB
md 442 3.29 MB
mlir 412 13.4 MB
other types: 75

Size of one row, upper bounds by quantile: 10% 1.02 KB, 50% 4.09 KB, 90% 32.8 KB, 99% 131 KB.

License and access

  • License: Per-repository open source licenses; license tier nc-research; training use: yes; redistribution: yes; files may be shipped: not judged yet.
  • License statement: not recorded on the card.
  • Login: no.
  • How to fetch: not recorded on the card; the row's uri is fetched as it is.

Freshness

  • Rows are counted for the release of 2026-11-17; the card's own sample and counts date from the card's writing (it has no snapshot date of its own).

Readiness

  • Readiness level: not judged yet.
  • Card review: unknown (a draft card has not been reviewed as a whole).
  • Personal information read from samples: not judged yet.

Query this source at a fixed release

Pinned to the release tag release-2026-11-17.3.

SET VARIABLE files = (SELECT list('hf://datasets/Battam/training-data-foundation@release-2026-11-17.3/' || f) FROM (SELECT unnest(data_files) AS f FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-17.3/meta/catalog.parquet' WHERE source = 'gpu-kernel-trees'));
SELECT uri, ext, size_bytes, lower(hex(sha256)) AS sha256
FROM read_parquet(getvariable('files'))
WHERE source = 'gpu-kernel-trees'
LIMIT 10;

Sample files

No sample files are recorded on the card yet.

Parts not to use for tasks

  • Parts to leave out: not judged yet.
  • Benchmark overlap none: every data file is compared with the benchmark list, and the default profile of tools/fetch.py removes none of the rows.

This source has no card (cards/<source>.yaml) yet; values shown here as not judged yet have no judgment so far.