feat(flux): core runtime + Megatron adapter scaffolding

[jasainio]

Base of an 18-PR series splitting the Flux diffusion-training feature (training Flux, a DiT text-to-image diffusion model, on Primus/Megatron) out of one large branch for reviewability. Targets feat/flux/ci-env and auto-retargets to main once that merges. Every other content PR stacks on this one.

What this changes

The shared runtime scaffolding the rest of the feature builds on: core runtime state + train-runtime wiring, the Megatron adapter and base/pretrain trainers, and a patch auto-loader (patches/__init__.py) that imports every *_patches.py in the package on import, so each later layer only drops in its own patch file with no registry edit. Also carries the shared test root (tests/conftest.py, tests/utils.py) and a one-line .gitignore change (the bare data ignore → root-anchored /data/*, and nothing else) that stops Git from ignoring the in-repo data/ source and config directories the later layers add. Kept deliberately small so it can land first.

Dependencies

Sequenced after the CI-pins PR (feat/flux/ci-env); no functional dependency on the turbo bump.

Test plan

pytest tests/unit_tests/core tests/unit_tests/backends/megatron; lint/pre-commit clean. Validated locally on an AMD GPU container: 58 passed.

Files

39 (core runtime, Megatron adapter/trainers, base patch loader, shared test root, .gitignore).

[luiza-amd]

Last 2 commits pushed resolve an issue in some existing pretrainers which feat/flux/core introduced.

On this branch, train_runtime.py was changed to instantiate every trainer with a much richer argument set:

        trainer = trainer_cls(
            backend_args=backend_args,
            module_name=self.ctx.module_name,
            primus_config=self.ctx.primus_config,
            module_rank=self.ctx.rank,
            module_world_size=self.ctx.world_size,
            module_master_addr=self.ctx.master_addr,
            module_master_port=self.ctx.master_port,
        )

Previously this was just trainer = TrainerClass(backend_args=backend_args).

The Megatron trainers survive this because MegatronBaseTrainer has no __init__, so it inherits BaseTrainer.__init__(self, backend_args=None, *args, **kwargs), which cooperatively absorbs/filters the extra kwargs.

• primus/backends/maxtext/maxtext_pretrain_trainer.py — the one causing the run-unittest-jax CI failure
• primus/backends/torchtitan/torchtitan_pretrain_trainer.py
• primus/backends/hummingbirdxt/hummingbirdxt_posttrain_trainer.py
• primus/backends/megatron/megatron_sft_trainer.py
• primus/backends/megatron_bridge/megatron_bridge_base_trainer.py
• primus/backends/megatron_bridge/megatron_bridge_pretrain_trainer.py
• primus/backends/megatron_bridge/megatron_bridge_posttrain_trainer.py

Each of these trainers had a rigid def __init__(self, backend_args: Any) signature that would raise TypeError under the new train_runtime.py call convention (which now passes module_name, primus_config, module_rank, module_world_size, module_master_addr, module_master_port).

Fix applied

Widen MaxTextPretrainTrainer.__init__ (and the other non-Megatron trainers) to accept and forward the extra kwargs, letting BaseTrainer filter them:

def __init__(self, backend_args: Any = None, **kwargs):
    super().__init__(backend_args=backend_args, **kwargs)

This is safe: for trainers that are not BaseModule subclasses, BaseTrainer.__init__ takes the else: super().__init__() branch and discards the extra kwargs; for BaseModule subclasses it forwards them via super().__init__(**kwargs).