fix: identify GPU by PCI bus id when probing hw decompression

[aminaramoon]

Problem

query_hw_decompression took a device ordinal and resolved it with rmm::cuda_set_device_raii:

rmm::cuda_set_device_raii set_device{rmm::cuda_device_id{static_cast<int>(cuda_ordinal)}};
return rmm::detail::hwdecompress::is_supported();

Ordinals are not a stable identity across APIs. NVML enumerates in PCI-bus order, while the CUDA runtime defaults to CUDA_DEVICE_ORDER=FASTEST_FIRST. The value passed in was gpu.id — a position in this discovery's own CUDA_VISIBLE_DEVICES-filtered list — which need not name the same device to CUDA on a heterogeneous host (mixed SKUs, or a display GPU alongside compute GPUs).

The mismatch was latent because rmm::detail::hwdecompress::is_supported() only calls cudaDriverGetVersion. It answers "is the driver >= 12.8" and never asks a per-device question, so the ordinal selected nothing and the set_device was decoration. Any real per-device query would have made the wrong ordinal load-bearing.

Change

Query CU_DEVICE_ATTRIBUTE_MEM_DECOMPRESS_ALGORITHM_MASK against the device resolved by cuDeviceGetByPCIBusId:

bool query_hw_decompression(std::string const& pci_bus_id)
{
  auto const& api = load_cuda_driver_api();
  if (!api.available || pci_bus_id.empty()) { return false; }

  int device = 0;
  if (api.device_get_by_pci_bus_id(&device, pci_bus_id.c_str()) != cuda_driver_success) { return false; }

  int algorithm_mask = 0;
  if (api.device_get_attribute(&algorithm_mask,
                               cu_device_attribute_mem_decompress_algorithm_mask,
                               device) != cuda_driver_success) { return false; }
  return algorithm_mask != 0;
}

This is immune to both the FASTEST_FIRST reordering and CUDA_VISIBLE_DEVICES remapping, takes its device explicitly (no context created, calling thread's current device untouched), and reports actual silicon capability rather than a driver version. For MIG entries the bus id is the parent physical GPU's, which is the correct scope — the decompression engine is a property of the physical device.

Notes on the implementation:

• The driver API is reached via dlopen("libcuda.so.1") + dlsym rather than a link dependency, so the library still loads on hosts without an NVIDIA driver, matching how NVML is already treated here.
• CUdevice/CUresult are spelled as int to avoid pulling in <cuda.h>; both match the driver ABI.
• Attribute 136 is written literally with a comment so the file still builds against pre-12.8 toolkit headers.
• The runtime API is not an option: this CUDA version exposes no cudaDevAttrMemDecompress* equivalent, so the driver API is the only route.

This removes the only RMM use in topology_discovery.cpp (its sole source file), so rmm::rmm is dropped from the three topology targets and ${CMAKE_DL_LIBS} added. Side effect: CUCASCADE_TOPOLOGY_ONLY=ON now configures and builds — it never calls find_package(rmm), so linking rmm::rmm had it failing at generate time.

Behavior change

hw_decompression_available now reports false on pre-Blackwell GPUs that previously reported true on any >= 12.8 driver. The old code answered a different question. The field has no in-tree readers as of this branch, but downstream consumers relying on the permissive answer should be checked.

Verification

On a 2x RTX 6000 Ada host:

• Full release build clean under -Wall -Wextra -Wpedantic -Wconversion with warnings-as-errors; CUCASCADE_TOPOLOGY_ONLY=ON builds clean too.
• cucascade_topology_discovery_tests: 47 assertions in 6 test cases, all pass.
• readelf -d on libcucascade_topology_discovery.so: libdl, libstdc++, libgcc_s, libc — no libcuda, no rmm.
• pre-commit (clang-format, codespell, cmake-format, cmake-lint) passes.
• Live probe including CUDA_VISIBLE_DEVICES=1 and UUID-form masking: id renumbers to 0 but pci stays 03:00.0 and resolves to the correct physical device.
• Both GPUs now report hw_decomp=0. Confirmed a true negative rather than a swallowed error via a direct driver probe: attr136_rc=0 mask=0. Ada has no decompression engine; that is Blackwell hardware.

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[wence-]

This is only necessary if you want to be able to compile without cuda.h. Is that required?

[aminaramoon]

I guess topology discovery could, but you do have a point as cucascade has cuda as a dependency. so technicallyt it will never happen. I can change that.