SCALE on AMD is missing the CUDA device SM-resource partitioning API

[diskdog]

SCALE 1.7.1 and the 2026.07.02 nightly, AMD target (gfx1100), compile-only. NVIDIA target and nvcc 13.1 are fine.

Companion to the Green Contexts gap (same CUDA 12.4 feature family, reported separately): the device-resource descriptor API is missing, both the functions and the types. The missing CUdevResource type cascades in real code: std::vector<CUdevResource> can't form its element type, error recovery falls the variable back to int, and later member access on it fails in turn, e.g. request for member 'size' in 'resources', which is of non-class type 'int'. One root cause, a chain of downstream errors.

Missing: CUdevResource, CUdevResourceDesc, CU_DEV_RESOURCE_TYPE_SM, cuDeviceGetDevResource, cuDevSmResourceSplitByCount, cuDevResourceGenerateDesc.

repro.cu:

#include <cuda.h>
int main() {
  CUdevResource res{};
  cuDeviceGetDevResource(0, &res, CU_DEV_RESOURCE_TYPE_SM);
  CUdevResource split[4];
  unsigned int n = 4;
  cuDevSmResourceSplitByCount(split, &n, &res, nullptr, 0, 2);
  CUdevResourceDesc desc;
  cuDevResourceGenerateDesc(&desc, split, n);
  return 0;
}

nvcc -std=c++17 -c repro.cu                          ok
targets/gfx1100/bin/nvcc -std=c++17 -c repro.cu      error: unknown type name 'CUdevResource'
llvm/bin/nvcc -arch=sm_75 -std=c++17 -c repro.cu     ok (NVIDIA target has it)

Both SCALE versions fail identically (longstanding, not a regression).

Likely cause: same as the Green Contexts gap. targets/gfx1100/include/cuda.h predates CUDA 12.4 while the AMD target advertises CUDA 12.9, so version-guarded callers enable the API.

Found building pytorch v2.12.1: the guarded uses in c10/cuda/driver_api.h fail the c10_cuda targets alongside the Green Contexts symbols.

[ChrisKitching]

This is a tricky one, because I don't think AMD offer this feature at all. I suppose the answer is to behave like nvidia CUDA does when these APIs are used with a GPU that can't do it (I believe such devices exist: IIRC this feature is mostly only for nvidia's datacenter cards).