since new version: Flux throws error when for train! / update! even on quick start problem

[dorn-gerhard]

I updated to the new version Flux 0.14.6 (and also CUDA 5.1.1) and got the illegal memory access error (see below) when using train! or update
machine: Windows Laptop
GPU: NVIDIA GeForce GTX 1050
CUDA Version: 12.0
Julia Version: 1.9.4
packages in environment:

st
Status `C:\Users\Gerhard\.julia\environments\GPU\Project.toml`
  [052768ef] CUDA v5.1.1
  [587475ba] Flux v0.14.6
  [92933f4c] ProgressMeter v1.9.0
  [02a925ec] cuDNN v1.2.1
  [10745b16] Statistics v1.9.0

I am aware of the new breaking changes in Flux and CUDA API, so
I just tested the quick start example and got the following error:

ERROR: CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)
Stacktrace:
 [1] throw_api_error(res::CUDA.cudaError_enum)
   @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\libcuda.jl:27
 [2] isdone
   @ C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\stream.jl:111 [inlined]
 [3] spinning_synchronization(f::typeof(CUDA.isdone), obj::CuStream)
   @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\synchronization.jl:79
 [4] device_synchronize(; blocking::Bool, spin::Bool)
   @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\synchronization.jl:171
 [5] device_synchronize()
   @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\synchronization.jl:169
 [6] top-level scope
   @ C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\src\initialization.jl:210

caused by: CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)
Stacktrace:
  [1] throw_api_error(res::CUDA.cudaError_enum)
    @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\libcuda.jl:27
  [2] nonblocking_synchronize(val::CuContext)
    @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\synchronization.jl:163
  [3] device_synchronize(; blocking::Bool, spin::Bool)
    @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\synchronization.jl:174
  [4] device_synchronize
    @ C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\synchronization.jl:169 [inlined]
  [5] CuModule(data::Vector{UInt8}, options::Dict{CUDA.CUjit_option_enum, Any})
    @ CUDA C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\module.jl:40
  [6] CuModule
    @ C:\Users\Gerhard\.julia\packages\CUDA\YIj5X\lib\cudadrv\module.jl:23 [inlined]
  [7] link(job::GPUCompiler.CompilerJob, compiled::WARNING: Error while freeing DeviceBuffer(256 bytes at 0x0000000204607c00):
CUDA.CuError(code=CUDA.cudaError_enum(0x000002bc), details=CUDA.Optional{String}(data=nothing))

Downgrading to Flux v0.14.5 resolved the problem.

Any hints, what I am doing wrong with the new version?

[ToucheSir]

Can you try running the MWE in FluxML/Zygote.jl#1473 and see if you get a similar error? This looks like a problem on the CUDA.jl side, but to help them we should reduce it as much as possible.

[dorn-gerhard]

thank you for the hint - you are right, CUDA@5.1.1 triggers the following error:

ERROR: WARNING: Error while freeing DeviceBuffer(400 bytes at 0x0000000205200a00):
CUDA.CuError(code=CUDA.cudaError_enum(0x000002bc), details=CUDA.Optional{String}(data=nothing))

Stacktrace:
  [1] throw_api_error(res::CUDA.cudaError_enum)

[henry2004y]

I have the same issue. With CUDA.jl v5.1.1 and Flux v0.14.7 on Win11 using Julia v1.10-rc3, all the demos with GPU failed with error an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS).

[ToucheSir]

Per FluxML/Zygote.jl#1473 (comment), the fix hasn't landed in a tagged version of CUDA.jl yet.

[CarloLucibello]

this should be solved