Skip to content

Commit 8ed0139

Browse files
committed
docs: stop saying the same thing twice
The hardware section repeated what 'What you need' already said and what troubleshooting now covers. It keeps only what neither says: that the multi-GPU paths assume eight XGMI-connected devices, and that the single place ROCm appears at all is the optional kernel rebuild.
1 parent 487cd47 commit 8ed0139

1 file changed

Lines changed: 12 additions & 19 deletions

File tree

README.md

Lines changed: 12 additions & 19 deletions
Original file line numberDiff line numberDiff line change
@@ -563,29 +563,22 @@ what that costs a real model isn't measured here.
563563

564564
---
565565

566-
## Hardware and requirements
566+
## Hardware
567567

568-
The CPU lane runs on any x86-64 Linux host with Docker or Podman, and needs no
569-
GPU at all.
568+
[What you need](#what-you-need) covers the short version. Two things it leaves
569+
out.
570570

571-
The GPU lane needs an AMD GPU reachable at `/dev/kfd` and `/dev/dri`. The kernels
572-
are compiled for gfx950, which is MI355X on CDNA4, and other architectures will
573-
load the wrong ISA. Multi-GPU paths assume 8 XGMI-connected devices.
571+
The multi-GPU paths, meaning the collectives and anything measured against RCCL,
572+
assume eight XGMI-connected devices. Everything else runs on one.
574573

575-
No ROCm is needed anywhere, at build time or run time, and the devcontainer image
576-
contains none of it.
574+
No ROCm is needed anywhere, at build time or at run time, and the devcontainer
575+
image contains none of it. The one place ROCm appears at all is
576+
`kernels/build.sh`, which borrows an amdgpu-capable clang to recompile the kernel
577+
source, and you only need that if you change a kernel. The compiled object is
578+
committed.
577579

578-
The host user has to be in the `render` group before the container starts so that
579-
`keep-groups` can hand through access to the device nodes. If `just probe`
580-
reports the device unavailable, reconnect after the group change takes effect.
581-
582-
One thing that catches people out. The `--node <n>` argument throughout is a KFD
583-
topology node id rather than an ordinal GPU index, and `mainarch probe` lists the
584-
ones on your machine. The examples all use `--node 2` because that's where the
585-
first GPU lands on the reference host, and yours may well differ. `just demo`
586-
works it out for you.
587-
588-
Fuller detail and the validation tiers live in `docs/hardware-support.md`.
580+
`docs/hardware-support.md` has the validation tiers and what each one actually
581+
proves.
589582

590583
---
591584

0 commit comments

Comments
 (0)