Strix Halo (gfx1151) - 3x speedup on Hunyuan Video with 7.10 + HSA_USE_SVM=0 fix

[bkpaine1]

Hardware

• AMD Ryzen AI MAX+ 395 (Strix Halo)
• 128GB unified memory
• gfx1151

Software

• TheROCk 7.10 nightly
• PyTorch 2.9.1+rocm7.10.0
• ComfyUI + HunyuanVideo 1.5

The Problem

VAE decode on Hunyuan Video would trigger SVM thrashing, leading to GPU ring timeout and MODE2 reset:

svm_range_restore_work hogged CPU... 4099 times
ring gfx_0.0.0 timeout
VRAM is lost due to GPU reset!

The Fix

export HSA_USE_SVM=0

This disables Shared Virtual Memory management that was fighting with the 128GB unified memory architecture.

Results

Metric	Before (7.9)	After (7.10 + fix)
Sampling	40s/it	10-20s/it
VAE Decode	CRASH	<1 min stable
Per-clip time	15-18 min	5-8 min
Full video (20 clips)	5+ hours	~70 min

Full Launch Script

#!/bin/bash
# STRIX HALO (gfx1151) | TheROCk 7.10
unset HSA_OVERRIDE_GFX_VERSION
export HIP_VISIBLE_DEVICES=0
export HSA_ENABLE_SDMA=0
export HSA_USE_SVM=0  # THE FIX
export PYTORCH_ALLOC_CONF="backend:native,expandable_segments:True,garbage_collection_threshold:0.9,max_split_size_mb:512"
export TORCHINDUCTOR_CACHE_DIR="$HOME/.cache/torch_inductor_710"
export TORCHINDUCTOR_FX_GRAPH_CACHE=1

python main.py --listen 0.0.0.0 --port 8189 --highvram --disable-smart-memory

Notes 25 frames 480X832 (tiktok) previous on 7.9 was 48s/it

• Inductor autotuning now works on gfx1151 - first run benchmarks kernels, subsequent runs use cache
• bf16 VAE had tensor mismatches, fp16 works fine
• Tiled VAE (tile_size=256) for safety, but non-tiled may work now

model weight dtype torch.float16, manual cast: None

model_type FLOW
Requested to load HunyuanVideo15
loaded completely; 95367431640625005117571072.00 MB usable, 15881.76 MB loaded, full load: True
93%|█████████████████████████████████████████████████████████████████████████████████████████████████ | 14/15 [05:25<00:24, 24.18s/it]

Credit

Debugged this with Claude (Anthropic) over multiple sessions analyzing dmesg logs and testing configurations. The SVM thrashing pattern in the logs led us to try HSA_USE_SVM=0.

---

cc @lamikr (if that's the right handle for TheROCk maintainer)