MIGraphX EP places almost none of a conformer encoder (parakeet-tdt-0.6b-v3 ONNX) - CPU-speed inference

Environment

• onnxruntime 1.24.4 ROCm build (Arch python-onnxruntime-opt-rocm), host ROCm 7.2.4
• GPU: gfx1100 (RX 7900 XTX)
• Model: NVIDIA parakeet-tdt-0.6b-v3 ONNX export (encoder-model.onnx, conformer
  with relative-position attention; the export istupakov/onnx-asr consumes)

Symptom

Session creation with ["MIGraphXExecutionProvider", "CPUExecutionProvider"]
"succeeds": a 2–4 s "Model Compile" (far too fast for a 0.6B encoder), then
inference runs at CPU speed with ~7 % GPU utilization and near-zero extra VRAM.
The encoder is evidently partitioned almost entirely to the CPU EP. Fixed-shape
add_free_dimension_override_by_name for all dynamic axes changes nothing, so
this is op support (likely the relative-position attention subgraph), not
dynamic shapes.

Ask

1. Which ops/subgraph patterns in this encoder are unsupported? (Happy to run
   any diagnostic build/env-var and attach logs.)
2. Conformer-style ASR encoders are a major model family — coverage here would
   make MIGraphX viable for a wide class of speech models.

Repro script + model layout available; the model is a public HF download.

Found while batch-transcribing ~650h of audio with parakeet-tdt-0.6b-v3 on an AMD gfx1100 workstation.