Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio diffusion: CFG-parallel router keeps the wrapped forward's signature (FLUX.3 under CFG-parallel) (#14918)

Daniel Hanb3004af536a3
diff --git a/studio/backend/core/inference/diffusion_multigpu.py b/studio/backend/core/inference/diffusion_multigpu.py
index 36d0eef7dda..afd9f220d1d 100644
--- a/studio/backend/core/inference/diffusion_multigpu.py
+++ b/studio/backend/core/inference/diffusion_multigpu.py
@@ -2286,6 +2286,12 @@ class CfgParallelRouter:
torch = _torch()
self.module_ref = weakref.ref(module)
self.orig_forward = module.forward
+ # FLUX.3's modular denoise step passes only kwargs named in forward's signature: expose the wrapped one.
+ try:
+ import inspect
+ self.__signature__ = inspect.signature(self.orig_forward)
+ except (TypeError, ValueError):
+ pass
self.inner: Any = None
self._pairing = False
self.had_override = "forward" in module.__dict__
diff --git a/studio/backend/tests/test_diffusion_multigpu.py b/studio/backend/tests/test_diffusion_multigpu.py
index 5ac7c592ef1..6ca47457142 100644
--- a/studio/backend/tests/test_diffusion_multigpu.py
+++ b/studio/backend/tests/test_diffusion_multigpu.py
@@ -2413,6 +2413,46 @@ def test_router_stays_on_the_source_for_a_forced_eager_render_and_the_compile_th
assert torch.equal(out, reference) and router.stats["replica_calls"] == 3
+def test_router_keeps_the_forward_signature_callers_filter_on():
+ """FLUX.3's modular denoise step filters kwargs by forward's signature; a bare router dropped img_ids/txt_ids."""
+ import inspect
+
+ class Dit(torch.nn.Module):
+ def __init__(self):
+ super().__init__()
+ self.lin = torch.nn.Linear(4, 4)
+
+ def forward(
+ self,
+ hidden_states,
+ encoder_hidden_states = None,
+ timestep = None,
+ img_ids = None,
+ txt_ids = None,
+ return_dict = True,
+ ):
+ assert img_ids is not None and txt_ids is not None, "the ids were filtered out"
+ return (self.lin(hidden_states) + img_ids.sum() + txt_ids.sum(),)
+
+ dit = Dit()
+ router = mg.CfgParallelRouter(dit, primary = "cpu", secondary = "cpu")
+ dit.forward = router
+ fields = {"img_ids": torch.zeros(2, 4), "txt_ids": torch.zeros(3, 4), "guidance_hint": 1.0}
+ # the call diffusers' modular_pipelines/flux3/denoise.py makes
+ accepted = set(inspect.signature(dit.forward).parameters)
+ kwargs = {k: v for k, v in fields.items() if k in accepted}
+ assert sorted(kwargs) == ["img_ids", "txt_ids"]
+ x = torch.randn(1, 4)
+ out = dit(
+ hidden_states = x,
+ encoder_hidden_states = None,
+ timestep = torch.ones(1),
+ return_dict = False,
+ **kwargs,
+ )[0]
+ assert torch.equal(out, dit.lin(x))
+
+
def test_replica_blocks_get_their_own_compile_guard():
from core.inference.diffusion_speed import guard_compiled_blocks