Need help with AMD ROCm on Windows

[kotn3l]

Hi all!

First of all I'd like to say thanks to all contributors to this, the upscaling quality is truly outstanding. I'm using it mainly to upscale textures for high-res renders.

Anyways, I've been having mixed results on Windows 11 with my RX9070XT. When it works, the quality is amazing, but most of the time I'm unable to upscale some textures. I'm not sure if it's because of my environment, or because of poor AMD support? I'd like to ask for help in regards to this.

Total VRAM 16304 MB, total RAM 32694 MB
pytorch version: 2.10.0a0+rocm7.11.0a20251130
Set: torch.backends.cudnn.enabled = False for better AMD performance.
AMD arch: gfx1201
ROCm version: (7, 2)
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 9070 XT : native
Enabled pinned memory 14712.0

⚠️  SeedVR2 optimizations check: SageAttention ❌ | Flash Attention ❌ | Triton ✅
💡 Optional: pip install sageattention flash-attn
📊 Initial CUDA memory: 15.64GB free / 15.92GB total

This error I see very often:

[18:44:47.080]  ━━━━━━━━ Phase 2: DiT upscaling ━━━━━━━━
[18:44:53.233] 🚀 Materializing DiT weights to CUDA:0: J:\comfyui\comfy\ComfyUI\models\SEEDVR2\seedvr2_ema_3b_fp16.safetensors
[18:45:10.466] 🚀 Materializing DiT: 635 parameters, 6468.73MB total
[18:45:10.482] 🚀 DiT materialized directly from meta with loaded weights
[18:45:10.488] ✅ Initialized 64 non-persistent buffers
[18:45:10.489] 🔧 Applying DiT compatibility wrapper
[18:45:10.489] 🔧 Stabilizing RoPE computations for numerical stability
[18:45:10.490] ✅ Stabilized 64 RoPE modules
[18:45:10.490] 🔧 Applying sdpa attention mode and torch.bfloat16 compute dtype to model
[18:45:10.491] ✅ Applied sdpa and compute_dtype=torch.bfloat16 to 32 modules
[18:45:10.492] 🎯 Model precision: DiT=torch.float16, VAE=torch.bfloat16, compute=torch.bfloat16
[18:45:10.492] 🔄 DiT already on CUDA:0, skipping movement
[18:45:10.493] 📊 After DiT loading for upscaling:
[18:45:10.493] 📊   [VRAM] 7.03GB allocated / 7.07GB reserved / Peak: 7.03GB / 8.51GB free / 15.92GB total
[18:45:10.493] 📊   [RAM] 1.20GB process / 12.06GB others / 18.66GB free / 31.93GB total
[18:45:10.494] 📊   Memory changes: VRAM +6.38GB, RAM -1.35GB
[18:45:10.494] 📊 Resetting VRAM peak memory statistics
[18:45:10.494] 🎬 Upscaling batch 1/1
[18:45:10.496] 🚀 Using seed: 3848114342 for deterministic generation
EulerSampler:   0%|                                                                                                                                                                       | 0/1 [00:00<?, ?it/s][18:45:37.965] ❌ [ERROR] Forward pass error: unflatten: Provided sizes [158] don't multiply up to the size of dim 0 (58) in the input tensor
[18:45:37.966] ℹ️ torch.float16 model - no conversion applied
[18:45:37.966] ❌ [ERROR] Error in Phase 2 (Upscaling): unflatten: Provided sizes [158] don't multiply up to the size of dim 0 (58) in the input tensor
[18:45:37.967] 🧹 Cleaning up DiT components
[18:45:37.968] 🔄 Moving DiT from CUDA:0 to CPU (releasing GPU memory)
[18:45:39.448] 🧹 DiT model deleted
[18:45:39.448] 🧹 Cleaned up text embeddings: texts_pos, texts_neg
[18:45:39.449] 🧹 Starting full cleanup
[18:45:39.449] 🧹 Cleaning up VAE components
[18:45:39.449] 🔄 Moving VAE from CUDA:0 to CPU (releasing GPU memory)
[18:45:39.621] 🧹 VAE model deleted
[18:45:39.622] 🧹 Clearing memory caches (deep)...
[18:45:40.347] ✅ Completed full cleanup
!!! Exception during processing !!! unflatten: Provided sizes [158] don't multiply up to the size of dim 0 (58) in the input tensor

My workflow:

Basically I'm doing a 4x upscale, but limiting max resolution based resolution -- long story short, sometimes the upscale works, sometimes it doesn't, I can't really figure out why. Am I missing some dependencies / AMD optimizations? Any help is appreciated, thanks!

[urixl]

Maybe this string is the answer:
Provided sizes [158] don't multiply up to the size of dim 0 (58)
Seek the part where you set something to 158, or a result of some settings that leads to 158.

[kotn3l]

Yeah, that is indeed the error message, but that comes from SeedVR2, not something I set. It's something internally in the model.

[urixl]

Does it happen with some files that have specific properties, like width, height, alpha channel?

[kotn3l]

Yes, I mean it does happen with some texture inputs, but nothing specific like alpha channel etc. But if I wanna upscale the same texture that has failed before, it will most likely fail again.

[adrientoupet]

Can you please share the full debug log and upload the failing texture so we can troubleshoot further? Thanks

[kotn3l]

Can you please share the full debug log and upload the failing texture so we can troubleshoot further? Thanks

Yes of course! Thank you for looking into this.

seedvr2_log.txt