mirror of
https://github.com/ryan4yin/nix-config.git
synced 2026-11-13 12:13:46 +01:00
feat(aquamarine): make the iGPU usable for media and OpenVINO (#474)
* feat(aquamarine): make the iGPU usable for media and OpenVINO The R9700s stay reserved for inference: VA-API/QSV transcoding moves to the iGPU media engine, and `openvino-gpu` works around ROCm's OpenCL loader shadowing the plugin's `ocl-icd` (WA-038). Assisted-by: DeepSeek Harness 0.2.1-rc.2 (deepseek-v4.1-flash) * docs(aquamarine): document OpenVINO on the iGPU How to run models on the iGPU: the `openvino-gpu` wrapper, the encoder staticization Whisper needs on `GPU`, compile/steady timings and memory, and the language and cache caveats. Assisted-by: DeepSeek Harness 0.2.1-rc.2 (deepseek-v4.1-flash)
This commit is contained in:
ryan4yin | 二花 authored and 
GitHub committed
1 parent
b5d71abd39
commit
4258538588
7 files changed
+203
-12
No files matched your search
+9
-8
@@ -66,14 +66,15 @@ only delete inactive duplicates.
|
||||
|
||||
A known limitation with no fix yet.
|
||||
|
||||
| ID | What & where | Why | Removal condition | Added | Revisit | Status |
|
||||
| ------ | ---------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------------------------------------------------- | ------ |
|
||||
| WA-016 | subscription rule snapshots — `modules/nixos/desktop/networking/mihomo/` (`rules_from:`) | mihomo's `proxy-providers` imports proxies only, so the `rules` and `proxy-groups` a subscription ships with are dropped by the core; the generator re-imports them from a local YAML copy that has to be refreshed by hand when the provider changes its routing policy | Drop the snapshots once mihomo can import rules from a proxy-provider, or the providers publish their rules as rule-providers | 2026-10-04 | when mihomo gains rule import | active |
|
||||
| WA-022 | qemu-user binfmt removed — `hosts/12kingdoms-shoukei/hardware-configuration.nix` | the x86-on-arm module owns the x86 ELF handlers and this host does not need riscv64, but dropping `boot.binfmt.emulatedSystems` also drops `nix.settings.extra-platforms` and qemu-user, so `podman run --platform=linux/amd64`, x86 chroots and local builds of uncached x86 derivations need a remote x86_64 builder | Restore `boot.binfmt.emulatedSystems` (without x86_64/i686, which the module asserts against) if such a workload is needed | 2026-10-07 | when an amd64 container/chroot or uncached x86 build is needed | active |
|
||||
| WA-025 | brcmfmac 4-way-handshake offload authentication — upstream kernel/wpa_supplicant, no code in this repo | Since hostap commit 416386060, wpa_supplicant 2.11+ expects `NL80211_CMD_PORT_AUTHORIZED` from drivers that offload the 4-way handshake; brcmfmac never reports it, so authentication intermittently fails on Apple Silicon (BRCM4378/4387/4388). nixpkgs already carries the Broadcom revert in wpa_supplicant 2.12; the proper fix, Janne Grunau's `wifi: brcmfmac: cfg80211: Report port_authorized for 4-way HS offload` (2026-09), is still under review. Entering the key again after such a failure only creates a duplicate profile, so cancel and retry instead | Fix lands via a nixpkgs kernel update (or by carrying the patch through `nixpkgs-patched`) | 2026-10-07 | on every kernel bump | active |
|
||||
| WA-033 | USB-C/XHCI port `_PLD` ACPI noise — this board's DSDT, no code in this repo | every boot logs 16 `ACPI Error: Aborting method \_SB.PC00.TXHC.RHUB.SS01._PLD ... (AE_NOT_FOUND)` lines, plus `\_SB.PC02.XHCI.RHUB.HS01._PLD`; the ports work, and these messages are unrelated to the GPU's ATPX failure, which names `\_SB.PC00.GFX0.ATPX` instead | Drop if a BIOS update defines the `_PLD` descriptors | 2026-10-14 | on a BIOS update | active |
|
||||
| WA-034 | IPv6-first resolution is dead on the wired link — `hosts/idols-aquamarine/default.nix` (`networking.getaddrinfo.precedence`) | the proxy gateway answers DNS in fake-IP mode and its IPv6 path is dead: an AAAA lookup returns a fake address whose connection fails at once (measured: `curl -6` fails in 8 ms, `curl -4` answers in 0.7 s). glibc clients are covered by the RFC 6724 table with the IPv4-mapped row raised; nix's libcurl and other non-glibc resolvers ignore `/etc/gai.conf`. Disabling IPv6 with `boot.kernel.sysctl` did not work here — networkd re-enables it on the wired link for `DHCP = "ipv6"`/`IPv6AcceptRA` — so that setting was dropped on 2026-10-14 | Drop once the gateway serves proxied names over IPv6 as well (the adjacent fix is `dns.ipv6: false` in mihomo, see WA-030) | 2026-10-14 | on a mihomo or gateway change | active |
