llmflux 2.0.0

LLMFlux v2.0.0

on Python PyPI

This is a major release and contains breaking changes — see CHANGELOG.md for the full list with rationale.

The headline feature is multi-node vLLM inference (#137, #141): --nodes N now shards a model across all N nodes instead of starting a server on the first and leaving the rest idle, using vLLM's own rendezvous under a single SPMD srun step — no Ray, no container change. See docs/MULTINODE.md.

Breaking changes

• Config.__init__ gained workspace as its first positional parameter, shifting the directory arguments. Pass them as keyword arguments.
• Batch and serve jobs now exit with the processor's / server's status instead of always 0.
• Five public Config members were removed, along with SlurmConfig.mem and the unprefixed DATA_INPUT_DIR / DATA_OUTPUT_DIR environment variables.
• llmflux connect now validates the node and port it reads from connection.json, and a directory passed to --input now raises.

The CLI surface is unchanged — same subcommands and flags as v1.0.0 — so for command-line users this bump is about behaviour, not interface.

Notable fixes

• llmflux run works again where the container directory is admin-owned and read-only (#138).
• llmflux benchmark no longer downloads its dataset into the installed package (#146).
• llmflux serve --engine ollama now exits with the server's status and publishes one resolved endpoint host, matching the vLLM path.

What's Changed

• remove mem env var by @minump in #119
• remove hardcoded config workspace path and take from env by @minump in #121
• flashinfer cache in hostvars by @minump in #125
• security: validate node/port from connection.json to prevent SSRF (#109) by @joshfactorial in #118
• fix: find images in a directory in vision_to_jsonl by @joshfactorial in #133
• change: report dropped images and raise max_image_size to 25MB by @joshfactorial in #135
• docs: remove duplicated changelog entry, restore lost PR reference by @joshfactorial in #136
• feat: send API key as bearer token from LLMClient by @joshfactorial in #130
• Pin the generated batch stage to the submitting interpreter by @dadamsncsa in #143
• feat: multi-node vLLM inference by @dadamsncsa in #141
• Release 2.0.0 by @dadamsncsa in #145

Full Changelog: v1.0.0...v2.0.0