SCALE on AMD is missing the CUDA Green Contexts driver 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.

SCALE's AMD target advertises CUDA 12.9 (CUDA_VERSION preprocesses to 12 * 1000 + 9 * 10) but does not declare the Green Contexts driver API that CUDA gained in 12.4. Code that gates on the advertised version, like pytorch's c10/cuda/driver_api.h, enables the API and then fails to compile.

Missing: CUdevResourceDesc, cuGreenCtxCreate, cuGreenCtxDestroy, cuGreenCtxStreamCreate, cuCtxFromGreenCtx, CU_GREEN_CTX_DEFAULT_STREAM. (The CUgreenCtx type itself is declared in redscale_impl/hostapi.h; it is the create / destroy / stream entry points and the default-stream constant that are absent.)

repro.cu:

#include <cuda.h>
static_assert(CUDA_VERSION >= 12040, "both compilers advertise >= 12.4");
int main() {
  CUgreenCtx gc = nullptr;
  CUdevResourceDesc desc = nullptr;
  CUresult r = cuGreenCtxCreate(&gc, desc, 0, CU_GREEN_CTX_DEFAULT_STREAM);
  if (r == CUDA_SUCCESS) cuGreenCtxDestroy(gc);
  CUstream s;
  cuGreenCtxStreamCreate(&s, gc, CU_STREAM_NON_BLOCKING, 0);
  CUcontext c;
  cuCtxFromGreenCtx(&c, gc);
  return 0;
}

nvcc -std=c++17 -c repro.cu                          ok
targets/gfx1100/bin/nvcc -std=c++17 -c repro.cu      error: unknown type name 'CUdevResourceDesc'; did you mean 'cudaResourceDesc'?
llvm/bin/nvcc -arch=sm_75 -std=c++17 -c repro.cu     ok (NVIDIA target has it)

Both SCALE versions fail identically, so this is a longstanding gap rather than a regression. Same family as the SM-resource partitioning API gap reported separately; split out because the symbol sets fail different pytorch translation units.

Likely cause: targets/gfx1100/include/cuda.h predates the 12.4 Green Contexts additions while redscale_impl/version.h pins the advertised version at 12.9. Either declaring the API or advertising the version SCALE actually covers would unbreak version-guarded callers.

Found building pytorch v2.12.1 (scale-validation pytorch entry bumped from v2.9.0): fails CUDAAllocatorConfig.cpp, CUDACachingAllocator.cpp, PeerToPeerAccess.cpp, driver_api.cpp, CUDAGreenContext.cpp.

[ChrisKitching]

Well, I guess this answers the question of which CUDA API we'll trip over first when trying to look at newer pytorch.

Green Contexts are probably possible to implement mostly-vacuously in SCALE, since our current implementation of contexts is already such that they are extremely lightweight. It's mostly just an exercise in nested reference counting and API plumbing.