wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-05-27 17:37:39 +08:00
Code Issues Actions 17 Packages Projects Releases Wiki Activity

multigpu: use unet_manual_cast for SelectModelDevice compute dtype (#14108)

Browse Source
This commit is contained in:
Jedrzej Kosinski 2026-05-25 20:03:37 -07:00 committed by GitHub
parent da49b7d0b6
commit 88956e77af
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
comfy_extras/nodes_multigpu.py
View File

@ -48,17 +48,14 @@ class MultiGPUCFGSplitNode(io.ComfyNode):
return io.NodeOutput(model) return io.NodeOutput(model)
def _force_fp32_cpu_compute(patcher: ModelPatcher): def _force_supported_compute_dtype(patcher: ModelPatcher, device: torch.device):
"""Force fp32 inference dtype for CPU. """Cast compute dtype to one the device supports; no-op if already supported."""
weight_dtype = patcher.model_dtype()
PyTorch's CPU conv2d kernels fall back to software emulation for fp16/bf16 cast_dtype = comfy.model_management.unet_manual_cast(weight_dtype, device)
and run ~500-600x slower than fp32, which makes a normal-sized workflow if cast_dtype is None:
look frozen for hours. Routing through set_model_compute_dtype leaves the return
weights as-is and casts at use, so peak memory does not blow up.""" logging.info(f"Select Model Device: using {cast_dtype} compute dtype on {device} (model weight dtype was {weight_dtype}).")
dtype = patcher.model_dtype() patcher.set_model_compute_dtype(cast_dtype)
if dtype in (torch.float16, torch.bfloat16):
logging.info(f"Select Model Device: using fp32 compute dtype for CPU inference (model dtype was {dtype}).")
patcher.set_model_compute_dtype(torch.float32)
def _remember_base_devices(patcher: ModelPatcher): def _remember_base_devices(patcher: ModelPatcher):
@ -229,8 +226,7 @@ class SelectModelDeviceNode(io.ComfyNode):
logging.warning(f"Select Model Device: cannot retarget model, passing through unchanged. ({e})") logging.warning(f"Select Model Device: cannot retarget model, passing through unchanged. ({e})")
return io.NodeOutput(model) return io.NodeOutput(model)
if resolved is not None: if resolved is not None:
if resolved.type == "cpu": _force_supported_compute_dtype(model, resolved)
_force_fp32_cpu_compute(model)
_prune_multigpu_collision(model, model.load_device) _prune_multigpu_collision(model, model.load_device)
return io.NodeOutput(model) return io.NodeOutput(model)