Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix `torch.clamp` / `torch.clamp_min` / `torch.clamp_max` return gradient 1.0 at boundary points where `F.relu6` and `F.hardtanh` return 0.0 for the same mathematical function

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…

Code Example

import torch
import torch.nn.functional as F

# F.relu6(x) = torch.clamp(x, 0, 6)  — identical functions
for x_val, label in [(0.0, "lower boundary"), (6.0, "upper boundary")]:
    x1 = torch.tensor(x_val, requires_grad=True)
    F.relu6(x1).backward()

    x2 = torch.tensor(x_val, requires_grad=True)
    torch.clamp(x2, 0.0, 6.0).backward()

    print(f"x={x_val} ({label}):  F.relu6={x1.grad.item()}  torch.clamp={x2.grad.item()}")

print()

# F.hardtanh(x, -1, 1) = torch.clamp(x, -1, 1)  — identical functions
for x_val, label in [(-1.0, "lower boundary"), (1.0, "upper boundary")]:
    x1 = torch.tensor(x_val, requires_grad=True)
    F.hardtanh(x1, -1.0, 1.0).backward()

    x2 = torch.tensor(x_val, requires_grad=True)
    torch.clamp(x2, -1.0, 1.0).backward()

    print(f"x={x_val} ({label}):  F.hardtanh={x1.grad.item()}  torch.clamp={x2.grad.item()}")

---

x=0.0 (lower boundary):  F.relu6=0.0  torch.clamp=1.0
x=6.0 (upper boundary):  F.relu6=0.0  torch.clamp=1.0

x=-1.0 (lower boundary):  F.hardtanh=0.0  torch.clamp=1.0
x=1.0  (upper boundary):  F.hardtanh=0.0  torch.clamp=1.0

---

x=0.0 (lower boundary):  F.relu6=0.0  torch.clamp=0.0
x=6.0 (upper boundary):  F.relu6=0.0  torch.clamp=0.0

x=-1.0 (lower boundary):  F.hardtanh=0.0  torch.clamp=0.0
x=1.0  (upper boundary):  F.hardtanh=0.0  torch.clamp=0.0

---

import torch
import torch.nn.functional as F

ops = [
    ("F.relu(x)",                  lambda x: F.relu(x)),
    ("F.relu6(x) at x=0",         lambda x: F.relu6(x)),
    ("F.relu6(x) at x=6",         lambda x: F.relu6(x)),
    ("F.hardtanh(x,-1,1) at x=1", lambda x: F.hardtanh(x, -1.0, 1.0)),
    ("torch.clamp(x,0,6) at x=0", lambda x: torch.clamp(x, 0.0, 6.0)),
    ("torch.clamp(x,0,6) at x=6", lambda x: torch.clamp(x, 0.0, 6.0)),
    ("torch.clamp_min(x,0)",       lambda x: torch.clamp_min(x, 0.0)),
    ("torch.clamp_max(x,6)",       lambda x: torch.clamp_max(x, 6.0)),
]
xvals = [0.0, 0.0, 6.0, 1.0, 0.0, 6.0, 0.0, 6.0]

for (name, fn), xv in zip(ops, xvals):
    x = torch.tensor(xv, requires_grad=True)
    fn(x).backward()
    print(f"{name:<35} grad={x.grad.item()}")

---

F.relu(x)                           grad=0.0   ✓
F.relu6(x) at x=0                   grad=0.0   ✓
F.relu6(x) at x=6                   grad=0.0   ✓
F.hardtanh(x,-1,1) at x=1           grad=0.0   ✓
torch.clamp(x,0,6) at x=0           grad=1.0   ✗
torch.clamp(x,0,6) at x=6           grad=1.0   ✗
torch.clamp_min(x,0)                 grad=1.0   ✗
torch.clamp_max(x,6)                 grad=1.0   ✗

---

PyTorch version: 2.13.0.dev20260512+cu130
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.4 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version: 18.1.3 (1ubuntu1)
CMake version: version 3.28.3
Libc version: glibc-2.39

Python version: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39
Is CUDA available: True
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5090
Nvidia driver version: 590.48.01

[pip3] numpy==2.4.4
[pip3] torch==2.13.0.dev20260512+cu130
[pip3] triton==3.7.0+git88b227e2
RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

torch.clamp, torch.clamp_min, and torch.clamp_max return gradient 1.0 at their boundary points. F.relu6 and F.hardtanh are the same mathematical function as torch.clamp but return gradient 0.0 at their boundaries. Two implementations of the same function disagree on the gradient.

