torch::install_torch() succeeds (all files present) but library(torch) fails with "Lantern is not loaded" / system.file() "no file found" — Windows, v0.17.0, cu128

[lundc01]

Environment:

• torch version: 0.17.0
• OS: Windows 10/11
• CUDA build: cu128 (also reproduced with cu129 before switching)
• R: 4.6.1

Problem:
install_torch() completes successfully every time — both LibTorch (~3.2GB)
and Lantern (~2.3MB) download to their exact reported byte sizes, with no
warnings. But library(torch) / torch::cuda_is_available() then fails with:

Error: Lantern is not loaded. Please use install_torch() to install
additional dependencies.
Caused by error in system.file(): ! no file found

What I've ruled out:

• Confirmed lantern.dll, torch.dll, torch_cuda.dll, and all CUDA/cuDNN DLLs
  ARE physically present in .../win-library/4.6/torch/lib/ (verified via
  File Explorer screenshot)
• .libPaths() correctly includes the folder where these files live
• Not an antivirus/quarantine issue (Windows Defender Protection History
  empty; exclusions added for the install folder regardless)
• Not a path-length issue (tested with TORCH_HOME set to a short path,
  C:/torchlib — same error)
• Not RStudio-specific (identical error in plain RGui, outside RStudio)
• Not a stale/corrupted previous install (identical error after
  remove.packages() + manual folder deletion + fresh install)
• Not a download/network issue (both files match exact reported byte size
  every time, multiple independent attempts)
• Tried creating a "deps" subfolder (matching an older, possibly outdated
  internal path convention) as a copy of "lib" — no change
• rlang::last_trace() reports no error was recorded; options(error =
  traceback) also produces no additional stack trace — the error appears
  to be caught and re-messaged internally before reaching R's standard
  error-reporting mechanisms, so I can't get further diagnostic detail
  from the R session itself.

Happy to provide any additional diagnostics on request — just not sure
what else to check from the outside at this point.

[cregouby]

Hello @lundc01,

torch 0.17.0 has a install_torch_sitrep() that may help.

Would you share the command output ?

[maxwell-geospatial]

I believe I am having a similar problem using both R 4.6.1 and R 4.5.3 with torch 0.17.0 and CUDA 12.8. I have tried the standard install method, installing from binaries for R 4.5.3, and when using the cuda12.8 R package. Even if I can get LibTorch and Lantern installed into torch/lib and do not get the "Lantern is not loaded" error when loading torch, R crashes as soon as cuDNN is used, such as when running a simple 2D convolution operation. I thought that system-wide CUDA installs may have been an issue; however, uninstalling all system-wide installs of CUDA and reinstalling my NVIDIA drivers (am using an RTX 2080 Ti) did not make a difference.

[lundc01]

Hello @lundc01,

torch 0.17.0 has a install_torch_sitrep() that may help.

Would you share the command output ?

Here is the output from the sitrep:

── System Information ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
• OS: "Windows" (10 x64 build 26200)
• R Version: "R version 4.6.1 (2026-06-24 ucrt)"
• Platform: "x86_64-w64-mingw32"
• Architecture: "x86-64"

── Package Information ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
[1] "2.8.0+cu128"

[cregouby]

Hello @lundc01,

I've update the sitrep in #1480 to better consider windows users in here : https://github.com/cregouby/torch/blob/bugfix/sitrep/R/install.R

Could you source that file (or copy paste the [install_torch_sitrep()](https://torch.mlverse.org/docs/reference/install_torch_sitrep) function ) and rerun it in your environment ?

[cregouby]

I believe I am having a similar problem using both R 4.6.1 and R 4.5.3 with torch 0.17.0 and CUDA 12.8. I have tried the standard install method, installing from binaries for R 4.5.3, and when using the cuda12.8 R package. Even if I can get LibTorch and Lantern installed into torch/lib and do not get the "Lantern is not loaded" error when loading torch, R crashes as soon as cuDNN is used, such as when running a simple 2D convolution operation. I thought that system-wide CUDA installs may have been an issue; however, uninstalling all system-wide installs of CUDA and reinstalling my NVIDIA drivers (am using an RTX 2080 Ti) did not make a difference.

Hello @maxwell-geospatial,

Would you be kind to open a specific issue to provide your own updated install_torch_sitrep() output, so that conversations are easier to follow, and specific issue easier to close ?
Thanks.

[lundc01]

Hello @lundc01,

I've update the sitrep in #1480 to better consider windows users in here : https://github.com/cregouby/torch/blob/bugfix/sitrep/R/install.R

Could you source that file (or copy paste the [install_torch_sitrep()](https://torch.mlverse.org/docs/reference/install_torch_sitrep) function ) and rerun it in your environment ?