|
||||
| WA-035 | SSH retry + multiplexing to LAN peers, ARP fallback tuning — `home/base/tui/ssh.nix`, `hosts/idols-ai/neighbour-probe.nix` | the desktop reaches the LAN over a Mesh WiFi backhaul, which drops unicast to a peer whose MAC it has aged out of its forwarding table while still flooding broadcast: a capture shows seven SYNs from one socket addressed to the cached MAC with no answer, three unicast ARP probes unanswered, then a broadcast ARP answered in 0.4 ms after ~12 s of failed connects. The ssh settings retry a short connect and keep a master warm so the entry stays fresh; the neighbour-probe tunings walk the kernel to that broadcast fallback in ~1 s | Drop once the backhaul is wired or the mesh stops dropping unknown unicast | 2026-10-16 | when the backhaul changes | active |
|
||||
| ID | What & where | Why | Removal condition | Added | Revisit | Status |
|
||||
| ------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------------------------------------------------- | ------ |
|
||||
| WA-016 | subscription rule snapshots — `modules/nixos/desktop/networking/mihomo/` (`rules_from:`) | mihomo's `proxy-providers` imports proxies only, so the `rules` and `proxy-groups` a subscription ships with are dropped by the core; the generator re-imports them from a local YAML copy that has to be refreshed by hand when the provider changes its routing policy | Drop the snapshots once mihomo can import rules from a proxy-provider, or the providers publish their rules as rule-providers | 2026-10-04 | when mihomo gains rule import | active |
|
||||
| WA-022 | qemu-user binfmt removed — `hosts/12kingdoms-shoukei/hardware-configuration.nix` | the x86-on-arm module owns the x86 ELF handlers and this host does not need riscv64, but dropping `boot.binfmt.emulatedSystems` also drops `nix.settings.extra-platforms` and qemu-user, so `podman run --platform=linux/amd64`, x86 chroots and local builds of uncached x86 derivations need a remote x86_64 builder | Restore `boot.binfmt.emulatedSystems` (without x86_64/i686, which the module asserts against) if such a workload is needed | 2026-10-07 | when an amd64 container/chroot or uncached x86 build is needed | active |
|
||||
| WA-025 | brcmfmac 4-way-handshake offload authentication — upstream kernel/wpa_supplicant, no code in this repo | Since hostap commit 416386060, wpa_supplicant 2.11+ expects `NL80211_CMD_PORT_AUTHORIZED` from drivers that offload the 4-way handshake; brcmfmac never reports it, so authentication intermittently fails on Apple Silicon (BRCM4378/4387/4388). nixpkgs already carries the Broadcom revert in wpa_supplicant 2.12; the proper fix, Janne Grunau's `wifi: brcmfmac: cfg80211: Report port_authorized for 4-way HS offload` (2026-09), is still under review. Entering the key again after such a failure only creates a duplicate profile, so cancel and retry instead | Fix lands via a nixpkgs kernel update (or by carrying the patch through `nixpkgs-patched`) | 2026-10-07 | on every kernel bump | active |
|
||||
| WA-033 | USB-C/XHCI port `_PLD` ACPI noise — this board's DSDT, no code in this repo | every boot logs 16 `ACPI Error: Aborting method \_SB.PC00.TXHC.RHUB.SS01._PLD ... (AE_NOT_FOUND)` lines, plus `\_SB.PC02.XHCI.RHUB.HS01._PLD`; the ports work, and these messages are unrelated to the GPU's ATPX failure, which names `\_SB.PC00.GFX0.ATPX` instead | Drop if a BIOS update defines the `_PLD` descriptors | 2026-10-14 | on a BIOS update | active |
|
||||
