We now have the initial gpucomm-fs scaffold: a minimal content-addressed store (CAS) CLI (init/put/ls/get) with BLAKE3 hashing + JSON metadata, plus CI.
Goal
Make this repo a binary-aware artifact store + (optional) filesystem layer for GPU-related artifacts (datasets, weights, PTX/CUBIN, benchmark bundles).
Proposed next milestones
- Artifact metadata v0.2
- Define a stable metadata schema (kind, project, cuda, sm arch, commit sha, build flags)
put --meta validation + show <hash>
- Integrity + lifecycle
verify (hash check + meta presence)gc / prune (remove unreferenced objects)
- Sharing / transport
pack/unpack or push/pull to an HTTP/S3 backend
- Filesystem layer (optional)
- FUSE mount (read-only first) mapping virtual paths → objects/meta
Decisions needed
- Do we prioritize remote backend (push/pull) before FUSE?
- What are the canonical artifact
kind values for gpucomm?
Acceptance criteria for next PR
- New
show + verify commands with tests
- CI stays green (
cargo fmt, cargo test)
We now have the initial
gpucomm-fsscaffold: a minimal content-addressed store (CAS) CLI (init/put/ls/get) with BLAKE3 hashing + JSON metadata, plus CI.Goal
Make this repo a binary-aware artifact store + (optional) filesystem layer for GPU-related artifacts (datasets, weights, PTX/CUBIN, benchmark bundles).
Proposed next milestones
put --metavalidation +show <hash>verify(hash check + meta presence)gc/prune(remove unreferenced objects)pack/unpackorpush/pullto an HTTP/S3 backendDecisions needed
kindvalues for gpucomm?Acceptance criteria for next PR
show+verifycommands with testscargo fmt,cargo test)