Tested
PyTorch "The NVIDIA driver on your system is too old": PyPI torch 2.11+ on Linux needs 580
Since torch 2.11.0 (23 March 2026), the Linux wheel that pip install torch fetches is a CUDA 13 build, and PyTorch says CUDA 13 wheels need NVIDIA driver 580 or newer. Torch 2.15.0 (28 October 2026) is the first release with no CUDA 12 build at all. The 12020 in "found version 12020" is not a torch version: it is the CUDA version your driver supports, 12.2. Two ways out: a driver of 580 or newer, or torch 2.14.1 from the cu126 index, the last CUDA 12 build, which needs a driver of 525 or newer and has no kernels for Blackwell GPUs. We ran that build only on driver 580.178.04.
Check your driver, your GPU and your torch
Run these three lines on the machine where torch fails, in the Python environment that fails. The first prints your GPU, driver version and compute capability. The second prints the CUDA version your driver supports. The third prints torch's version, the CUDA version it was built with, and whether torch can use the GPU.
nvidia-smi --query-gpu=name,driver_version,compute_cap --format=csv,noheader
nvidia-smi | grep -o 'CUDA Version: [0-9.]*'
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"If nvidia-smi rejects compute_cap as a field, the first line prints only that error, without the driver version. Then run nvidia-smi --query-gpu= instead. We saw this with a made-up field name on driver 580.178.04, not on a driver that rejects compute_cap.
On our host, with torch 2.15.0 from PyPI, they printed:
NVIDIA GeForce RTX 3090, 580.178.04, 8.6
CUDA Version: 13.0
2.15.0+cu132 13.2 TrueWith torch 2.14.1 from the cu126 index in a new environment on the same host:
NVIDIA GeForce RTX 3090, 580.178.04, 8.6
CUDA Version: 13.0
2.14.1+cu126 12.6 TrueOn our driver, the CUDA Version: 13.0 that nvidia-smi printed matched the number torch reads from the driver, 13000. We did not compare the two on any other driver. Use the table to read your own three lines.
| What the three lines show | What it means | What to do |
|---|---|---|
torch.version.cuda starts with 13 and the driver is below 580 | This page. The wheel is a CUDA 13 build and NVIDIA lists 580 as the minimum driver for CUDA 13.x. | Way out A. Way out B too, if the driver is 525 or newer and compute_cap is below 10.0 |
compute_cap is 5.0 or higher and below 7.5 | PyTorch's list calls these GPUs Maxwell, Pascal and Volta. The torch 2.15.0 wheels have no kernels for them. | Way out B, torch 2.14.1 for CUDA 12.6, on a driver of 525 or newer |
| Anything else | Not a case this page tested. A compute_cap below 5.0 is below the cu126 build's list too, which starts at 5.0. | If torch printed a found version, read it in the next section |
On a driver that is too old, torch also prints a warning before the last line, and the last line ends in False. The next section shows that output, from a simulated old driver. With torch 2.15.0 and no NumPy in the environment, a "Failed to initialize NumPy" warning came first in our runs; it is about NumPy, not the driver.
Read your "found version"
Two calls print the message, and they print it differently. torch. does not raise: it prints the message as a UserWarning that starts with CUDA initialization: and returns False. Torch counts devices once per process, so the warning appears once. The first call that initializes CUDA, such as torch.zeros(1).cuda(), raises it as a RuntimeError from torch._C._cuda_init(). Both come from the same text in c10/ in torch 2.15.0. With a driver that reports 12020 it reads:
The NVIDIA driver on your system is too old (found version 12020). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver.We had no driver older than 580, so we reproduced both calls with real torch 2.15.0 and a stand-in for the driver library: a small C library we wrote that reports a driver supporting CUDA 12.2 and supports nothing else, loaded in place of libcuda.so.1. It is a simulation of an old driver, not an old driver. is_available() printed this, with the path to your environment shortened to ...:
.../torch/cuda/__init__.py:241: UserWarning: CUDA initialization: The NVIDIA driver on your system is too old (found version 12020). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver. (Triggered internally at /__w/pytorch/pytorch/c10/cuda/CUDAFunctions.cpp:119.)
