[randomizedcoder]

Part 14/14. Part of splitting upstream PR 573885 (rocmPackages: 7.2.4 -> 10.1) into smaller PRs, as asked in its review.

• Switches rocmPackages to ROCm 10.1
• Target: staging (mass rebuild)

Changes

• rocmPackages = recurseIntoAttrs rocmPackages_10
  • rocmPackages_10 stays a real attribute, so it works with allowAliases = false
• Removes the 7.2.4 scope and the 50 package directories that rocm-modules/10/ replaces
  • Kept, still shared: aotriton, hipcc, hiprt, mivisionx, mscclpp, rocm-docs-core, rocm-tests, rocprof-compute-viewer, update.nix
• torch:
  • adds libdrm to the ROCm build inputs and propagates it (ROCm 10's rocm_smi-config.cmake needs it)
  • adds rocmPackages.hipfile: torch 2.14 requires it with ROCm 7.14 and later; it reaches OpenSplat and torchcodec through cxxdev
  • drops the OMP_TOOL workaround from upstream PR 573885 (fixed in rocprofiler-sdk, rocmPackages_10: add profiling and debugging tools, RCCL #21)
• vllm: adds hipfile, because it lists its ROCm inputs itself instead of using torch's cxxdev
• ollama: uses the rocm_v7_2 backend label (there is no ROCm 10 label yet), with a TODO
• Release notes, including the new rocmPackages.hipfile
• Carries the prep PRs it needs (rocmPackages.rocm-bandwidth-test: remove #10, opensplat: add pkg-config for ROCm 10 #15, torchcodec: add libdrm for ROCm 10 #16, rocmPackages.rocprof-trace-decoder: take comgr and runtime from scope #17); they drop out once merged

Testing

• After the switch, every rocmPackages derivation (default, llvm, gfx906, gfx1102 scopes) is identical to rocmPackages_10's before it
• No file refers to a removed path; the ROCm consumers evaluate with aliases on and off
• Built on x86_64-linux with the default GPU targets: all 73 rocmPackages derivations plus consumers, 204 derivations in total
  • blender, btop, zluda, adaptivecpp, whisper-cpp, stable-diffusion-cpp, onnxruntime, magma, ctranslate2, ollama, llama-cpp
  • torch (Python 3.13 and 3.14), torchaudio, torchcodec, xformers, vllm, amd-aiter, bitsandbytes
• nixpkgs-review: 85 packages built, including torchWithRocm for Python 3.13 and 3.14; one failure, opensplat (below)

Hardware smoke tests:

check	gfx1102 (Radeon Pro W7500)	gfx906 (MI50) ¹
rocminfo, amd-smi	pass	pass
hipcc vector add	pass	pass
import torch; torch.cuda.device_count() = 1	pass	pass
fp32 / fp16 matmul (max rel err)	9.6e-7 / 4.5e-4	5.5e-7 / 4.5e-4
conv2d forward + backward (MIOpen)	pass (≤ 1.6e-6)	pass (≤ 1.6e-5)
torch.profiler GPU kernels	pass	pass
rocprofv3 --kernel-trace	7 kernels	7 kernels
import vllm	pass	pass
llama.cpp lists ROCm0	pass	pass

¹ Run with ROCR_VISIBLE_DEVICES=0: that machine also holds a gfx803 card that HIP skips but amdsmi counts.

Known failure, not from this PR

• opensplat has been broken on master in every variant since torch 2.14.1 (2026-10-22)
• OpenSplat 1.1.4 builds as C++17; torch 2.14.1's headers require C++20
• OpenSplat 1.1.5 fixes it (upstream PR 553190)
• Its configure step passes with ROCm 10.1; compilation then stops on the C++20 check

Series

Preparatory PRs, each against master: #10, #11, #15, #16, #17.

Part	PR	Contents	Base
1	#18	rocmPackages_10: init	master
2	#19	llvm (LLVM 24)	#18
3	#20	runtime	#19
4	#21	profiling and debugging tools, RCCL	#20
5	#22	rocPRIM, rocRAND, rocFFT	#21
6	#23	hipBLASLt	#22
7	#24	Tensile and rocBLAS	#23
8	#25	sparse, solvers, rocALUTION, rocWMMA, hipfort	#24
9	#26	composable_kernel	#25
10	#27	rocMLIR	#26
11	#28	MIOpen	#27
12	#29	MIGraphX, RPP, MIVisionX	#28
13	#30	remaining packages	#29
▶ 14	#31	rocmPackages: 7.2.4 -> 10.1 (the switch) (this PR)	#30

---

🤖 Generated with Claude Code

[randomizedcoder]

Opened upstream as NixOS#579527.