Open syl20bnr opened 4 months ago
The upstream issue in PyTorch seems to be fixed (https://github.com/pytorch/pytorch/issues/124009).
@syl20bnr can we try updating it again and see if this is fixed? We should try it before the upcoming release.
Yep, I'll look into it next week,
Looking at it while I am refactoring our CI.
The issue is still happening but with another DLL:
INTEL MKL ERROR: The specified module could not be found. mkl_vml_def.1.dll. Intel MKL FATAL ERROR: cannot load mkl_vml_def.1.dll.
It would appear tch 0.17 is now available. It features libtorch-2.4
FWIW: This is probably not related, I'm trying to get burn working using libtorch on my Radeon GPU. pyTorch-2.3 and 2.4 work fine, yet burn appers not to work when changing tch from 0.15 to 0.17.
Running benches/custom_gelu.rs (target/benchmarks/release/deps/custom_gelu-82b2276b553d5723)
thread 'main' panicked at /home/oleid/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tch-0.17.0/src/wrappers/tensor_generated.rs:8361:40:
called `Result::unwrap()` on an `Err` value: Torch("Could not run 'aten::empty.memory_format' with arguments from the 'CUDA' backend. This could be because the operator doesn't exist for this backend, or was omitted during the selective/custom build process (if using custom build). If you are a Facebook employee using PyTorch on mobile, please visit https://fburl.com/ptmfixes for possible resolutions. 'aten::empty.memory_format' is only available for these backends: [CPU, Meta, QuantizedCPU, QuantizedMeta, MkldnnCPU, SparseCPU, SparseMeta, SparseCsrCPU, SparseCsrMeta, BackendSelect, Python, FuncTorchDynamicLayerBackMode, Functionalize, Named, Conjugate, Negative, ZeroTensor, ADInplaceOrView, AutogradOther, AutogradCPU, AutogradCUDA, AutogradHIP, AutogradXLA, AutogradMPS, AutogradIPU, AutogradXPU, AutogradHPU, AutogradVE, AutogradLazy, AutogradMTIA, AutogradPrivateUse1, AutogradPrivateUse2, AutogradPrivateUse3, AutogradMeta, AutogradNestedTensor, Tracer, AutocastCPU, AutocastXPU, AutocastCUDA, FuncTorchBatched, BatchedNestedTensor, FuncTorchVmapMode, Batched, VmapMode, FuncTorchGradWrapper, PythonTLSSnapshot, FuncTorchDynamicLayerFrontMode, PreDispatch, PythonDispatcher].\n
[...]
Update
tch
once the upstream fix is released inPytorch 2.3.1
andtch
is updated.See compilation bug issue: https://github.com/LaurentMazare/tch-rs/issues/870