[Bug] MiniMaxH3: DynamicVRAM causes infinite hang at "Model Initializing..." on RTX 4070 12GB; custom nodes surface underlying CUDA illegal memory access

[Milor123]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

MiniMaxH3 should begin sampling and generate video when loaded with DynamicVRAM enabled on a 12GB NVIDIA GPU.

Actual Behavior

The process hangs indefinitely at 0%| 0/20 [00:00<?, ?it/s, Model Initializing ... ] and never progresses. No CPU or GPU activity is observed after this point.

When custom nodes (e.g., ComfyUI-MultiGPU) are loaded, the same hang manifests instead as an immediate CUDA error: an illegal memory access was encountered during stream synchronization (prefetch_queue_pop → ensure_offload_stream → wait_stream/record_event), suggesting the custom nodes' CUDA synchronization calls surface an already-corrupt context.

Steps to Reproduce

1. Start ComfyUI with default flags (--lowvram or no memory flags).
2. Load any MiniMaxH3 workflow (tested with int8_convrot and NVFP4 formats).
3. Queue the prompt.
4. Observe: model stages successfully (~11944MB staged), sampling starts, then hangs forever at step 0/20.

Environment

Component	Version
OS	Windows 11
ComfyUI	0.33.0
comfy-aimdo	0.4.13
comfy-kitchen	0.2.31
PyTorch	2.10.0+cu130
Python	3.13.5
GPU	NVIDIA GeForce RTX 4070 (12 GB VRAM)
Driver	610.88
System RAM	48 GB (48445 MB)
CPU	Intel Core i5-14500
Installation	Manual (git clone + uv venv)

Debug Logs

Without custom nodes (--disable-all-custom-nodes):

[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 11956MB Staged. 0 patches attached. Force pre-loaded 210 weights: 1175 KB.
  0%| 0/20 [00:00<?, ?it/s,   Model Initializing ...  ]

Process hangs indefinitely here. Tested for 5+ minutes with zero progress.

With custom nodes (NVFP4 format):
Crash occurs at comfy/model_prefetch.py:117 during prefetch_queue_pop:

File "comfy\model_prefetch.py", line 117, in prefetch_queue_pop
    offload_stream, fully_faulted = comfy.ops.cast_modules_with_vbar(...)
File "comfy\ops.py", line 138, in ensure_offload_stream
    offload_stream = comfy.model_management.get_offload_stream(device)
File "comfy\model_management.py", line 1470, in get_offload_stream
    ss[stream_counter].wait_stream(current_stream(device))
torch.AcceleratorError: CUDA error: an illegal memory access was encountered

With custom nodes (int8_convrot format):
Crash occurs inside comfy_kitchen quantized kernel:

File "comfy_kitchen\tensor\w4a8_int8.py", line 119, in w4a8_int8_linear
    return impl(**kwargs)
File "comfy_kitchen\backends\cuda\__init__.py", line 2157, in w4a8_int8_linear
    out = torch.empty(m, n, dtype=out_dtype, device=x.device)
torch.AcceleratorError: CUDA error: an illegal memory access was encountered

Is the issue caused by a custom node?

• I have confirmed that the issue is NOT caused by a custom node. I tested with --disable-all-custom-nodes and the hang persists (it simply becomes silent instead of crashing).

Workarounds Attempted

Workaround	Result
--disable-all-custom-nodes	Hangs silently at step 0/20 (no crash, no progress)
--disable-async-offload	Hangs identically
CUDA_LAUNCH_BLOCKING=1	Hangs identically
--disable-pinned-memory	Hangs/crashes identically
--vram-headroom 1	Hangs/crashes identically
--disable-dynamic-vram	OOM (12GB insufficient for this model)
Different model format (int8_convrot → NVFP4)	Same failure mode

Additional Context

The workflow is the (1 mp), 15 duration:

• Two failure modes, same root cause: Without custom nodes = silent hang. With custom nodes = cudaErrorIllegalAddress crash. The custom nodes are not the cause; they perform soft_empty_cache / mem_get_info calls that surface an already-corrupt CUDA context.
• This appears to be the same underlying issue as 0.00 MB usable + hangs at Model Initializing with Flux GGUF on 8GB NVIDIA GPU (lowvram + comfy-aimdo) #14573 (hang at Model Initializing with DynamicVRAM on low-VRAM NVIDIA), MiniMax H3 reproducibly causes GPU is lost / TDR black-screen on RTX 5070 Ti when the system has 64 GB RAM; identical workload is stable when Windows is capped to 32 GB #15488 (GPU lost / illegal memory access with MiniMax H3 on high-RAM Windows systems), and MiniMax H3 crashes after Clean/Unload on Windows with AIMDO DynamicVRAM #15352 (AIMDO crash post-unload with MiniMax H3).
• The fact that CUDA_LAUNCH_BLOCKING=1 does not change the behavior rules out a simple async kernel race condition. The corruption appears to happen during the initial VBAR/stream setup or weight prefetch phase before any kernel even executes.

[rattus128]

We have an ongoing issue of corrupted cuda contexts. Do you by any change have multiple Nvidia GPUs in the system? (even of you are only use one?)

[Milor123]

We have an ongoing issue of corrupted cuda contexts. Do you by any change have multiple Nvidia GPUs in the system? (even of you are only use one?)

Not, I have my iGPU intel as main, and my dGPU rtx 4070 only, no more devices, and with LTX 2.5 I can generate videos of 2.0 MP very good.