#5494 [Other] Windows: recognise the Radeon Pro V620 (PCI id 0x73A1) as gfx1030

closed · @LouisClt · 0 コメント · GitHub で見る

Setup & installMulti-GPUAMD / HIPNVIDIA / CUDADocumentationWindowsLinux

本文

## Problem and result

On Windows, Strata identifies an AMD card by its PCI device id, then by its name. The Radeon Pro V620 (Navi 21, gfx1030,
32 GB) has device id `0x73A1`, which is in neither table, and Windows names it "AMD Radeon Pro V620", which no name rule maps.
So setup finds no architecture for it, and `strata update` / Control do not count it.

Seen on one PC with a V620, an RX 7700S (gfx1102) and a Radeon 780M. `python strata.py update --check` on `origin/main` (146df3c4):

```
selected for this PC: 0.1.58 windows-x86_64-hip-rdna3
- 0.1.58 windows-x86_64-hip-rdna2 has code for gfx1030, gfx1031, this PC's GPU is gfx1102
- 0.1.58 windows-x86_64-hip-rx9070 has code for gfx1200, gfx1201, this PC's GPU is gfx1102
```

With this branch:

```
selected for this PC: 0.1.58 windows-x86_64-hip-rdna3
- 0.1.58 windows-x86_64-hip-rx9070 has code for gfx1200, gfx1201, this PC's GPU is gfx1030, gfx1102
```

The package chosen does not change on this PC (see Limits); the V620 is now counted and the rdna2 package is no longer refused.

The card itself runs: the PC's owner installed with the id added to setup's table by hand, and the server log of that install
shows `GPU 0: AMD Radeon Pro V620 (gfx1030)` and the requests listed in the docs row below.

## Related discussion

None that covers this fix. #4342 mentions Radeon V620s on Windows that the app would not set up with the HIP engine; the cause
was not established there (that PC also has a Quadro display card). #4668 reports 2x V620 on Linux, where the card is found
through KFD.

## Approach

`0x73A1: "gfx1030"` in setup.py's `_WIN_AMD_DID` and in `engine_component.WIN_AMD_DEVICE_ARCH` (`test_engine_component.py`
keeps the two equal). No name rule is added: the id is what this PC reports, and a name rule would also catch V620 variants
with other ids that nobody has run. Other cards are unchanged. On Linux the same table is the fallback when KFD cannot be read
(`hip_archs_from_pci`), so a V620 is also identified there.

docs/AMD_HIP.md gets one row in "Other gfx103x and RDNA2-adjacent cards (community reports)" with what the server log of that
PC showed, labelled as read from the log.

## Validation

**Passed:**
- `python -m pytest tests/terminal/test_gfx_selection.py tests/terminal/test_engine_component.py tests/terminal/test_update_feed.py`
  (this branch, Windows 11 26300, Python 3.11.9): 98 passed, 2 skipped.
- The new test `test_the_radeon_pro_v620_is_a_gfx1030` fails on `origin/main` code (`['', 'gfx1102'] != ['gfx1030', 'gfx1102']`)
  and passes on this branch.
- `python tools/test_setup_amd.py`: 53 tests OK.
- `python strata.py update --check` on the PC above: the output in Problem and result.

**Failed:**
- `python -m pytest tests/terminal --continue-on-collection-errors` (this branch, same environment): 278 failed, 2124 passed,
  87 skipped, 21 collection errors. The 274 test ids in its FAILED summary were run again on `origin/main` (146df3c4) in the
  same environment: all 274 fail there too. The collection errors were not compared. Most failures are `ModuleNotFoundError`
  for `regex`, `jinja2` and `rich`, which this Python does not have.

**Not run:**
- setup.py from this branch was not run again end to end on the V620 PC; the owner's install used the same table entry added by hand.
- No engine code changes, so no CUDA, HIP, SYCL or Metal build.
- The `rdna2` engine package on the V620: not installed; the numbers in the docs row are from engines 0.1.41 and 0.1.44.1.
- Linux: the KFD fallback path was not run.

## Limits

- Only the id `0x73A1` is covered. Other Navi 21 workstation and server cards with ids not in the table (for example the
  Radeon PRO W6800, #3988) are not changed here.
- On a PC with an RX 6000-class card and an RX 7000-class card, `strata update` prefers the `rdna3` package, and
  `--backend hip-rdna2` is ignored once both cards are identified. That is a separate problem and not part of this PR.

Developed with an AI coding assistant; every command result above was run on the PC described.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

関連リンク

インストール・モデル・リリースへのサイト内リンク。