V0.2.1 small patch release for infmax v3 support

[RogerLiu312]

sflow v0.2.1 Release Notes

Release date: April 2026
Previous release: v0.2.0 (March 2026)

---

Highlights

sflow v0.2.1 is a documentation and workflow polish release for the InfMax v3 migration path. It documents the branch behavior for CSV-driven execution, self-contained YAML batch submission, replica variable domains, node placement, and probe orchestration.

---

User-Facing Changes

CLI and Batch Workflows

• sflow run --bulk-input now has documented single-row CSV execution. Use --row with exactly one selector to run a specific CSV row.
• Advanced --row selectors are documented for run, compose, and batch: repeated flags, comma lists, Python-style slices with exclusive end, open-ended slices, and negative indices such as --row=-1.
• sflow batch --bulk-submit is documented for submitting self-contained YAML files, folders, or glob patterns without CSV merging.
• Auto-derived node counts are documented. Single-job and bulk-submit batch modes can derive --nodes from the Slurm backend; bulk-input mode requires either --nodes or a CSV node-count column.
• --sflow-version is documented for pinning the git ref installed by generated sbatch scripts.
• Expression-aware --sbatch-extra-args is documented. Extra sbatch directives can resolve ${{ variables.X }} or shorthand ${{ X }} from config defaults, CLI --set, and CSV row values.

Variables and Replica Sweeps

• Variable domain metadata is documented through ${{ variables.NAME.domain }}.
• Replica sweep behavior is clarified: ${{ variables.NAME }} resolves to the per-replica value, while ${{ variables.NAME.domain }} remains the full domain list.
• Domain overrides via --set are documented: JSON-style list values update the variable domain, and the variable value becomes the first list item.

Resources and Placement

• resources.nodes.exclude is documented for removing nodes from the placement pool before applying indices, count, or GPU packing.
• Negative node indices are clarified, including the fact that negative indices are resolved after exclude filtering.
• Default Slurm placement is documented: when a task does not set resources.nodes, sflow passes the full backend allocation to srun.
• GPU packing behavior is documented, including multi-node expansion when a GPU request is an exact multiple of gpus_per_node.

Probes

• Probe timing defaults are documented, including timeout: 1200 for readiness probes and each_check_timeout: 30.
• HTTP probes (http_get and http_post) are documented with examples.
• Multiple readiness probes are documented as AND semantics: all readiness probes must trigger before a task becomes ready.
• Failure probes are documented as fail-fast signals that mark tasks as failed by probe and cancel downstream work.
• Replica HTTP probe deduplication is documented for parallel replicas with identical HTTP probes.

---

Documentation Updated

• docs/user/cli.md
• docs/user/variables.md
• docs/user/resources.md
• docs/user/probes.md
• docs/user/quick-reference.md
• docs/user/configuration.md
• docs/user/architecture.md