#5683 [Benchmark] the machine line names a disabled iGPU and 2.0 GiB of VRAM instead of the two cards the engine used

closed · @Kompasby · 1 commentaires · Sur GitHub

BenchmarksSetup & installMulti-GPUAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindows

Description

### Area

[Benchmark] (measurements and reports)

### Desktop app, Terminal or source?

Terminal (the `strata` command)

### What happened

Running the standard suite with `strata bench run --url ...` on a Windows PC with two discrete AMD cards writes the wrong GPU into both the summary and the report's `README.md`.

Summary printed by the runner:

```text
Machine: AMD Radeon(TM) Graphics; AMD Ryzen 7 9700X 8-Core Processor; 64 GB RAM; Windows 10.0.26300.
```

The same line in `<out>/README.md`:

```text
- GPU: AMD Radeon(TM) Graphics (2.0 GiB dedicated)
```

Neither card that actually ran the model appears. `AMD Radeon(TM) Graphics` is the name of the processor's integrated adapter, and that adapter is **disabled in the BIOS** on this machine. The engine in the same session used an RX 6800 XT (gfx1030) and an RX 6600 (gfx1032) in a layer split, and its own log names them. `Win32_VideoController` on the same machine lists only the two discrete cards.

Expected instead: the machine facts name the accelerator(s) the run used — both of them for a split — with their real VRAM, or at least not an adapter that is disabled and not in use.

The measurements themselves were fine (21/21 requests, sanity 12/12); only the hardware attribution is wrong.

### Smallest steps or command that reproduce it

1. Windows 11 (10.0.26300), Ryzen 7 9700X, integrated graphics disabled in the BIOS, RX 6800 XT 16 GiB + RX 6600 8 GiB installed.
2. Start the engine with a layer split across the two cards (`--layer-split auto`). Its log says `GPU 0: AMD Radeon RX 6800 XT (gfx1030)` and `layer split: CUDA1 runs layers 39-47`.
3. Run:
   ```text
   strata bench run --url http://127.0.0.1:8080 --suite standard-v1 --skip-model-hash --track custom --out <DIR> --output human
   ```
4. Read the `Machine:` line of the summary and the `GPU:` line of `<DIR>/README.md`.

Both runs I made that day (different page-size states of the same machine) wrote the same wrong GPU line.

### Reproduction status

Reproduced locally

### Strata version

strata 0.2.11 (build 0.1.57; python 3.12.15), source build; engine 0.1.57 for `gfx1030;gfx1032;gfx1036`

### Graphics card, memory and system

RX 6800 XT 16 GiB (gfx1030) + RX 6600 8 GiB (gfx1032); integrated gfx1036 disabled in the BIOS; 63.7 GiB RAM; Windows 11 10.0.26300; driver 32.0.21045.16010

### Model, size and settings

Qwen3.8-Flash-Next IQ3_S · 131,072 context · KV int8 · pictures on · expert arena in system RAM (40,290 MiB), 7,362 experts in the VRAM caches

### Logs

```text
# summary printed by the runner
Machine: AMD Radeon(TM) Graphics; AMD Ryzen 7 9700X 8-Core Processor; 64 GB RAM; Windows 10.0.26300.

# the report's README.md, same run
- GPU: AMD Radeon(TM) Graphics (2.0 GiB dedicated)

# the engine's own log, same session
strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)
strata generate: GPU 0: AMD Radeon RX 6800 XT (gfx1030)
strata generate: PCIe probe: 28.3 GB/s host->device (best of 28.1 28.2 28.3 28.2) -> pcie_frac 0.55
strata generate: layer split: CUDA1 runs layers 39-47, expert cache 1835 slots (3.86 GiB)

# Win32_VideoController on the same machine — only the two discrete cards
AMD Radeon RX 6600     | Status OK
AMD Radeon RX 6800 XT  | Status OK
```

### Two extra observations (not part of the bug itself)

- On this machine `Win32_VideoController.AdapterRAM` reports **4,095 MB for both cards** — the documented 32-bit cap of that field. A fix that starts from that value would still print wrong VRAM, so the source may need to be DXGI 2, `hipDeviceProp_t.totalGlobalMem`, or the engine's own device list.
- The report this came from is PR #5682 (`bench/results/2026-11-09-community-rx6800xt-rx6600-iq3s/`); the real cards are described in prose there, because the generated line was wrong.
- Possibly the same root cause as #3418 (the app's single-GPU picker on a Ryzen iGPU + discrete Radeon): both would come from an adapter enumeration that includes the integrated device. Not verified.

Sur le site

Liens install, modèles, releases.