Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 5 additions & 5 deletions .github/workflows/e2e-selfhosted.yml
Original file line number Diff line number Diff line change
Expand Up @@ -329,7 +329,7 @@ jobs:
# instantly (observed on run 29320025393). Installing in place keeps the
# baked paths valid, and each scenario's data/runtimes symlink resolves to
# this same real tree.
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# Build the rocm and rocmd binaries ONCE and reuse them for both the
Expand Down Expand Up @@ -510,7 +510,7 @@ jobs:
# is missing)` and fails (diagnosed on the box 2026-07-15). Installing in
# place, directly into the persistent shared dir, keeps the baked paths
# valid for all scenarios and for the end-of-run version probe.
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# Build the rocm and rocmd binaries once; reuse them for pre-warm + suite.
Expand Down Expand Up @@ -831,7 +831,7 @@ jobs:
# every later serve sees status=unusable and fails (diagnosed on the Linux
# box 2026-07-15; the Windows scenario-8 cold-download failure is the same
# class). Install in place so the baked paths stay valid for all scenarios.
$prewarm = "$env:RUNNER_WORKSPACE\e2e-prewarm-multi-arch-v2"
$prewarm = "$env:RUNNER_WORKSPACE\e2e-prewarm-next-v1"
$env:E2E_SHARED_RUNTIMES_DIR = "$prewarm\data\runtimes"

# E2E_SHARED_CACHE_DIR: every Linux lane sets this; the harness's own
Expand Down Expand Up @@ -1415,7 +1415,7 @@ jobs:
# correctness, not speed: `install sdk` bakes ABSOLUTE paths into the
# runtime manifest, so installing into a per-scenario temp dir leaves
# every later serve pointing at a deleted install root.
prewarm="/root/work/e2e-prewarm-multi-arch-v2"
prewarm="/root/work/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# Build the rocm and rocmd binaries once; reuse them for pre-warm + suite.
Expand Down Expand Up @@ -1598,7 +1598,7 @@ jobs:
# optimization — `install sdk` bakes absolute paths into the runtime
# manifest, so installing anywhere temporary breaks every later serve.
# See e2e-gpu for the full rationale.
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# See the e2e-gpu lane for why the e2e-test-hooks feature must match
Expand Down
14 changes: 7 additions & 7 deletions .github/workflows/nightly.yml
Original file line number Diff line number Diff line change
Expand Up @@ -438,7 +438,7 @@ jobs:
# The shared dir IS the pre-warm's own data/runtimes; NEVER move it — a
# post-install mv invalidates the absolute paths install sdk bakes into
# the runtime manifest and every serve fails instantly (run 29320025393).
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# Build both binaries once; reuse them for pre-warm + suite.
Expand Down Expand Up @@ -529,7 +529,7 @@ jobs:
# (see e2e-selfhosted.yml e2e-gpu).
runner_root="$(dirname "$(dirname "$RUNNER_WORKSPACE")")"
export E2E_SHARED_UV_CACHE_DIR="$runner_root/uv-cache"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# See the e2e-gpu lane in e2e-selfhosted.yml for why the
Expand Down Expand Up @@ -605,7 +605,7 @@ jobs:
# (see e2e-selfhosted.yml e2e-gpu).
runner_root="$(dirname "$(dirname "$RUNNER_WORKSPACE")")"
export E2E_SHARED_UV_CACHE_DIR="$runner_root/uv-cache"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# See the e2e-gpu lane in e2e-selfhosted.yml for why the
Expand Down Expand Up @@ -687,7 +687,7 @@ jobs:
# (see e2e-selfhosted.yml e2e-gpu).
runner_root="$(dirname "$(dirname "$RUNNER_WORKSPACE")")"
export E2E_SHARED_UV_CACHE_DIR="$runner_root/uv-cache"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# See the e2e-gpu lane in e2e-selfhosted.yml for why the
Expand Down Expand Up @@ -774,7 +774,7 @@ jobs:
run: |
export CARGO_TARGET_DIR="$RUNNER_WORKSPACE/e2e-target"
export E2E_SHARED_CACHE_DIR="$RUNNER_WORKSPACE/e2e-shared"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-multi-arch-v2"
prewarm="$RUNNER_WORKSPACE/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# See the e2e-gpu lane in e2e-selfhosted.yml for why the
Expand Down Expand Up @@ -1021,7 +1021,7 @@ jobs:
# absolute paths. On the ephemeral DevLab Dispatch pool this host is
# destroyed after the job, so the paths only need to stay valid for
# this run's scenarios, not across runs.
$prewarm = "$env:RUNNER_WORKSPACE\e2e-prewarm-multi-arch-v2"
$prewarm = "$env:RUNNER_WORKSPACE\e2e-prewarm-next-v1"
$env:E2E_SHARED_RUNTIMES_DIR = "$prewarm\data\runtimes"

