Repository navigation
Flow-Factory v0.5.0
This release completes the declared TP/SP/Ulysses-CP paths across 19 registered model adapters and
14 trainers, with model-owned layouts, shared preparation/state ownership and a change-aware
validation harness. It also makes sharded target materialization the qualified FSDP2 default and
reduces host loading memory through checkpoint slices on supported Diffusers/modular loaders.
Highlights
- Compose TP, SP and CP with logical data parallelism on the supported FSDP2 paths. SP is enabled
by default for SD3.5 and Wan T2V under TP; other adapters opt in explicitly. - Keep model-specific packed projections, references, causal prefixes, multimodal streams and
output boundaries in their adapter owners. The model table records parameter and activation
coverage separately. - Use common objective and execution contracts across online, offline and multi-role trainers.
Reward/optimization overlap preserves logical work and state ownership without demanding
deterministic ready-order trajectories. - Prepare qualified targets on meta and materialize only local FSDP2 shards, including TP/SP/CP
compositions. Bagel/SenseNova retain native host loaders; checkpoint slices additionally reduce
host memory for supported Diffusers/modular loaders. - Reuse authenticated, source-equivalent PR evidence after merges while running fresh CPU gates.
Validation follows changed owners and purposeful two/four/eight-GPU configurations.
Upgrade
Python 3.11 or newer and released Diffusers >=0.42.1 are required. The container qualification tuple
uses Python 3.12, Torch 2.10.0, Accelerate 1.15.0 and Diffusers 0.42.1; CP admits the corresponding
2.10.x / 1.15.x / 0.42.x release lines. Flux3 and Bagel no longer need Diffusers Git revision pins.
Optional model kernels and reward dependencies remain separate extras.
Exact training-state continuation requires compatible saved runtime identities and the same
physical topology/placements. Older TP state without placement fingerprints requires weights-only
restart or an explicit migration. Canonical model weights remain separate from optimizer state.
Read installation,
loading and migration, and
model coverage before selecting a recipe.
Scope and evidence
The coverage promise concerns the declared layouts and contracts, not every Cartesian combination
of model, algorithm, backend, precision and topology. DDP/ZeRO-2/FSDP2 retain their documented
backend roles. EP/DeepEP, model-admitted Ring, arbitrary in-flight recovery and topology-changing
optimizer continuation remain future work. Capacity measurements are tied to their recorded model,
input geometry and source; this release makes no universal throughput or convergence claim.
Package/source hashes, installation checks, the resolved container environment, image identity and
smoke evidence are attached to the GitHub Release. Original numerical and lifecycle receipts remain
indexed in the source's release acceptance and model/algorithm handoffs. Merged receipts
retain their original execution commits; a release tag does not relabel them as fresh executions.