review: rename FLUX2_DUAL_GPU env var to DIFFSYNTH_DUAL_GPU

Per @gemini-code-assist review on modelscope#1436: model-neutral env var name so
a single signal controls dual-GPU mode for both FLUX.2 and Wan paths.
This change:

- FLUX2_DUAL_GPU            -> DIFFSYNTH_DUAL_GPU
- FLUX2_DUAL_GPU_SPLIT_AT   -> DIFFSYNTH_DUAL_GPU_SPLIT_AT

Function names (enable_flux2_dual_gpu) and the helper module name
(flux2_dual_gpu_diffsynth.py) keep the model-specific naming since
they're the entry points users explicitly import for FLUX.2 training.

Coordinated with PR modelscope#1436 which adopts the same DIFFSYNTH_DUAL_GPU
gate on the Wan side. Once both merge, one env var enables dual-GPU
for either model family; runner.py recognizes the gate regardless of
which family the training script targets.

Author: genno-whittlery

diffsynth/diffusion/runner.py

@@ -29,8 +29,8 @@ def launch_training_task(
     dataloader = torch.utils.data.DataLoader(dataset, shuffle=True, collate_fn=lambda x: x[0], num_workers=num_workers)
     # Dual-GPU model-parallel: skip the device move that would undo our
     # manual split, and tell accelerate not to touch the model's device.
-    _flux2_dual_gpu = os.environ.get("FLUX2_DUAL_GPU", "false").lower() == "true"
-    if not _flux2_dual_gpu:
+    _diffsynth_dual_gpu = os.environ.get("DIFFSYNTH_DUAL_GPU", "false").lower() == "true"
+    if not _diffsynth_dual_gpu:
         model.to(device=accelerator.device)
         model, optimizer, dataloader, scheduler = accelerator.prepare(model, optimizer, dataloader, scheduler)
     else:

examples/flux2/model_training/flux2_dual_gpu_diffsynth.py

@@ -26,8 +26,8 @@
   point and at ``norm_out`` (the boundary back to cuda:0).
 
 Env vars:
-    FLUX2_DUAL_GPU=true                    enable dual-GPU path
-    FLUX2_DUAL_GPU_SPLIT_AT=24             override split index
+    DIFFSYNTH_DUAL_GPU=true                    enable dual-GPU path
+    DIFFSYNTH_DUAL_GPU_SPLIT_AT=24             override split index
                                            (default: num_single // 2)
 
 Usage in DiffSynth-Studio's training script
@@ -38,7 +38,7 @@
     # ...build pipeline / training module normally...
     training_module = Flux2ImageTrainingModule(...)
 
-    # Activate the split (env-gated; no-op when FLUX2_DUAL_GPU is unset).
+    # Activate the split (env-gated; no-op when DIFFSYNTH_DUAL_GPU is unset).
     # Call AFTER LoRA injection (switch_pipe_to_training_mode) so PEFT
     # has wrapped target modules. The split places the wrapped modules
     # and PEFT LoRA params follow the base layer's device automatically.
@@ -63,13 +63,13 @@
 # ─── Env-gated public surface ───────────────────────────────────────────────
 
 def is_dual_gpu_enabled() -> bool:
-    """True iff ``FLUX2_DUAL_GPU=true`` in the environment."""
-    return os.getenv("FLUX2_DUAL_GPU", "false").lower() == "true"
+    """True iff ``DIFFSYNTH_DUAL_GPU=true`` in the environment."""
+    return os.getenv("DIFFSYNTH_DUAL_GPU", "false").lower() == "true"
 
 
 def get_split_at(num_single_blocks: int) -> int:
-    """Single-blocks split index. Override via ``FLUX2_DUAL_GPU_SPLIT_AT``."""
-    override = os.getenv("FLUX2_DUAL_GPU_SPLIT_AT")
+    """Single-blocks split index. Override via ``DIFFSYNTH_DUAL_GPU_SPLIT_AT``."""
+    override = os.getenv("DIFFSYNTH_DUAL_GPU_SPLIT_AT")
     if override is not None:
         return int(override)
     return num_single_blocks // 2
@@ -78,7 +78,7 @@ def get_split_at(num_single_blocks: int) -> int:
 def enable_flux2_dual_gpu(dit: nn.Module) -> nn.Module:
     """Distribute the DiffSynth Flux2DiT across cuda:0 and cuda:1.
 
-    When ``FLUX2_DUAL_GPU`` is unset this is a no-op pass-through.
+    When ``DIFFSYNTH_DUAL_GPU`` is unset this is a no-op pass-through.
 
     Call after the Flux2DiT (typically ``training_module.pipe.dit``) is
     loaded and after any PEFT LoRA injection has run. PEFT places LoRA
@@ -92,15 +92,15 @@ def enable_flux2_dual_gpu(dit: nn.Module) -> nn.Module:
 
     if torch.cuda.device_count() < 2:
         raise RuntimeError(
-            f"FLUX2_DUAL_GPU=true requires ≥2 CUDA devices, found "
+            f"DIFFSYNTH_DUAL_GPU=true requires ≥2 CUDA devices, found "
             f"{torch.cuda.device_count()}."
         )
 
     num_single = len(dit.single_transformer_blocks)
     split_at = get_split_at(num_single)
     if not 0 < split_at < num_single:
         raise RuntimeError(
-            f"FLUX2_DUAL_GPU_SPLIT_AT={split_at} out of range "
+            f"DIFFSYNTH_DUAL_GPU_SPLIT_AT={split_at} out of range "
             f"(dit has {num_single} single blocks)."
         )
 

examples/flux2/model_training/train.py

@@ -138,7 +138,7 @@ def flux2_parser():
         # pre-allocate on cuda:0 before we get a chance to split it.
         device="cpu" if (args.initialize_model_on_cpu or is_dual_gpu_enabled()) else accelerator.device,
     )
-    # Dual-GPU split: gated on FLUX2_DUAL_GPU env var. Distributes
+    # Dual-GPU split: gated on DIFFSYNTH_DUAL_GPU env var. Distributes
     # pipe.dit across cuda:0 and cuda:1 at the single_transformer_blocks
     # midpoint. Called AFTER LoRA injection (which happened inside
     # Flux2ImageTrainingModule.__init__) so PEFT LoRA params follow