One PEP 503 index per GPU architecture. The wheels themselves are
GitHub release assets; these pages point at them, with a
#sha256= on every link so pip verifies what it downloads.
pip install \ --index-url https://sbates130272.github.io/rocm-bake/data/wheels/gfx1151/ \ --extra-index-url https://pypi.org/simple \ "rocm-torch[device-gfx1151]"
| architecture | torch | triton | vllm | rocm-torch |
platform tags |
|---|---|---|---|---|---|
gfx1151 | 2.15.0a0+git31a77fc.rocm10.1.0 | 3.9.0+gitaad2a60d.rocm10.1.0 | 0.33.0+rocm101 | 10.1.0+gfx1151.2d896c4b | 5 manylinux · 4 linux_x86_64 · 3 pure |
gfx1201 | 2.15.0a0+git31a77fc.rocm10.1.0 | 3.9.0+gitaad2a60d.rocm10.1.0 | 0.33.0+rocm101 | 10.1.0+gfx1201.d2cb4667 | 5 manylinux · 4 linux_x86_64 · 3 pure |
gfx1250 | 2.15.0a0+git31a77fc.rocm10.1.0 | 3.9.0+gitaad2a60d.rocm10.1.0 | 0.33.0+rocm101 | 10.1.0+gfx1250.2d896c4b | 5 manylinux · 4 linux_x86_64 · 3 pure |
gfx950 | 2.15.0a0+git31a77fc.rocm10.1.0 | 3.9.0+gitaad2a60d.rocm10.1.0 | 0.33.0+rocm101 | 10.1.0+gfx950.2d896c4b | 5 manylinux · 4 linux_x86_64 · 3 pure |
A version says which ROCm a wheel was built against. Nothing in a
filename says which GPU: torch for two
architectures is the same filename and differs by tens of MB of device
code. That is what the per-arch index path and the
rocm-torch metapackage carry, and why installing through
the index is the only correct way to get these.
The extra index is not optional — ordinary
dependencies resolve from PyPI. Our wheels win over same-named PyPI
projects because the metapackage pins exact versions whose local
segment (+rocm…) upstream cannot match.
Platform tags differ by wheel and are not a choice: auditwheel grants a manylinux tag only where the compiled code's glibc symbols allow one. Some wheels here reference symbols newer than any defined policy and stay linux_x86_64; some contain no CPU code at all. Each architecture's index lists the measured tag and the glibc behind it, under Platform tags.