softab

Pyannote Benchmarks on AMD Strix Halo (gfx1151)

Last Updated: 2026-02-02 Hardware: AMD Ryzen AI Max+ 395, Radeon 8060S (40 CU, gfx1151), 128GB unified memory Test Audio: 11s JFK speech (test-audio.wav) Model: pyannote/speaker-diarization-3.1

Executive Summary

• Best config: ROCm 6.2 with gfx1100 fallback (~2.5-2.7s for 11s audio)
• ROCm 7.2 / TheRock images have dependency issues (lightning import errors)
• Requires HF_TOKEN for model download (gated model)
• 4x realtime speaker diarization on GPU

Benchmark Results (11s audio → diarization time)

Image	PyTorch	GFX	Time	Realtime	Status
pyannote-rocm62-gfx1151	2.5.1+rocm6.2	gfx1100	2.54s	4.3x	✅ PASS
pyannote-rocm62-working-gfx1151	2.5.1+rocm6.2	gfx1100	2.59s	4.2x	✅ PASS
pyannote-rocm72-amd-gfx1151	-	-	-	-	❌ Import error
pyannote-rocm644-gfx1151	-	-	-	-	❌ Import error
pyannote-therock-gfx1151	-	-	-	-	❌ Import error

Realtime = audio_duration / processing_time (higher = faster)

Key Findings

1. Only ROCm 6.2 Images Work

• ROCm 7.2, 6.4.4, and TheRock images fail with lightning.pytorch import errors
• Likely dependency version mismatch with pyannote-audio

2. Uses gfx1100 Fallback

• Working images use gfx1100 (not native gfx1151)
• hipBLASLt automatically disabled (“unsupported architecture”)
• Still achieves good performance via fallback

3. Performance Characteristics

• ~2.5s for 11s audio = 4.3x realtime
• Consistent across runs (low variance)
• GPU memory usage: ~2-3GB for speaker-diarization-3.1

Image Compatibility

Status	Images
✅ Working	pyannote-rocm62-gfx1151, pyannote-rocm62-working-gfx1151
❌ Import Error	pyannote-rocm72-, pyannote-rocm644-, pyannote-therock-*
❓ Untested	pyannote-rocm62-minimal-gfx1151

Error Details (ROCm 7.x images)

ImportError: cannot import name 'is_oom_error' from 'lightning.pytorch.utilities.memory'

This is a pyannote-audio / pytorch-lightning version mismatch.

Prerequisites

HuggingFace Token

Pyannote models are gated - requires accepted license and token:

1. Accept license at https://huggingface.co/pyannote/speaker-diarization-3.1
2. Get token from https://huggingface.co/settings/tokens
3. Pass via HF_TOKEN environment variable

Monkeypatch for Token Handling

Older pyannote uses use_auth_token, newer huggingface_hub uses token:

import huggingface_hub
_orig = huggingface_hub.hf_hub_download
def _patched(*args, **kwargs):
    if 'use_auth_token' in kwargs:
        kwargs['token'] = kwargs.pop('use_auth_token')
    return _orig(*args, **kwargs)
huggingface_hub.hf_hub_download = _patched

Running Benchmarks

# Using benchmark script
export HF_TOKEN="your_token_here"
./benchmarks/pyannote-simple.sh softab:pyannote-rocm62-gfx1151 /data/models/test-audio.wav

# Direct run
podman run --rm \
  --device=/dev/kfd --device=/dev/dri \
  --ipc=host \
  --security-opt seccomp=unconfined \
  --security-opt label=disable \
  -e HF_TOKEN="$HF_TOKEN" \
  -v /data/models:/audio:ro \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  softab:pyannote-rocm62-gfx1151 \
  python3 -c "
import torch
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained('pyannote/speaker-diarization-3.1')
pipeline.to(torch.device('cuda'))
diarization = pipeline('/audio/test-audio.wav')
for turn, _, speaker in diarization.itertracks(yield_label=True):
    print(f'[{turn.start:.1f}s - {turn.end:.1f}s] {speaker}')
"

Comparison with Other Workloads

Workload	Best Backend	Realtime Factor	Notes
Whisper (11s)	ROCm 7.2 HIP	24.6x	Transcription
Pyannote (11s)	ROCm 6.2	4.3x	Diarization
PyTorch ViT	TheRock 7.11	-	725 img/s

Pyannote is slower than Whisper because diarization involves:

• Multiple neural network passes (segmentation, embedding, clustering)
• Speaker embedding extraction for each segment
• Clustering algorithm on embeddings

Known Issues

1. ROCm 7.x dependency mismatch - pyannote-audio incompatible with newer lightning
2. Requires monkeypatch - Token parameter name changed in huggingface_hub
3. TF32 disabled warning - pyannote disables TF32 for reproducibility
4. hipBLASLt warning - Automatically falls back to hipblas

Recommended Setup

For speaker diarization on Strix Halo:

# Use the working ROCm 6.2 image
softab:pyannote-rocm62-gfx1151

# Or the explicitly named "working" variant
softab:pyannote-rocm62-working-gfx1151

---

See also:

• Whisper Benchmarks - Speech-to-text comparison
• PyTorch Benchmarks - General PyTorch performance