FastH3 Omni (INT8 ConvRot) on AMD Strix Halo, Windows 11 + ROCm 10.1: setup, timings, gotchas

#5
by ShamanFlamingo - opened

FastH3 Omni runs fine on AMD with these ComfyUI weights. Here is a full data point in case it helps other AMD users (and maybe a future ROCm row in the cookbook πŸ™‚).

Machine

  • GMKtec EVO-X2: Ryzen AI Max+ 395, Radeon 8060S iGPU (gfx1151), 128 GB LPDDR5X unified memory; PyTorch sees about 110 GB as VRAM
  • Native Windows 11 (no WSL), Python 3.12.11
  • torch 2.13.0+rocm10.1.0, ComfyUI 0.42.4, comfy-kitchen 0.2.39
  • Launch flags: --disable-pinned-memory --use-split-cross-attention --bf16-vae (Dynamic VRAM on)

Models

  • Diffusion: fasth3_omniref_8step_pruned_int8_convrot.safetensors (this repo)
  • Text encoder: qwen3vl_32b_minimax_h3_int8_convrot.safetensors (NVFP4 AWQ is Blackwell-only)
  • VAE: minimax_h3_video_vae_fp16 + minimax_h3_audio_vae_fp32
  • Graph: native "MiniMax H3 Reference to Video" node, MiniMaxH3SigmaShift 12 / 3, ModelAttentionBackend = comfy kitchen attention, BlockSparseAttention (vsa, keep 10 %, start 0.2, min_tokens 12288), res_multistep / simple, 8 steps, 24 fps

Timings (end to end, request β†’ MP4 with audio)

Same task as Hao AI Lab's local comparison (1 image, 832Γ—480, 124 frames, 8 steps)

  • Warm, 3 seeds: 332.5 s mean (333.1 / 332.2 / 332.3 s), sampling 34.5 s/step (276 s)
  • Cold (first run after model swap): 372 s
  • For reference, from Hao AI Lab's post: 2Γ— DGX Spark 130 s Β· 1Γ— Spark 192 s Β· RTX 5090 62 s Β· RTX 3090 FP8 16.9 min Β· M5 Max 19.5 min

Other single runs (indicative, not averaged)

Refs Resolution Γ— frames Time
1 image 1344Γ—768 Γ— 124 (5 s) 922 s
2 images 832Γ—480 Γ— 124 465 s
1 voice (no image) 832Γ—480 Γ— 124 321 s
1 image + 1 reference video 832Γ—480 Γ— 124 1162 s
5 images, euler/simple 1024Γ—576 Γ— 362 (15 s) 2359 s (249 s/step)

The iGPU's strength here is memory: the TE (β‰ˆ26 GB), the DiT (β‰ˆ21 GB) and the VAEs all stay resident, with no offload, and 10 s at 768p ran without OOM.

Gotchas on gfx1151 / ROCm 10.1 (Windows)

  1. SDPA Flash / Efficient kernels have no valid image for gfx1151 in torch 2.13+rocm10.1 ("device kernel image is invalid"). This only hits direct SDPA calls, i.e. the audio VAE encoding voice/video references. Fix: force the MATH backend (torch.backends.cuda.enable_flash_sdp(False), enable_mem_efficient_sdp(False), plus ComfyUI's SDPA_BACKEND_PRIORITY = [MATH]). With that, the voice-reference run went from crash to 321 s.
  2. No Triton wheel for ROCm on Windows, so anything Triton-based is unavailable.
  3. PYTORCH_*_ALLOC_CONF=expandable_segments:True causes OOM on Windows/HIP: leave it unset.
  4. comfy kitchen attention is much faster than PyTorch attention here (about βˆ’60 % per step).
  5. Heavy CPU load on the same chip throttles the iGPU (down to about 1.4 GHz): keep the CPU idle while benchmarking.
  6. ComfyUI caches identical graphs: change the seed between benchmark runs.

Happy to run specific prompts or reference mixes on this box if useful. Next step on my side: trying FastVideo itself natively on ROCm.

Thanks a lot for sharing this. Did you try running this workflow with additional LoRa's like character replacement LoRa?

Sign up or log in to comment