| WA-034 | IPv6-first resolution is dead on the wired link — `hosts/idols-aquamarine/default.nix` (`networking.getaddrinfo.precedence`) | the proxy gateway answers DNS in fake-IP mode and its IPv6 path is dead: an AAAA lookup returns a fake address whose connection fails at once (measured: `curl -6` fails in 8 ms, `curl -4` answers in 0.7 s). glibc clients are covered by the RFC 6724 table with the IPv4-mapped row raised; nix's libcurl and other non-glibc resolvers ignore `/etc/gai.conf`. Disabling IPv6 with `boot.kernel.sysctl` did not work here — networkd re-enables it on the wired link for `DHCP = "ipv6"`/`IPv6AcceptRA` — so that setting was dropped on 2026-10-14 | Drop once the gateway serves proxied names over IPv6 as well (the adjacent fix is `dns.ipv6: false` in mihomo, see WA-030) | 2026-10-14 | on a mihomo or gateway change | active |
|
||||
| WA-035 | SSH retry + multiplexing to LAN peers, ARP fallback tuning — `home/base/tui/ssh.nix`, `hosts/idols-ai/neighbour-probe.nix` | the desktop reaches the LAN over a Mesh WiFi backhaul, which drops unicast to a peer whose MAC it has aged out of its forwarding table while still flooding broadcast: a capture shows seven SYNs from one socket addressed to the cached MAC with no answer, three unicast ARP probes unanswered, then a broadcast ARP answered in 0.4 ms after ~12 s of failed connects. The ssh settings retry a short connect and keep a master warm so the entry stays fresh; the neighbour-probe tunings walk the kernel to that broadcast fallback in ~1 s | Drop once the backhaul is wired or the mesh stops dropping unknown unicast | 2026-10-16 | when the backhaul changes | active |
|
||||
| WA-038 | OpenVINO GPU plugin needs an `ocl-icd` loader — [hosts/idols-aquamarine/hardware-intel-igpu.nix](./hosts/idols-aquamarine/hardware-intel-igpu.nix) (`openvino-gpu` wrapper) | `/run/opengl-driver/lib/libOpenCL.so.1` comes from ROCm's clr, which provides OpenCL 2.2 symbols, while OpenVINO's GPU plugin links the OpenCL 3.0 loader (`ocl-icd`): the plugin fails to load and OpenVINO silently omits GPU from `available_devices`. The wrapper runs a command with `ocl-icd` first on `LD_LIBRARY_PATH` and `OCL_ICD_VENDORS` pinned to the Intel ICD | Drop once nixpkgs' openvino GPU plugin carries an explicit `ocl-icd` RUNPATH, or ROCm's clr ships an OpenCL 3.0 loader | 2026-10-18 | on an openvino or ROCm bump | active |
|
||||
|
||||
## WIP
|
||||
|
||||
|
||||
@@ -0,0 +1,135 @@
|
||||
# The Intel iGPU on aquamarine
|
||||
|
||||
The Arrow Lake-S iGPU (i915) does this host's video work and can run OpenVINO models; the R9700s
|
||||
stay on inference. [hardware-intel-igpu.nix](./hardware-intel-igpu.nix) installs the stack and the
|
||||
`openvino-gpu` wrapper. Measured 2026-10-18, OpenVINO 2026.4.0 / openvino-genai 2026.4.0.0.
|
||||
|
||||
## Devices
|
||||
|
||||
`openvino-gpu` puts the OpenCL 3.0 loader (`ocl-icd`) ahead of ROCm's clr and pins the Intel ICD;
|
||||
without it the GPU plugin fails to load and GPU is missing from the device list (WA-038). Use it for
|
||||
every OpenVINO run:
|
||||
|
||||
```bash
|
||||
openvino-gpu python -c 'import openvino as ov; c = ov.Core(); print(c.available_devices)'
|
||||
# ['CPU', 'GPU', 'NPU']
|
||||
openvino-gpu python -c 'import openvino as ov; print(ov.Core().get_property("GPU", "FULL_DEVICE_NAME"))'
|
||||
# Intel(R) Graphics (iGPU)
|
||||
```
|
||||
|
||||
VA-API uses `/dev/dri/renderD128`:
|
||||
|
||||
```bash
|
||||
nix shell nixpkgs#libva-utils -c vainfo --display drm --device /dev/dri/renderD128
|
||||
nix shell nixpkgs#ffmpeg -c ffmpeg -hwaccel vaapi -hwaccel_device /dev/dri/renderD128 \
|
||||
-i in.mkv -c:v hevc_vaapi -qp 26 out.mkv
|
||||
```
|
||||
|
||||
## Running OpenVINO GenAI models
|
||||
|
||||
Models with static or bounded-dynamic shapes compile on `GPU`; unbounded dynamic shapes fail with
|
||||
`dynamic shape without upper bound`. Inspect a model before choosing the device:
|
||||
|
||||
```bash
|
||||
openvino-gpu python - /path/to/model.xml <<'PY'
|
||||
import sys
|
||||
import openvino as ov
|
||||
|
||||
for port in ov.Core().read_model(sys.argv[1]).inputs:
|
||||
print(port.get_any_name(), port.get_partial_shape())
|
||||
PY
|
||||
```
|
||||
|
||||
### Whisper needs a static encoder
|
||||
|
||||
The published `whisper-large-v3-turbo-int8-ov` IRs are fully dynamic. The NPU pipeline reshapes them
|
||||
internally and the CPU path runs them as-is, but the GPU path cannot compile the encoder. Reshape it
|
||||
once, with batch bounded `0..1` so OpenVINO GenAI's `{0, 128, 3000}` input reset still matches, then
|
||||
run the normal pipeline:
|
||||
|
||||
```bash
|
||||
# Store paths change on every rebuild, so read them from the running service.
|
||||
MODEL=$(systemctl cat whisper-asr | sed -n 's/^Environment="WHISPER_MODEL_DIR=\([^"]*\)"/\1/p')
|
||||
ENVPY=$(systemctl cat whisper-asr | sed -n 's/^ExecStart=\([^ ]*\).*/\1/p')
|
||||
OUT=/tmp/whisper-static
|
||||
|
||||
rm -rf "$OUT" && mkdir -p "$OUT"
|
||||
cp "$MODEL"/*.json "$OUT"/
|
||||
for f in "$MODEL"/*.bin "$MODEL"/*.xml; do
|
||||
case "$(basename "$f")" in
|
||||
openvino_encoder_model.*) ;;
|
||||
*) ln -s "$f" "$OUT/$(basename "$f")" ;;
|
||||
esac
|
||||
done
|
||||
|
||||
openvino-gpu "$ENVPY" - <<PY
|
||||
import openvino as ov
|
||||
|
||||
# 128 mel bins for large-v3/turbo, 3000 frames = 30 s window, batch bound 0..1.
|
||||
enc = ov.Core().read_model("$MODEL/openvino_encoder_model.xml")
|
||||
enc.reshape({"input_features": ov.PartialShape([ov.Dimension(0, 1), 128, 3000])})
|
||||
ov.save_model(enc, "$OUT/openvino_encoder_model.xml", compress_to_fp16=False)
|
||||
PY
|
||||
```
|
||||
|
||||
Transcribe (mono 16 kHz 16-bit WAV):
|
||||
|
||||
```bash
|
||||
openvino-gpu "$ENVPY" - "$OUT" /tmp/jfk.wav en <<'PY'
|
||||
import sys
|
||||
import wave
|
||||
|
||||
import numpy as np
|
||||
import openvino_genai
|
||||
|
||||
models, wav_path, language = sys.argv[1:4]
|
||||
with wave.open(wav_path) as wav:
|
||||
samples = (
|
||||
np.frombuffer(wav.readframes(wav.getnframes()), dtype=np.int16).astype(np.float32)
|
||||
/ 32768.0
|
||||
)
|
||||
pipe = openvino_genai.WhisperPipeline(models, "GPU", CACHE_DIR="/tmp/whisper-gpu-cache")
|
||||
config = pipe.get_generation_config()
|
||||
config.task = "transcribe"
|
||||
config.language = f"<|{language}|>"
|
||||
print(pipe.generate(samples, config).texts[0].strip())
|
||||
PY
|
||||
```
|
||||
|
||||
## Timings and memory
|
||||
|
||||
Steady state, pinned language, 3-5 iterations after warmup. NPU through the `whisper-asr` service;
|
||||
iGPU and CPU standalone. Audio: 11.0 s English, 5.6 s Chinese.