# See e2e-gpu-strix-windows in e2e-selfhosted.yml: the harness's own
Expand Down Expand Up @@ -1459,7 +1459,7 @@ jobs:
# correctness, not speed: `install sdk` bakes ABSOLUTE paths into the
# runtime manifest, so installing into a per-scenario temp dir leaves
# every later serve pointing at a deleted install root.
prewarm="/root/work/e2e-prewarm-multi-arch-v2"
prewarm="/root/work/e2e-prewarm-next-v1"
export E2E_SHARED_RUNTIMES_DIR="$prewarm/data/runtimes"

# See the e2e-gpu lane in e2e-selfhosted.yml for why the
Expand Down
15 changes: 10 additions & 5 deletions MANIFEST.md
Original file line number Diff line number Diff line change
Expand Up @@ -659,12 +659,17 @@ ROCm distribution. Two install formats are supported:

- **Wheel format** — Python wheel packages (`rocm`, `torch`, `torchvision`,
`torchaudio`) are resolved from AMD-hosted PyPI-compatible indexes and
installed via `uv` into a managed virtual environment. Release channel wheels
are served from `https://repo.amd.com/rocm/whl-multi-arch`. Nightly channel
wheels are served from `https://rocm.nightlies.amd.com/whl-multi-arch`.
installed via `uv` into a managed virtual environment. An unpinned release
install, and a pin of stable ROCm 10 or newer, resolves wheels from
`https://stable.repo.amd.com/rocm/whl-next`. A pin of an older stable release,
such as `--version 7.14.1`, uses `https://repo.amd.com/rocm/whl-multi-arch`.
Nightly channel wheels are served from
`https://rocm.nightlies.amd.com/whl-multi-arch`.
- **Tarball format** — Prebuilt SDK tarballs are downloaded from AMD-hosted
artifact storage. Release channel tarballs are served from
`https://repo.amd.com/rocm/tarball/`. Nightly tarballs are served from
artifact storage. An unpinned release install, and a pin of stable ROCm 10
or newer, uses `https://stable.repo.amd.com/rocm/core/tarball/`. A runtime
installed from `https://repo.amd.com/rocm/tarball/` stays on that catalog
when it is updated. Nightly tarballs are served from
`https://rocm.nightlies.amd.com/tarball/`. The metadata index files for
tarball installs are optionally verified with an RSA/SHA-256 detached
signature (controlled by `ROCM_CLI_METADATA_PUBLIC_KEY_PATH` /
Expand Down
45 changes: 27 additions & 18 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -252,11 +252,13 @@ This downloads TheRock ROCm wheels and a matching PyTorch stack into a managed
environment. Running `rocm install sdk` creates a managed runtime alongside any
legacy install.

ROCm CLI supports ROCm 7.14, 10.0, and 10.1. Without `--version`,
`rocm install sdk` currently installs ROCm 7.14. To install ROCm 10.0 or 10.1,
add `--version`. If ROCm CLI can't tell which exact GPU arch this computer has,
or you want a different one, also add `--family` with the raw `gfx` code from
`rocm examine`. For example: `--version 10.1.0 --family gfx1200`.
ROCm CLI supports ROCm 7.14 and the ROCm 10 line. Without `--version`,
`rocm install sdk` on the release channel installs the newest stable ROCm 10
release. That release needs an exact GPU arch. If ROCm CLI can't tell which
arch this computer has, or you want a different one, add `--family` with the
raw `gfx` code from `rocm examine`. For example: `--family gfx1200`.
To install ROCm 7.14.1 instead, pass `--version 7.14.1`. `--channel nightly`
with no version stays on the nightly stream.