return torch._C._cuda_getDeviceCount() > 0
Falseand torch.zeros(1).cuda() raised this, ending with the same text:
Traceback (most recent call last):
File ".../torch/cuda/__init__.py", line 616, in _lazy_init
torch._C._cuda_init()
RuntimeError: The NVIDIA driver on your system is too old (found version 12020). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver.The number in the message is the value of cudaDriverGetVersion: driver_version() in the same source file calls it. NVIDIA's runtime API documentation says the function "returns the latest version of CUDA supported by the driver", encoded as 1000 × major + 10 × minor, with CUDA 9.2 written as 9020. The table maps the numbers people search for. The third column is what NVIDIA's CUDA Toolkit release notes list as the minimum driver for that toolkit's first release on Linux x86_64.
| Found version | CUDA version | NVIDIA's minimum driver for that toolkit's first release | Against NVIDIA's minimums, 525 for CUDA 12.x and 580 for CUDA 13.x |
|---|---|---|---|
| 9010 | 9.1 | Not in the tables we read | Below 525 |
| 11040 | 11.4 | 470.42.01 | Below 525 |
| 11070 | 11.7 | 515.43.04 | Below 525 |
| 12020 | 12.2 | 535.54.03 | 525 or newer, below 580 |
| 12040 | 12.4 | 550.54.14 | 525 or newer, below 580 |
| 12060 | 12.6 | 560.28.03 | 525 or newer, below 580 |
| 12070 | 12.7 | No CUDA 12.7 toolkit in NVIDIA's table | 525 or newer, below 580 |
| 12090 | 12.9 | 575.51.03 | 525 or newer, below 580 |
| 13000 | 13.0 | Branch R580, no version in the Linux table | 580 or newer |
A driver that reports one of the first eight numbers is below the minimum NVIDIA lists for CUDA 13.x, which is 580, so it is below what a CUDA 13 build of torch needs. Found versions from 12000 to 12090 sit between that and NVIDIA's minimum for CUDA 12.x, 525. That range is the one a CUDA 12 build of torch, way out B, is meant for. The numbers 9010, 11040 and 11070 are below 525, so by NVIDIA's table way out B is not for them: use way out A. We did not run a driver from either range. Our own driver, 580.178.04, reported 13000.
NVIDIA's release notes and compatibility guide both give the minimum for CUDA 13.x as >= 580. The guide's example of an older driver running a newer 13.x toolkit uses 580.65.06, which it says "was released with CUDA 13.0". LeRobot's 0.7.0 release notes say "580.65 or newer". We found no NVIDIA table that makes 580.65.06 the minimum, so this page says 580. Our driver is newer than 580.65.06, so we cannot say whether an earlier 580 release works.
NVIDIA's documentation says the function returns 0 when no driver is installed. In torch's source a driver version of 0 or below takes another branch: the .cuda() path raises "Found no NVIDIA driver on your system", and torch. returns False without a warning.
Which torch needs which driver
PyPI's own metadata and the download.pytorch.org index pages show which CUDA each release carries. Both lists are for Linux x86_64 wheels, read on 22 November 2026.
| torch | Uploaded to PyPI | CUDA build that PyPI gives on Linux | CUDA builds on download.pytorch.org | NVIDIA's minimum for that CUDA family |
|---|---|---|---|---|
| 2.10.0 | 21 January 2026 | 12.8 (nvidia-cuda-runtime-cu12 12.8.90) | cu126, cu128, cu129, cu130 | 525, CUDA 12.x |
| 2.11.0 | 23 March 2026 | 13.0 (cuda-toolkit 13.0.2) | cu126, cu128, cu129, cu130 | 580, CUDA 13.x |
| 2.12.0 | 13 May 2026 | 13.0 (cuda-toolkit 13.0.2) | cu126, cu130, cu132 | 580, CUDA 13.x |
| 2.12.1 | 17 June 2026 | 13.0 (cuda-toolkit 13.0.2) | cu126, cu129, cu130, cu132 | 580, CUDA 13.x |
| 2.13.0 | 8 July 2026 | 13.0 (cuda-toolkit 13.0.3) | cu126, cu129, cu130, cu132 | 580, CUDA 13.x |
| 2.14.0 | 2 September 2026 | 13.0 (cuda-toolkit 13.0.3) | cu126, cu130, cu132 | 580, CUDA 13.x |
| 2.14.1 | 30 September 2026 | 13.0 (cuda-toolkit 13.0.3) | cu126, cu130, cu132 | 580, CUDA 13.x |
| 2.15.0 | 28 October 2026 | 13.2 (cuda-toolkit 13.2.2) | cu132, cu134 | 580, CUDA 13.x |
On download.pytorch.org, 2.15.0 exists only as cu132 and cu134. The newest release with a cu126 build is 2.14.1, with a cu129 build 2.13.0, and with a cu128 build 2.11.0. PyTorch's notice says "PyTorch 2.14 is the last release to ship any CUDA 12.x wheels". Its 2.15.0 release notes say cu126 and cu130 are no longer built, and: "CUDA 13 wheels need NVIDIA driver 580 or newer and a GPU of Turing (sm_75) or newer, so Maxwell, Pascal and Volta GPUs and hosts with a CUDA 12.x driver should stay on 2.14."