|
||||
|
||||
| Device | English | Chinese | CPU-s/request | Memory in use |
|
||||
| ------ | ------------------ | ------- | ------------- | -------------------- |
|
||||
| NPU | 1.47 s (1.46-1.57) | 1.36 s | 0.055 / 0.042 | ~2.2 GiB (resident) |
|
||||
| iGPU | 1.49 s | 1.41 s | 0.23 / 0.14 | ~2.1 GiB (on demand) |
|
||||
| CPU | 1.23 s | 1.18 s | 9.35 / 8.99 | ~1.1 GiB (on demand) |
|
||||
|
||||
| Device | Cold compile / first use | Peak memory |
|
||||
| ------ | ----------------------------------------------------------------------------------- | ----------- |
|
||||
| NPU | ~70 s VPUX compile on a cold `/var/cache/whisper-asr`; warm starts import the blobs | 6.2 GB |
|
||||
| iGPU | load 2.2 s instead of 0.7 s warm; first inference 1.8 s instead of 1.5 s | 2.65 GB |
|
||||
| CPU | load 0.95 s; first inference 1.85 s instead of 1.23 s | ~2.0 GB |
|
||||
|
||||
## Language switching
|
||||
|
||||
Pinned and auto-detect (no `language`) both work on the iGPU and CPU: alternating en/zh in one
|
||||
process or service is correct, 1.44-1.79 s on iGPU, 1.20-1.34 s on CPU. On the NPU pinned works
|
||||
(1.46/1.36 s); auto-detect works while the idle pipeline is fresh (>=4 s between requests), but an
|
||||
auto request right after a pinned one inherited the pinned language and translated the English audio
|
||||
to Chinese, correct again after the ~4 s rebuild. A pinned request during that rebuild took 5.25 s.
|
||||
Pin the language per client before mixing modes.
|
||||
|
||||
## Cache
|
||||
|
||||
A `CACHE_DIR` written by a failed GPU compile produced repeated text; deleting it and recompiling
|
||||
fixed the output. Wipe the cache after changing the model or after a failed attempt.
|
||||
|
||||
## Which device
|
||||
|
||||
- **Whisper/ASR: NPU.** Same latency as the iGPU, lowest CPU cost, leaves the iGPU free.
|
||||
- **iGPU: OpenVINO models that compile on `GPU`, and ASR fallback.** It occupies the GPU for the
|
||||
whole request, so transcodes and the console wait.
|
||||
- **CPU: emergency only.** Fastest single request, but it keeps ~7.5 cores busy.
|
||||
@@ -9,7 +9,7 @@ that name was decommissioned; this is a different, physical machine that reuses
|
||||
| Part | Model | Notes |
|
||||
| ------ | ----------------------------------------------------- | ------------------------------------------------------- |
|
||||
| Board | Colorful CVN Z890 ARK FROZEN V20 (BIOS 1013) | moved from `ai` on 2026-10-13, so its EFI vars come too |
|
||||
| CPU | Intel Core Ultra 7 270K Plus | Arrow Lake; NPU unused here |
|
||||
| CPU | Intel Core Ultra 7 270K Plus | Arrow Lake; the NPU serves `whisper-asr` (./asr.nix) |
|
||||
| Memory | 2x48G JUHOR (96G), 4800 MT/s | the 2x16G kit stayed with `ai` |
|
||||
| GPU | 2x AMD Radeon AI PRO R9700, 32G each (RDNA4, gfx1201) | 64G of VRAM total |
|
||||
| Root | Fanxiang S790 2TB | was `ai`'s `/data` disk until 2026-10-13 |
|
||||
@@ -26,9 +26,14 @@ Two things about this board set the ceiling for the GPU work:
|
||||
R9700 has no xGMI link between cards, and a dual-R9700 report has RCCL deadlocking with vLLM TP=2
|
||||
even with `NCCL_P2P_DISABLE=1`. See [ROCM.md](./ROCM.md).
|
||||
|
||||
The iGPU needs no configuration: the kernel binds `i915` and
|
||||
`hardware.enableRedistributableFirmware` is on, so a monitor plugged into the board's HDMI/DP ports
|
||||
gets the text console (`multi-user.target`, no display manager) when SSH is unreachable.
|
||||
The iGPU carries display and video work: the kernel binds `i915`, and
|
||||
[hardware-intel-igpu.nix](./hardware-intel-igpu.nix) installs `intel-media-driver` (VA-API),
|
||||
`vpl-gpu-rt` (oneVPL/QSV) and `intel-compute-runtime` (OpenCL/Level-Zero), so transcoding can use
|
||||
its media engine while the R9700s stay on inference. VA-API targets `/dev/dri/renderD128`, and
|
||||
OpenVINO jobs run through `openvino-gpu <command>`; [IGPU.md](./IGPU.md) records the wrapper, the
|
||||
static-shape workaround Whisper needs on `GPU`, and the measured NPU/iGPU/CPU timings. A monitor
|
||||
plugged into the board's HDMI/DP ports still gets the text console (`multi-user.target`, no display
|
||||
manager) when SSH is unreachable.