Running `rocm install sdk` when a managed runtime is already the active default
asks first, because the new install takes over as the active default; see
Expand Down Expand Up @@ -455,9 +457,10 @@ rocm update [--apply] [--runtime KEY] [--activate] [--dry-run]
```

`install sdk` downloads TheRock ROCm wheels into a Python environment managed
by ROCm CLI. ROCm CLI supports ROCm 7.14, 10.0, and 10.1. Without `--version`,
it currently installs ROCm 7.14. To install 10.0 or 10.1, follow the steps under
ROCm 10 and newer, below.
by ROCm CLI. On the release channel, with no `--version`, it installs the
newest stable ROCm 10 release. Pass `--version 7.14.1` to install that older
release instead. `--channel nightly` with no version stays on the nightly
stream. ROCm 10 needs an exact GPU arch; see ROCm 10 and newer, below.

To remove ROCm CLI and what it manages, see
[Logs and cleanup](#logs-and-cleanup).
Expand Down Expand Up @@ -531,14 +534,16 @@ active default runtime, not about what a named prefix loses.

#### ROCm 10 and newer

ROCm 10 and newer ship from a different source layout. You opt in by pinning the
version with `--version`. These versions select their device payload by exact
GPU arch, so when ROCm CLI can't resolve one for this computer (it detects only
a family of GPUs, or none), or you want to target a different GPU, also pass
`--family` with the raw `gfx` code instead of a family label:
ROCm 10 and newer ship from a different source layout than ROCm 7. On the
release channel, `rocm install sdk` with no `--version` installs the newest
stable release from that layout and prepares Python 3.14 for it. These
versions select their device payload by exact GPU arch, so when ROCm CLI can't
resolve one for this computer (it detects only a family of GPUs, or none), or
you want to target a different GPU, also pass `--family` with the raw `gfx`
code instead of a family label:

```
rocm install sdk --version 10.1.0 --family gfx1200 --dry-run
rocm install sdk --family gfx1200 --dry-run
```

A family label such as `--family gfx120X-all` is rejected for those versions
Expand All @@ -565,10 +570,14 @@ dependency metadata alongside the pinned torchaudio. Before it creates or
changes a managed runtime, it validates that every selected framework package
carries the same ROCm build identifier.

Nothing about this happens on its own. Without a `--version` of 10 or newer,
`install sdk` resolves the same release and nightly sources as before. It doesn't
quietly retry against the ROCm 10 sources when a lookup finds nothing; it tells
you what it couldn't find instead.
An older release stays on the earlier source. `rocm install sdk --version 7.14.1`
installs 7.14.1 from that source and uses Python 3.12. `--channel nightly` with
no version stays on the nightly stream. Pinning a ROCm 10 version still works,
for example `rocm install sdk --version 10.1.0 --family gfx1200`. Updating a
runtime keeps the source layout that installed it, so `rocm update` does not
move a ROCm 7 runtime onto ROCm 10. The install does not quietly retry the
other source when a lookup finds nothing; it tells you what it couldn't find
instead.

#### Driver installation

Expand Down
13 changes: 11 additions & 2 deletions apps/rocm/src/main.rs
Original file line number Diff line number Diff line change
Expand Up @@ -777,10 +777,16 @@ enum BenchCommand {
#[derive(Subcommand, Debug)]
enum InstallTarget {
/// Install TheRock ROCm wheels into a Python folder managed by ROCm CLI.
///
/// With no `--version` on the release channel, this installs the newest
/// stable ROCm 10 release and needs an exact GPU arch (`--family gfx1200`).
/// `--channel nightly` with no version stays on the nightly index.
/// `--version 7.14.1` installs that release from the earlier index.
#[command(after_help = "EXAMPLES:\n \
rocm install sdk\n \
rocm install sdk --family gfx1200 --dry-run\n \
rocm install sdk --version 7.14.1 --family gfx110X-all --dry-run\n \
rocm install sdk --channel nightly --build-date 2025-01-15\n \
rocm install sdk --family gfx110X-all --dry-run\n \
rocm install sdk --devel")]
Sdk {
/// Package channel to install, such as release or nightly.
Expand All @@ -800,7 +806,10 @@ rocm install sdk --devel")]
build_date: Option<String>,
/// TheRock GPU package family to install, such as gfx110X-all.
///
/// A GPU target ROCm publishes no build for is refused rather than
/// ROCm 10, including `rocm install sdk` with no `--version` on the
/// release channel, rejects a family label and asks for an exact arch
/// such as `--family gfx1200`. A GPU target ROCm publishes no build
/// for is refused rather than
/// mapped to a near neighbour: `--family gfx90c` - the Vega iGPU in
/// Renoir, Lucienne, Cezanne and Barcelo APUs - fails and installs
/// nothing, instead of substituting another family. A plain `rocm
Expand Down
Loading
Loading