Minimal reproducer

import torch
import torch.nn.functional as F

# F.relu6(x) = torch.clamp(x, 0, 6)  — identical functions
for x_val, label in [(0.0, "lower boundary"), (6.0, "upper boundary")]:
    x1 = torch.tensor(x_val, requires_grad=True)
    F.relu6(x1).backward()

    x2 = torch.tensor(x_val, requires_grad=True)
    torch.clamp(x2, 0.0, 6.0).backward()

    print(f"x={x_val} ({label}):  F.relu6={x1.grad.item()}  torch.clamp={x2.grad.item()}")

print()

# F.hardtanh(x, -1, 1) = torch.clamp(x, -1, 1)  — identical functions
for x_val, label in [(-1.0, "lower boundary"), (1.0, "upper boundary")]:
    x1 = torch.tensor(x_val, requires_grad=True)
    F.hardtanh(x1, -1.0, 1.0).backward()

    x2 = torch.tensor(x_val, requires_grad=True)
    torch.clamp(x2, -1.0, 1.0).backward()

    print(f"x={x_val} ({label}):  F.hardtanh={x1.grad.item()}  torch.clamp={x2.grad.item()}")

Observed output

x=0.0 (lower boundary):  F.relu6=0.0  torch.clamp=1.0
x=6.0 (upper boundary):  F.relu6=0.0  torch.clamp=1.0

x=-1.0 (lower boundary):  F.hardtanh=0.0  torch.clamp=1.0
x=1.0  (upper boundary):  F.hardtanh=0.0  torch.clamp=1.0

Expected output

x=0.0 (lower boundary):  F.relu6=0.0  torch.clamp=0.0
x=6.0 (upper boundary):  F.relu6=0.0  torch.clamp=0.0

x=-1.0 (lower boundary):  F.hardtanh=0.0  torch.clamp=0.0
x=1.0  (upper boundary):  F.hardtanh=0.0  torch.clamp=0.0

Why 0.0 is the correct answer

1. PyTorch's own autograd documentation states:

"If the function is convex (at least locally), use the sub-gradient of minimum norm."

clamp(x, lo, hi) is convex. At a boundary point (e.g. x = lo), the subdifferential is the interval [0, 1]. The element of minimum norm is 0.0.

2. Internal consistency: F.relu6 and F.hardtanh already do the right thing.

F.relu6(x) = torch.clamp(x, 0, 6) and F.hardtanh(x, a, b) = torch.clamp(x, a, b) — these are not approximations, they are the same operation. Both return 0.0 at boundaries, consistent with the documented convention. torch.clamp returning 1.0 is therefore a bug in torch.clamp's backward, not a design choice.

3. This is consistent with the existing F.relu / torch.maximum convention.

F.relu(0.0) gradient = 0.0, also following the min-norm rule. torch.clamp_min(x, 0) and torch.clamp(x, min=0) are equivalent to F.relu but return 1.0 at x=0 — another inconsistency in the same family.

Full picture of affected ops

import torch
import torch.nn.functional as F

ops = [
    ("F.relu(x)",                  lambda x: F.relu(x)),
    ("F.relu6(x) at x=0",         lambda x: F.relu6(x)),
    ("F.relu6(x) at x=6",         lambda x: F.relu6(x)),
    ("F.hardtanh(x,-1,1) at x=1", lambda x: F.hardtanh(x, -1.0, 1.0)),
    ("torch.clamp(x,0,6) at x=0", lambda x: torch.clamp(x, 0.0, 6.0)),
    ("torch.clamp(x,0,6) at x=6", lambda x: torch.clamp(x, 0.0, 6.0)),
    ("torch.clamp_min(x,0)",       lambda x: torch.clamp_min(x, 0.0)),
    ("torch.clamp_max(x,6)",       lambda x: torch.clamp_max(x, 6.0)),
]
xvals = [0.0, 0.0, 6.0, 1.0, 0.0, 6.0, 0.0, 6.0]

for (name, fn), xv in zip(ops, xvals):
    x = torch.tensor(xv, requires_grad=True)
    fn(x).backward()
    print(f"{name:<35} grad={x.grad.item()}")
F.relu(x)                           grad=0.0   ✓
F.relu6(x) at x=0                   grad=0.0   ✓
F.relu6(x) at x=6                   grad=0.0   ✓
F.hardtanh(x,-1,1) at x=1           grad=0.0   ✓
torch.clamp(x,0,6) at x=0           grad=1.0   ✗
torch.clamp(x,0,6) at x=6           grad=1.0   ✗
torch.clamp_min(x,0)                 grad=1.0   ✗
torch.clamp_max(x,6)                 grad=1.0   ✗

Versions

PyTorch version: 2.13.0.dev20260512+cu130
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.4 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version: 18.1.3 (1ubuntu1)
CMake version: version 3.28.3
Libc version: glibc-2.39

Python version: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39
Is CUDA available: True
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5090
Nvidia driver version: 590.48.01

[pip3] numpy==2.4.4
[pip3] torch==2.13.0.dev20260512+cu130
[pip3] triton==3.7.0+git88b227e2

cc @ezyang @albanD @gqchen @nikitaved @soulitzer @Varal7 @bobrenjc93

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

pytorch - 💡(How to fix) Fix `torch.clamp` / `torch.clamp_min` / `torch.clamp_max` return gradient 1.0 at boundary points where `F.relu6` and `F.hardtanh` return 0.0 for the same mathematical function