flashhub-kernels

Load CUDA kernel packages published on the flashhub CDN (https://flashhub-cdn.aodianyun.com by default).

Kernel packages on the flashhub CDN are made to be:

• Portable: a kernel can be loaded from a local path or downloaded from the CDN into a cache, without being installed into PYTHONPATH.
• Variant-matched: each package publishes one build per environment (torch version, CUDA version, CPU architecture); the client picks the variant that fits the local runtime and downloads only that build.
• Verified: every artifact is checked against the sha256 digests in its metadata.json before loading.

🚀 Quick Start

Install the flashhub-kernels package (requires torch>=2.0):

pip install flashhub-kernels

Load a kernel from the flashhub CDN:

import torch

from flashhub_kernels import get_kernel

kernel = get_kernel("sdq/audio-codebook-primitives", version=1)

logits = torch.randn((4, 16), dtype=torch.bfloat16, device="cuda")
codebook = torch.randn((4, 16, 8), dtype=torch.bfloat16, device="cuda")
codes, emb = kernel.delayed_codebook_argmax_embed_bf16(
    logits, codebook, delay=2, boc=1
)

API

• get_kernel(repo_id, revision=None, version=None) — download from the CDN into the local cache (variant-matched to torch/CUDA/arch), import and return the kernel module. version selects v<version>; revision accepts "v1"/"1". Exactly one of revision/version must be given.
• get_local_kernel(repo_path) — import a kernel from a local repo dir.
• has_kernel(repo_id, version=None) — whether the variant exists on the CDN.
• get_kernel_variants(repo_id, version=None) — resolve every published variant against the current system (VariantAccepted/VariantRejected).
• get_loaded_kernels() — snapshot of kernels loaded in this process.
• configure(cdn_url=None, prefix=None, cache_dir=None) — override CDN settings for the process (defaults read from FLASHHUB_CDN_URL, FLASHHUB_CDN_PREFIX, FLASHHUB_CACHE_DIR).

CDN layout

repo_id is the full <org>/<pkg> (e.g. sdq/audio-codebook-primitives); the org is part of the object path.

<prefix>/<org>/<pkg>/v<version>/build/<variant>/
    metadata.json          # {"digest": {"files": {rel: b64-sha256}}}
    <py_pkg>/__init__.py
    <py_pkg>/_ops.py
    <py_pkg>/_<pkg>_cuda_<id>.so
<prefix>/<org>/<pkg>/v<version>/variants.json   # optional: published variant names

variant is derived from the runtime: torch<major><minor>-cxx11-cu<cuda>-<arch>-linux.