An open PyTorch proposal for 2.16, opened on 13 November 2026, keeps CUDA 13.2 as the PyPI build and removes no CUDA version. PyTorch's release table lists 22 December 2026 as the tentative 2.16 date. Proposal for 2.16 · Release table
On Python 3.10 there is no 2.15 wheel: PyTorch's release notes say "On Python 3.10, pip cannot find a 2.15 wheel." In our resolution for Python 3.10 on Linux x86_64, uv picked torch 2.14.1 with cuda-toolkit 13.0.3, which is still a CUDA 13 build.
Way out A: a driver of 580 or newer
Install a driver of 580 or newer on the machine that has the GPU, then run the three lines above again. We did not install a driver on any machine for this page, so it gives no install commands. NVIDIA's driver download page is where the address in torch's message, http://www.nvidia.com/, redirected when we read it. For your distribution's packages, see Ubuntu Server's guide to NVIDIA drivers or NVIDIA's CUDA installation guide for Linux. On a shared cluster or a cloud image the driver belongs to the administrator or the image: ask for a driver of 580 or newer, and meanwhile use way out B if the driver is 525 or newer and compute_cap is below 10.0.
NVIDIA's release notes pair CUDA 13.2, the toolkit torch 2.15.0 is built with, with the R595 driver branch, and say that existing CUDA 13.x applications run on drivers of 580 or newer under minor version compatibility. NVIDIA's compatibility guide lists three limits of minor version compatibility: a feature that needs a newer driver can fail with cuda, code compiled to PTX does not work on older drivers, and the application must be built for a target architecture. On our driver, 580.178.04, which reported CUDA 13.0, torch 2.15.0 built for CUDA 13.2 ran one matrix multiplication and one torch.compile call and gave the expected results. We ran nothing else with it, and no 580 driver older than 580.178.04.
Way out B: torch 2.14.1 for CUDA 12.6
Torch 2.14.1 is the newest release on the cu126 index, and PyTorch's notice calls 2.14 the last release with any CUDA 12.x wheel. NVIDIA's table lists 525 as the minimum driver for CUDA 12.x toolkits under minor version compatibility, which covers the drivers that report found versions from 12000 to 12090. The limits listed above for minor version compatibility apply here too. We ran this build only on driver 580.178.04, so we did not test it on any driver in that range. The cu126 build also carries kernels for Maxwell, Pascal and Volta, which the 2.15.0 wheels do not. Its list ends at sm_90, so it has no kernels for Blackwell GPUs (compute_cap 10.0 and higher): for those, way out A is the only one. We did not run a Blackwell GPU.
The cu126 index has matching companion packages. We read each wheel's metadata rather than assume the versions:
| Package | Newest on the cu126 index | What its wheel metadata asks of torch |
|---|---|---|
| torch | 2.14.1 | |
| torchvision | 0.29.1 | torch>=2.14.0 (0.29.0 asks for torch==2.14.0) |
| torchaudio | 2.11.0 | No torch requirement line |
With pip:
pip install torch==2.14.1 torchvision==0.29.1 torchaudio==2.11.0 --index-url https://download.pytorch.org/whl/cu126Replace the three version numbers only with versions that the cu126 index lists. Keep the torch pin: torchvision 0.29.1 accepts any torch from 2.14.0, so with another index in play pip can pick a newer torch. Our run was Python 3.12.12 on Linux x86_64; the cu126 index lists torch 2.14.1 for CPython 3.10 to 3.15 there.
With uv, the pip-style form takes the index from the backend name:
uv pip install --torch-backend=cu126 torch==2.14.1 torchvision==0.29.1 torchaudio==2.11.0In a uv project, add the index and point the packages at it, as uv's PyTorch guide shows. explicit = true keeps the index for the packages you name and leaves everything else on PyPI.
