Skip to content

Update fm_solvers_unipc.py Fix For CUSOLVER_STATUS_INTERNAL_ERROR #1110

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+24 −3
Open

Update fm_solvers_unipc.py Fix For CUSOLVER_STATUS_INTERNAL_ERROR #1110

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 24 additions & 3 deletions wanvideo/schedulers/fm_solvers_unipc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,28 @@
if is_scipy_available():
import scipy.stats

# === sm_120 hotfix: robust wrapper per torch.linalg.solve ===
def _safe_solve(A: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
"""
Sostituto robusto per torch.linalg.solve su GPU recenti (sm_120):
- cast a float32
- garantisce contiguità e RHS di forma (n,1) se vettore
- fallback CPU se il solver CUDA fallisce (CUSOLVER_STATUS_INTERNAL_ERROR)
Ritorna con il dtype originale di b.
"""
dev = A.device
A32 = A.to(torch.float32).contiguous()
b_is_vec = (b.ndim == 1)
b32 = b.to(torch.float32).contiguous()
if b32.ndim == 1:
b32 = b32.unsqueeze(-1)
try:
x = torch.linalg.solve(A32, b32)
except Exception:
x = torch.linalg.solve(A32.cpu(), b32.cpu()).to(dev)
if b_is_vec:
x = x.squeeze(-1)
return x.to(b.dtype)

class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
"""
Expand Down Expand Up @@ -494,8 +516,7 @@ def multistep_uni_p_bh_update(
if order == 2:
rhos_p = torch.tensor([0.5], dtype=x.dtype, device=device)
else:
rhos_p = torch.linalg.solve(R[:-1, :-1],
b[:-1]).to(device).to(x.dtype)
rhos_p = _safe_solve(R[:-1, :-1], b[:-1]).to(device).to(x.dtype)
else:
D1s = None

Expand Down Expand Up @@ -640,7 +661,7 @@ def multistep_uni_c_bh_update(
if order == 1:
rhos_c = torch.tensor([0.5], dtype=x.dtype, device=device)
else:
rhos_c = torch.linalg.solve(R, b).to(device).to(x.dtype)
rhos_c = _safe_solve(R, b).to(device).to(x.dtype)

if self.predict_x0:
x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0
Expand Down