[Bug][Linux] /v1/embeddings always fails with qds_device::wait() unexpected command state (Krackan Point, FLM 1.0.6)

[ripiox]

Problem Description

On Linux, every request to /v1/embeddings fails with the same internal XRT error, deterministically, while chat completions on the same server instance keep working fine.

Request:

curl http://127.0.0.1:52625/v1/embeddings \
  -d '{"model":"embed-gemma:300m","input":"hola"}'

Response (100% of requests):

{"error":"qds_device::wait() unexpected command state"}

Server log shows the request arriving and being accepted, then the error:

[LOG]  Target: /v1/embeddings
    "model": "embed-gemma:300m"
Embedding input[0]: hola
→ {"error":"qds_device::wait() unexpected command state"}

Notes:

• Happens with single-string input and with arrays.
• flm serve llama3.2:1b --embed 1 loads both models successfully (Embedding mode enabled: reserving additional 300MB); failure only occurs at the actual embed dispatch.
• Without --embed 1, chat on the same box is fine (96 tok/s prefill / 42 tok/s decode with llama3.2:1b).
• The error message text matches the generic XRT/NPU queue-state error also seen in std::runtime_error -- what(): qds_device::wait() unexpected command state Xilinx/mlir-aie#1751 and run_llm failed in ryzenai 1.4.0 (linux) amd/RyzenAI-SW#177.

Related issues (but not the same symptom): #732 and #661 report embeddings that complete but are wrong/collapsed on Linux; #647 reports non-determinism; #655 reports an IOCTL regression on the same chip family (Krackan Point). This report adds the case where the embed dispatch itself errors out on every request.

Environment

Item	Value
Hardware	AMD Ryzen AI 5 PRO 340 (Krackan Point), PCI 1022:17f0, Lenovo subsystem
OS	Omarchy (Arch Linux), Wayland
Kernel	7.2.5-4-omarchy
NPU FW	1.1.2.64
Device	/dev/accel/accel0 with 8 columns
amdxdna	in-tree, reported as 0.10 by flm validate
XRT	xrt 2.21.75-14 + xrt-plugin-amdxdna 2.21.75-2 (Arch extra)
FLM	fastflowlm 1.0.6-1 (Arch extra)
Memlock	unlimited (via ulimit wrapper)

flm validate output:

[Linux]  Kernel: 7.2.5-4-omarchy
[Linux]  NPU: /dev/accel/accel0 with 8 columns
[Linux]  NPU FW Version: 1.1.2.64
[Linux]  amdxdna version: 0.10
[Linux]  Memlock Limit: infinity

Steps to reproduce

flm pull llama3.2:1b
flm pull embed-gemma:300m
flm serve llama3.2:1b --embed 1

curl http://127.0.0.1:52625/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model":"embed-gemma:300m","input":"hola"}'

Expected

A vector array response:

{"data":[{"embedding":[...],...}]}

Actual

{"error":"qds_device::wait() unexpected command state"}

CPU governor/pmode (--pmode balanced) does not change the outcome; also reproduced with pmode default. Model files pass flm check after a fresh verified download.