[project]
name = "example"
version = "0"
requires-python = ">=3.12"
dependencies = [
"torch",
"torchvision",
]
[tool.uv.sources]
torch = [
{ index = "pytorch-cu126", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
]
torchvision = [
{ index = "pytorch-cu126", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
]
[[tool.uv.index]]
name = "pytorch-cu126"
url = "https://download.pytorch.org/whl/cu126"
explicit = trueOn our host uv lock and uv sync with this file installed torch 2.14.1+cu126 and torchvision 0.29.1+cu126. The lock file also holds torch 2.15.0 from PyPI for every platform other than Linux and Windows, where the marker does not apply. uv's --torch-backend=auto picks the index from the installed driver, and the pick depends on the uv version. On our driver, 580.178.04, uv 0.10.7 and 0.10.12 resolved torch 2.14.1 from the cu130 index; uv 0.11.33, 0.12.24 and 0.13.9 resolved torch 2.15.0 from the cu132 index. We did not run it on an older driver and cannot say what it picks there.
A package that pulls torch in
Read the package's torch requirement first: it is the requires_dist list in the package's PyPI metadata. LeRobot 0.7.0, released on 5 November 2026, asks for torch<2.16.0,>=2.7 and torchvision<0.30.0,>=, so torch 2.14.1 satisfies it and way out B exists for it. LeRobot's release notes say: "On Linux, torch and torchvision now come from PyPI as CUDA 13 builds (NVIDIA driver 580.65 or newer); the cu128 index is gone. For older drivers, install torch from a CUDA 12 index first (see the installation guide)." A package that required torch 2.15 or newer would have no way out B, and the driver would be the answer.
This is what lerobot==0.7.0 gave on our host:
| Command | torch and torchvision installed | How we tested it |
|---|---|---|
pip install lerobot==0.7.0 | 2.15.0 and 0.29.2, CUDA 13.2 | Resolution only, pip install --dry-run |
The same with --extra-index-url https:// and no pins | 2.15.0 and 0.29.2, CUDA 13.2 | Resolution only |
The same with a constraints file that pins the two +cu126 builds | 2.14.1+cu126 and 0.29.1+cu126 | Installed, pip check clean, GPU check passed |
pip install of torch and torchvision from the cu126 index, then pip install lerobot==0.7.0 | 2.14.1+cu126 and 0.29.1+cu126 | Installed, pip check clean, GPU check passed |
uv pip install --torch-backend=cu126 lerobot==0.7.0 | 2.14.1+cu126 and 0.29.1+cu126 | Installed, GPU check passed |
An extra index with no pin changed nothing: pip saw PyPI's torch 2.15.0 as the newest allowed one and took it. The second install in the fourth row left torch alone because pip reported torch<2.16.0,>=2.7 already satisfied by 2.14.1+cu126. That order is LeRobot's own advice:
pip install torch==2.14.1 torchvision==0.29.1 --index-url https://download.pytorch.org/whl/cu126
pip install lerobot==0.7.0If you cannot control the order, a constraints file keeps the pins. The temporary file goes away at the end:
C=$(mktemp)
printf 'torch==2.14.1+cu126\ntorchvision==0.29.1+cu126\n' > "$C"
pip install -c "$C" --extra-index-url https://download.pytorch.org/whl/cu126 lerobot==0.7.0
rm "$C"With uv:
uv pip install --torch-backend=cu126 lerobot==0.7.0Replace lerobot==0.7.0 with your package and version. Replace the torch and torchvision numbers with versions the cu126 index lists, and keep the +cu126 suffix in constraints: it is how that index names its builds.
Maxwell, Pascal and Volta
PyTorch's notice says: "If you are on a Maxwell, Pascal, or Volta GPU, PyTorch 2.14 is the last release with a prebuilt wheel for you." The notice gives two choices from 2.15 on: stay on 2.14 or earlier, or build PyTorch from source against a CUDA 12.x toolkit. This page covers the first; we did not try a source build. The compute_cap in the first check line is how you tell: the notice lists the three as sm_50, sm_60 and sm_70 and says CUDA 13.x does not support compute capabilities below Turing (sm_75), so a compute_cap below 7.5 is what this section is about. nvidia-smi --help-query-gpu describes the field as "The CUDA Compute Capability, represented as Major DOT Minor." We ran it only on driver 580.178.04. If an older driver rejects the field, NVIDIA's compute capability table lists the value for each GPU.