Here is the requested output:
── System Information ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
• OS: "Windows" (10 x64 build 26200)
• R Version: "R version 4.6.1 (2026-06-24 ucrt)"
• Platform: "x86_64-w64-mingw32"
• Architecture: "x86-64"

── Package Information ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
• torch package: "0.17.0"
• Install path: C:\Users\AppData\Local\R\win-library\4.6\torch

── File System Check ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
✔ Core installation files are present on disk.
C:\Users\AppData\Local\R\win-library\4.6\torch/lib and C:\Users\AppData\Local\R\win-library\4.6\torch/bin directory contents:
[lib]: asmjit.dll, asmjit.lib, c10.dll, c10.lib, c10_cuda.dll, c10_cuda.lib, caffe2_nvrtc.dll, caffe2_nvrtc.lib, cpuinfo.lib, cublas64_12.dll, cublasLt64_12.dll, cudart64_12.dll,
cudnn_adv64_9.dll, cudnn_cnn64_9.dll, cudnn_engines_precompiled64_9.dll, cudnn_engines_runtime_compiled64_9.dll, cudnn_graph64_9.dll, cudnn_heuristic64_9.dll, cudnn_ops64_9.dll, cudnn64_9.dll,
cufft64_11.dll, cufftw64_11.dll, cupti64_2025.1.1.dll, curand64_10.dll, cusolver64_11.dll, cusolverMg64_11.dll, cusparse64_12.dll, dnnl.lib, fbgemm.dll, fbgemm.lib, fmt.lib, kineto.lib,
lantern.dll, lantern.lib, libiomp5md.dll, libiompstubs5md.dll, libittnotify.lib, libprotobuf-lite.lib, libprotobuf.lib, libprotoc.lib, microkernels-prod.lib, nvJitLink_120_0.dll,
nvrtc-builtins64_128.dll, nvrtc64_120_0.alt.dll, nvrtc64_120_0.dll, pthreadpool.lib, sleef.lib, torch.dll, torch.lib, torch_cpu.dll, torch_cpu.lib, torch_cuda.dll, torch_cuda.lib,
torch_global_deps.dll, uv.dll, XNNPACK.lib, zlibwapi.dll
[bin]: protoc.exe

── CUDA Detection ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
ℹ System CUDA version: "12.8"
ℹ Installation kind: "cu128"
✔ Detected CUDA version is supported.

── Runtime Load Test ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
✖ torch failed to load. Starting automated diagnosis...
✖ DIAGNOSIS: Installation is incomplete (files missing).

── Environment Variables ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
ℹ No torch-specific environment variables set.

── Summary ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
✔ No issues detected. torch is working correctly.

[cregouby]

Hello @lundc01,

the diagnostic is that files are present on disk, but R cannot initialize them. It maybe you are missing Visual C++ Redistributable (VCR), or R's system.file() cannot resolve the torch package path, even though files exist on disk, Thus R cannot find the package internally. A check of.libPaths() could tell that the package might be installed in a location R doesn't scan, or the package metadata is corrupted. You may also have a PATH conflict (python / conda / uv / poetry / ... ) that leads to the wrong dll loading.

I've update the script (again) so that it would have provide you the complete diagnostic.

You may reinstall via
Sys.setenv(TORCH_INSTALL_DEBUG = '1'); torch::install_torch(reinstall = TRUE)

[lundc01]

Hi, thanks for continuing to dig into this! A quick recap since a few of these have already been tested pretty thoroughly on my end:

Visual C++ Redistributable: already installed and reinstalled (x64) — no change.
.libPaths(): confirmed correct — it includes the exact folder where the files live (verified via File Explorer too).
PATH conflict (python/conda/etc.): I have a separate tool (TagLab) using PyTorch on this machine. I directly tested by stripping all Python-related entries from PATH for the R session before loading torch — identical error persisted.

Two new things since my last update, which might be more useful:

I tried Sys.setenv(TORCH_VERIFY_LOAD = "FALSE") to skip the subprocess-based load check entirely. library(torch) then loads with no error message at all — but torch::cuda_is_available() still fails immediately with the same "Lantern is not loaded" error. So the problem doesn't seem to be in that verification subprocess itself, but in the actual Lantern loading that happens on first real use.
I've now confirmed this is not a download/installation issue through three independent methods, all producing byte-verified-complete files, and all ending in the exact same error once loading is attempted:
Fresh download via the default libcurl method
Fresh download via the wininet method
A manually downloaded (browser), file-size-verified-complete local zip via TORCH_URL

Quick clarification — I don't see a link to the updated script in your message. Could you share the specific URL (or commit hash) for the latest version? I want to make sure I'm running the exact version you intend, since I noticed install_torch_sitrep() was already part of the released v0.17.0 I have installed (as torch_sitrep()), and it gave a much simpler report than what I've seen described elsewhere — so I suspect there have been further changes since that release.