42 lines
4.0 KiB
Markdown
42 lines
4.0 KiB
Markdown
---
|
|
name: ryzenai-generic-failure-loader-order
|
|
description: "Debug AMD Ryzen AI OGA \"Exception during initialization: Generic Failure\" on Linux XDNA2 — the LD_LIBRARY_PATH-vs-activate ordering trap, its proof instruments, and the OGA 0.14 run contract."
|
|
---
|
|
|
|
Symptom (any `onnxruntime_genai` run on a Ryzen AI NPU, container or bare):
|
|
|
|
```
|
|
[E:onnxruntime:, inference_session.cc:2811 operator()] Exception during initialization: Generic Failure
|
|
```
|
|
|
|
Fails ~2 ms after `og.Model()`. No `openat` of `/dev/accel/accel0`, **nothing in the kernel log**, no XRT log (the INI logger is not built into AMD's shipped XRT), no core dump. It is NOT a platform, model-export, or driver-ABI failure.
|
|
|
|
## Root cause
|
|
|
|
`source <venv>/bin/activate` from `install_ryzen_ai.sh` **prepends** to `LD_LIBRARY_PATH`:
|
|
|
|
```
|
|
site-packages/flexml/flexml_extras/lib : site-packages/onnxruntime/capi : site-packages/voe/lib
|
|
```
|
|
|
|
`voe/lib` is the **compiler-side** bundle: it ships its own `libxrt_coreutil.so.2` and `libryzen_mm.so.1`. Set the path *before* activating and those copies win, XRT's XDNA shim binds against the wrong coreutil, `xrt_core::shim_loader::shim_loader()` → `get_userpf_device()` → `xrt::device(0)` throws, `ryzenai::Allocator::Alloc` throws next, and ORT flattens both into the bare string above.
|
|
|
|
**Rule: activate first, THEN set `LD_LIBRARY_PATH`, with `/opt/xilinx/xrt/lib` first.** Passing the var via `podman -e` and sourcing activate afterwards in the same `bash -lc` is the broken shape — that is exactly how a runner script differs from a hand-typed probe, and it is why diagnostics pass while the runner fails. Mounting XRT is not enough either: the tree must be **on** the path.
|
|
|
|
## Proof instruments, cheapest first
|
|
|
|
1. `ldd <deployment>/lib/libonnxruntime_providers_ryzenai.so | grep -E "libxrt_coreutil|libryzen_mm"` — the resolved path settles it alone: `/opt/xilinx/xrt/lib` = good, `.../voe/lib` = this bug.
|
|
2. `XILINX_XRT=/opt/xilinx/xrt /opt/xilinx/xrt/bin/xrt-smi examine` → `RyzenAI-npu5 | aie2p | 6x8` + `NPU Firmware Version`. Needs `libboost_filesystem1.83.0`, `libboost_program_options1.83.0`, `libuuid1` (26.04 names; the 24.04 docs' 1.74.0 does not exist). A "No such file or directory" for a file that exists = missing interpreter/dep, not a missing binary.
|
|
3. `strace -f -e trace=openat` — `/dev/accel/accel0` never opened ⇒ failure is pre-device ⇒ loader-level.
|
|
4. `gdb -q -batch` with `catch throw` + `bt 10` loop → names `shim_loader` / `Allocator::Alloc`.
|
|
|
|
Known dead ends (one run each, then stop): `--ulimit memlock`; XRT INI verbosity; `DD_LOG_LEVEL` / `RYZENAI_TRACE` / `RYZENAI_EP_TRACING` (print nothing before the throw); byte-scanning XRT `.so` for PCI IDs (ambiguous); libboost presence for the OGA path (only `xrt-smi` needs it); flag-by-flag container matrices (`--memory`, `--security-opt`, `--name`, extra mounts — all innocent).
|
|
|
|
## Verified facts on this stack
|
|
|
|
- Vendor XRT 2.25.37 userspace (`dpkg-deb -x`, never dpkg-installed) works against Ubuntu's in-tree `amdxdna` 0.7.0; no DKMS module needed. PCI `1022:17f0` rev 11 = XDNA2, reported as `aie2p 6x8`.
|
|
- AMD's `quicktest` passing does **not** prove NPU use (it can silently run on CPU); trust `xrt-smi` + logits, not "it didn't crash".
|
|
- `/dev/accel/accel0` is `root:render 0660`; rootless podman cannot read it (userns drops the gid; `keep-groups`, `keep-id:gid=990`, `--group-add render` all fail) → rootful container, or a udev rule like the GPU nodes'.
|
|
- OGA 0.14 (RAI 1.8) contract: `GeneratorParams` has only `set_search_options` (no `input_ids`, no `model.generate`); feed with `Generator.append_tokens(ids)`, loop `generate_next_token()`, read `get_sequence(0)`; `max_length` counts **prompt + generated**. Full Fusion (`*_npu_4K`) caps input+output at 4096.
|
|
- Validate content, not just exit code: `np.argsort(logits[0])[-8:]`. A real logits tensor after `append_tokens` proves the graph executed (prefill is folded into step 1, so a sub-millisecond first step + a 1-token sequence means it did not); if argmax is the sequence terminator (`
|