The wheel, not the driver, sets which GPUs it has kernels for. On our installs, torch. returned sm_75 sm_80 sm_86 sm_90 sm_100 sm_120 for torch 2.15.0 and sm_50 sm_60 sm_70 sm_75 sm_80 sm_86 sm_90 for torch 2.14.1 from the cu126 index.
At the first CUDA initialization, torch compares the GPU's compute capability with that list and warns. The text below is torch 2.15.0's, from torch/. The code that prints it is the same in 2.14.0 and 2.14.1, slightly different in 2.11.0 to 2.13.0, and not in 2.10.0. torch. only counts devices, so it does not run this check. We had no GPU older than Turing, so we ran torch's own check with a made-up capability of 6.1 on our RTX 3090 host. That simulates what torch prints, and says nothing about what a kernel launch does on a real older GPU, which we did not run. With torch 2.15.0 it printed two warnings, the first of which was:
Found GPU0 Example GPU which is of compute capability (CC) 6.1.
The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:
- 7.5 which supports hardware CC >=7.5,<8.0
- 8.0 which supports hardware CC >=8.0,<9.0
- 8.6 which supports hardware CC >=8.6,<9.0
- 9.0 which supports hardware CC >=9.0,<10.0
- 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
- 12.0 which supports hardware CC >=12.0,<13.0
Your installed torch==2.15.0+cu132 does not include kernels for this GPU. Reinstall the same version against a CUDA build that does, e.g.:
For CUDA 12.6 use pip install torch==2.15.0 --index-url https://download.pytorch.org/whl/cu126The second warning said that the GPU "with CUDA capability sm_61 is not compatible with the current PyTorch installation". The command in the last line of the first warning cannot work: the cu126 index has no torch 2.15.0, so use way out B with 2.14.1. For Volta, a capability of 7.0, only the first warning printed in our run. With torch 2.14.1 from the cu126 index, the same check printed nothing for 5.2, 6.1 and 7.0, and torch 2.15.0 printed nothing for 7.5.
Windows, WSL2 and containers
- Windows. The 2.15.0
win_amd64wheel for Python 3.12 on PyPI is a CPU build: itstorch/readsversion.py __version__ = '2.15.0+cpu'andcuda = None, and the wheel's CUDA requirements apply only on Linux. The cu132 index lists a Windows CUDA build of 2.15.0. We did not run torch on Windows, and we read only the cp312 wheel. - WSL2. NVIDIA's WSL guide says the driver installed on Windows is the one WSL uses, mapped into WSL as
libcuda.so, and says not to install a Linux driver inside WSL. A "too old" error in WSL2 is therefore a question about the Windows driver. We did not run WSL2. - Containers. NVIDIA's Container Toolkit guide lists installing the GPU driver on the host as a prerequisite. Run the three lines inside the container to see what it reports. NVIDIA's compatibility guide limits its forward-compatibility package to Data Center GPUs, select RTX "NGC Server Ready" cards and Jetson boards. We ran no container and did not test that package.
What we ran
We ran everything on one host: Linux x86_64, NVIDIA driver 580.178.04, an NVIDIA GeForce RTX 3090 with compute capability 8.6, Python 3.12.12, pip 25.0.1 and uv 0.10.7. Each install went into a new virtual environment. Each GPU step printed the torch version, the CUDA version, torch. and the result of a 64 × 64 matrix multiplication done on the GPU. A second check repeated each install listed below from a clean start, with Python 3.13.12, pip 25.3 and uv 0.13.9: each gave the same torch build, and the GPU check passed on each.
- torch 2.15.0 from PyPI (
pip install torch) installed as2.15.0+cu132with CUDA 13.2.is_available()was True and the multiplication gave the expected values. Onetorch.compilecall gave the same result as eager mode. - torch 2.14.1 from the cu126 index with pip installed as
2.14.1+cu126with CUDA 12.6, and torchvision 0.29.1 and torchaudio 2.11.0 imported next to it.is_available()was True, the multiplication was correct, and onetorch.compilecall matched eager mode. - The uv pip form with
--torch-backend=cu126and the uv project with the explicit index each installed2.14.1+cu126, and the multiplication was correct. - LeRobot 0.7.0 installed in the three ways in the table above, base package without extras. Each gave
2.14.1+cu126, the two pip installs passedpip check, and the GPU check passed. We did not import or run LeRobot itself. - The old-driver output above came from real torch 2.15.0 with a stand-in for the driver library, and the older-GPU output came from torch's own check with a made-up capability. Both are simulations and are labelled where they appear.