|
||||
|
||||
`iommu=pt` is set in [hardware-amd-gpu.nix](./hardware-amd-gpu.nix) because ROCm documents hangs on
|
||||
multi-GPU systems without IOMMU passthrough. The same file disables amdgpu runtime power management:
|
||||
|
||||
@@ -28,6 +28,7 @@ in
|
||||
# hardware
|
||||
./hardware-configuration.nix
|
||||
./hardware-amd-gpu.nix
|
||||
./hardware-intel-igpu.nix
|
||||
|
||||
./preservation.nix
|
||||
./secureboot.nix
|
||||
|
||||
@@ -0,0 +1,33 @@
|
||||
{
|
||||
pkgs,
|
||||
...
|
||||
}:
|
||||
|
||||
# ===============================================================================================
|
||||
# Intel Arrow Lake-S iGPU (i915). The R9700s are reserved for model inference, so the iGPU takes
|
||||
# video work and small OpenVINO jobs: iHD backs VA-API, oneVPL/QSV is Intel's video API, and
|
||||
# intel-compute-runtime provides the OpenCL/Level Zero devices OpenVINO's GPU plugin needs.
|
||||
# ===============================================================================================
|
||||
{
|
||||
hardware.graphics.extraPackages = with pkgs; [
|
||||
intel-media-driver
|
||||
vpl-gpu-rt
|
||||
intel-compute-runtime
|
||||
];
|
||||
|
||||
# OpenVINO's GPU plugin links the OpenCL 3.0 loader (ocl-icd), but
|
||||
# `/run/opengl-driver/lib/libOpenCL.so.1` comes from ROCm's clr (OpenCL 2.2), so the plugin
|
||||
# fails to load and OpenVINO silently drops GPU from `available_devices`. Run OpenVINO jobs
|
||||
# through this wrapper to put ocl-icd first and pin the Intel ICD (WORKAROUNDS.md, WA-038).
|
||||
# See IGPU.md for running models on the iGPU.
|
||||
environment.systemPackages = [
|
||||
(pkgs.writeShellApplication {
|
||||
name = "openvino-gpu";
|
||||
text = ''
|
||||
export LD_LIBRARY_PATH="${pkgs.ocl-icd}/lib''${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
|
||||
export OCL_ICD_VENDORS="${pkgs.intel-compute-runtime}/etc/OpenCL/vendors"
|
||||
exec "$@"
|
||||
'';
|
||||
})
|
||||
];
|
||||
}
|
||||
@@ -14,6 +14,8 @@
|
||||
stableKernel = true;
|
||||
btopShowsRadeon = true;
|
||||
btopReadsIgpu = true;
|
||||
igpuMediaStackInstalled = true;
|
||||
openvinoGpuWrapperInstalled = true;
|
||||
gpuMetricsExported = true;
|
||||
collectorFlagsUnique = true;
|
||||
inferenceUserCanRender = true;
|
||||
|
||||
@@ -58,6 +58,20 @@ in
|
||||
btopShowsRadeon = cfg.modules.btop.package == eval.pkgs.btop-rocm;
|
||||
# CAP_PERFMON inside the wrapper is what the i915 iGPU needs.
|
||||
btopReadsIgpu = cfg.modules.btop.perfmon;
|
||||
# The iGPU takes video work off the R9700s: VA-API, oneVPL and OpenCL/Level-Zero.
|
||||
igpuMediaStackInstalled =
|
||||
let
|
||||
packageNames = map (package: package.pname or "") (cfg.hardware.graphics.extraPackages or [ ]);
|
||||
in
|
||||
builtins.all (name: builtins.elem name packageNames) [
|
||||
"intel-media-driver"
|
||||
"vpl-gpu-rt"
|
||||
"intel-compute-runtime"
|
||||
];
|
||||
# OpenVINO's GPU plugin needs ocl-icd ahead of ROCm's clr loader (WA-038).
|
||||
openvinoGpuWrapperInstalled = builtins.any (
|
||||
package: (package.name or "") == "openvino-gpu"
|
||||
) cfg.environment.systemPackages;
|
||||
gpuMetricsExported = builtins.elem "drm" cfg.services.prometheus.exporters.node.enabledCollectors;
|
||||
# node_exporter rejects a repeated flag, and this list merges with the base
|
||||
# module's entries, so a duplicate restart-loops the unit.
|
||||
|
||||
Reference in new issue
Block a user