Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

UniPCMultistepScheduler fails in torch.linalg.solve under a float64 default dtype: the unit entry of rks takes the default dtype

未关闭 适合新手
  • 哪里不对？
#14,888 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
88/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python, pytorch
领域
machine-learning, testing-qa

调研方向

从 src/diffusers/schedulers/scheduling_unipc_multistep.py 中 multistep_uni_p_bh_update 和 multistep_uni_c_bh_update 的 rks 构造开始，然后阅读 tests/schedulers/test_scheduler_unipc.py 并运行默认 dtype 覆盖。完成标准是 scheduler 以 torch.float64 作为默认 dtype 运行时不会出现 linalg.solve dtype 错误，同时现有行为和测试保持不变。

由索引模型根据 Issue 内容生成。

描述

Describe the bug

multistep_uni_p_bh_update and multistep_uni_c_bh_update build rks from the sigmas, which are float32, and append torch.ones((), device=device), which takes the default dtype. Under torch.set_default_dtype(torch.float64) the torch.stack(rks) promotes R to float64 while b stays float32, and the first torch.linalg.solve(R, b) (the order-2 corrector on the second step with the default config) raises

RuntimeError: linalg.solve: Expected A and B to have the same dtype, but found A of type Double and B of type Float instead

whatever the dtype of the sample (float16, float32 and float64 all fail). DPMSolverMultistepScheduler, DEISMultistepScheduler and SASolverScheduler run under a float64 default dtype.

Fix, patch below: torch.ones_like(h) in both places, so the unit entry has the dtype and device of the other rks entries. Under the float32 default this is the same tensor as before: 540 configurations (solver_order 1 to 3, bh1/bh2, the three prediction_types, Karras sigmas or not, predict_x0, thresholding, 1/2/3/10/25 steps) are bit-identical to main. test_default_dtype_float64 fails on main with the error above and passes with the patch, the file's 64 tests pass, ruff is clean. I can open the PR if this looks right to you.

Reproduction
import torch

from diffusers import UniPCMultistepScheduler

torch.set_default_dtype(torch.float64)

scheduler = UniPCMultistepScheduler()
scheduler.set_timesteps(10)
sample = torch.rand(1, 3, 8, 8, generator=torch.Generator().manual_seed(0), dtype=torch.float32)
for t in scheduler.timesteps:
    sample = scheduler.step(sample * t / (t + 1), t, sample).prev_sample
print(sample.dtype, sample.abs().mean())
Logs
Traceback (most recent call last):
  File "repro.py", line 11, in <module>
    sample = scheduler.step(sample * t / (t + 1), t, sample).prev_sample
  File ".../diffusers/schedulers/scheduling_unipc_multistep.py", line 1191, in step
    sample = self.multistep_uni_c_bh_update(
  File ".../diffusers/schedulers/scheduling_unipc_multistep.py", line 1078, in multistep_uni_c_bh_update
    rhos_c = torch.linalg.solve(R, b).to(device).to(x.dtype)
RuntimeError: linalg.solve: Expected A and B to have the same dtype, but found A of type Double and B of type Float instead

With the patch: torch.float32 tensor(0.2337, dtype=torch.float32), the same value as under the float32 default dtype.

System Info
  • diffusers 0.40.0 (PyPI) and main at e0abab8
  • torch 2.14.0+cpu, numpy 2.2.6
  • Python 3.13, Linux

AI disclosure: I used an AI coding agent to help find this, to write the reproduction, the patch, the test and this report. I have read and checked the report, the patch and the test myself and I will answer questions personally.

Patch
git diff against main (2 files, +15 -2) 
diff --git a/src/diffusers/schedulers/scheduling_unipc_multistep.py b/src/diffusers/schedulers/scheduling_unipc_multistep.py
index 5c2cbcc..e8f5af1 100644
--- a/src/diffusers/schedulers/scheduling_unipc_multistep.py
+++ b/src/diffusers/schedulers/scheduling_unipc_multistep.py
@@ -903,7 +903,7 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
             rks.append(rk)
             D1s.append((mi - m0) / rk)
 
-        rks.append(torch.ones((), device=device))
+        rks.append(torch.ones_like(h))
         rks = torch.stack(rks)
 
         R = []
@@ -1038,7 +1038,7 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
             rks.append(rk)
             D1s.append((mi - m0) / rk)
 
-        rks.append(torch.ones((), device=device))
+        rks.append(torch.ones_like(h))
         rks = torch.stack(rks)
 
         R = []
diff --git a/tests/schedulers/test_scheduler_unipc.py b/tests/schedulers/test_scheduler_unipc.py
index ac7e1d3..92ff4e7 100644
--- a/tests/schedulers/test_scheduler_unipc.py
+++ b/tests/schedulers/test_scheduler_unipc.py
@@ -257,6 +257,19 @@ class UniPCMultistepSchedulerTest(SchedulerCommonTest):
 
                     assert sample.dtype == torch.float16
 
+    def test_default_dtype_float64(self):
+        # the unit entry of `rks` took the default dtype, the other entries the dtype of the sigmas
+        default_dtype = torch.get_default_dtype()
+        torch.set_default_dtype(torch.float64)
+        try:
+            sample = self.full_loop(solver_order=3)
+        finally:
+            torch.set_default_dtype(default_dtype)
+        result_mean = torch.mean(torch.abs(sample))
+
+        assert sample.dtype == torch.float64
+        assert abs(result_mean.item() - torch.mean(torch.abs(self.full_loop(solver_order=3))).item()) < 1e-3
+
     def test_full_loop_with_noise(self):
         scheduler_class = self.scheduler_classes[0]
         scheduler_config = self.get_scheduler_config()
主要语言
Python
星标
34.6k
派生
7.4k
平均合并
4 天 5 小时
30 天内合并 PR
47

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

huggingface/diffusers 的其他 Issue

查看 huggingface/diffusers 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。