The wheels, as the indexes published their SHA-256 hashes (pip checks them on download):
torch-2.15.0-cp312-cp312-manylinux_2_28_x86_64.whl
6d8c86f4e33a9f12a58149bd9b0d64351632b8eb44479be8ff85e5389932dd79
torch-2.14.1+cu126-cp312-cp312-manylinux_2_28_x86_64.whl
75727245346d8f92224a860bf40c56bd681edcfa0456d01d7dfcb0bf6f0cf86c
torchvision-0.29.1+cu126-cp312-cp312-manylinux_2_28_x86_64.whl
f846900988cbd64b055886c9ad7e7a44606c5710e8272921062c647f237eae34
torchaudio-2.11.0+cu126-cp312-cp312-manylinux_2_28_x86_64.whl
fe8ba48c690cacc58087603c4d853580006e3d1c0ebe22c0d75f99a004bc981aWe did not run: a driver older than 580, a 580 driver older than 580.178.04, any GPU older than Turing, a Blackwell GPU, Windows, WSL2, macOS, Linux aarch64, a container, another Linux distribution, torch on drivers 535, 550 or 570, or NVIDIA's forward-compatibility package. We did not test any distribution's driver install commands.
Sources
- PyPI metadata for torch 2.10.0 to 2.15.0. Read 22 November 2026, 06:47 UTC.
- PyPI metadata for LeRobot 0.7.0. Read 22 November 2026, 06:47 UTC.
- download.pytorch.org index of torch on cu126. Read 22 November 2026, 06:47 UTC.
- download.pytorch.org index of torchvision on cu126. Read 22 November 2026, 06:47 UTC.
- download.pytorch.org index of torchaudio on cu126. Read 22 November 2026, 06:47 UTC.
- PyTorch 2.15.0 release notes. Read 22 November 2026, 06:48 UTC.
- PyTorch notice: CUDA 12.6 wheels will no longer be published from PyTorch 2.15. Read 22 November 2026, 06:47 UTC.
- PyTorch proposal for the 2.15 CUDA matrix, issue 190385. Read 22 November 2026, 06:47 to 06:48 UTC.
- PyTorch proposal for the 2.16 CUDA matrix, issue 204834. Read 22 November 2026, 06:48 UTC.
- PyTorch release table and CUDA architecture matrix (RELEASE.md). Read 22 November 2026, 06:48 UTC.
- torch v2.15.0 source: c10/cuda/CUDAFunctions.cpp. Read 22 November 2026, 06:48 UTC.
- torch v2.15.0 source: torch/cuda/__init__.py. Read 22 November 2026, 06:48 UTC.
- torch v2.15.0 source: torch/csrc/cuda/Module.cpp. Read 22 November 2026, 06:49 UTC.
- torch v2.15.0 source: torch/csrc/Exceptions.h. Read 22 November 2026, 06:56 UTC.
- torch v2.15.0 source: CUDAHooks.cpp, in aten/src/ATen/cuda/detail. Read 22 November 2026, 06:49 UTC.
- NVIDIA CUDA Toolkit release notes. Read 22 November 2026, 06:50 UTC.
- NVIDIA CUDA Compatibility: minor version compatibility. Read 22 November 2026, 06:51 UTC.
- NVIDIA CUDA Compatibility: forward compatibility. Read 22 November 2026, 06:51 UTC.
- NVIDIA CUDA Compatibility: frequently asked questions. Read 22 November 2026, 06:51 UTC.
- NVIDIA CUDA Runtime API: cudaDriverGetVersion. Read 22 November 2026, 06:51 UTC.
- NVIDIA CUDA on WSL User Guide. Read 22 November 2026, 06:50 UTC.
- NVIDIA Container Toolkit installation guide. Read 22 November 2026, 07:02 UTC.
- NVIDIA driver downloads, the page torch's message points to. Read 22 November 2026, 07:14 UTC.
- NVIDIA CUDA installation guide for Linux. Read 22 November 2026, 06:57 UTC.
- NVIDIA CUDA GPU compute capability. Read 22 November 2026, 06:57 UTC.
- Ubuntu Server: NVIDIA drivers installation. Read 22 November 2026, 06:57 UTC.
- uv: Using uv with PyTorch. Read 22 November 2026, 06:52 UTC.
- LeRobot 0.7.0 release notes. Read 22 November 2026, 06:57 UTC.
Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Send corrections to hello@cybernative.ai. We date corrections on this page.