Use the generate config sampler instead of the undefined train_config in GenerateProcess

[rajakshay]

What

In GenerateProcess.__init__ (jobs/process/GenerateProcess.py), the sampler-building branch references self.train_config.noise_scheduler, but GenerateProcess never sets self.train_config (only training processes do).

Why it breaks

Running a generate job that reaches this branch crashes at construction:

AttributeError: 'GenerateProcess' object has no attribute 'train_config'

Fix

Use self.generate_config.sampler, which is already built earlier in the same __init__. GenerateConfig defines sampler (default 'ddpm'), which is exactly what get_sampler(...) expects.

Note on the previous attempt (#498)

This revives the fix reported in #416 and attempted in #498 (closed for inactivity). It intentionally uses generate_config.sampler rather than generate_config.noise_scheduler (as in #498): GenerateConfig has a sampler attribute but no noise_scheduler attribute, so .sampler actually resolves the crash instead of moving it to a different AttributeError.

Regression safety

This branch currently always crashes, so no working path is affected.

Validation

• Before: a FLUX generate job crashes at init with the AttributeError above.
• After: the generate job constructs, loads the model, and produces images (validated with a FLUX generate run that produced a full batch).

Refs #416. Supersedes the approach in #498.

[rajakshay]

Reproduction (no models or data, CPU-only)

get_model_class is stubbed so no weights load and so the model class has no get_train_scheduler (which forces the sampler else-branch containing the bug). The crash happens before any model would be loaded. Run from the repo root.

repro.py

import importlib
from collections import OrderedDict

# import_module returns the real submodule even though jobs.process re-exports the class
GP = importlib.import_module("jobs.process.GenerateProcess")


class DummyModel:  # no get_train_scheduler -> forces the sampler else-branch
    def __init__(self, *a, **k):
        pass


GP.get_model_class = lambda cfg: DummyModel  # no weights loaded


class FakeJob:
    name = "repro"
    meta = OrderedDict()
    device = "cpu"


config = OrderedDict({
    "output_folder": "/tmp/repro_out",
    "model": OrderedDict({"name_or_path": "stub", "is_flux": True}),
    "generate": OrderedDict({"prompts": ["a photo"], "sampler": "ddpm"}),
    "dtype": "float16",
})

try:
    GP.GenerateProcess(0, FakeJob(), config)
    print("PASSED: GenerateProcess constructed; sampler built from generate_config.sampler")
except AttributeError as e:
    print("REPRODUCED AttributeError:", str(e))
except Exception as e:
    print("OTHER:", type(e).__name__, "-", str(e).splitlines()[0])

Result:

# BEFORE (current main)
REPRODUCED AttributeError: 'GenerateProcess' object has no attribute 'train_config'

# AFTER (this PR)
PASSED: GenerateProcess constructed; sampler built from generate_config.sampler