XL Support?

[ZenRepublic]

yo yo! XL just dropped and comfy is the neatest way to train. is it supported?

[ZenRepublic]

error: [FL AceStep] Error processing sample sample_0202: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same
Error processing sample sample_0202: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same

[ZenRepublic]

Pasting from my conversation with Claude, sorry for the slop:

Preprocessing fails on ACE-Step 1.5-XL: device mismatch + async CUDA corruption
Environment

ComfyUI (recent), ACE-Step 1.5-XL checkpoint
RTX 5090, torch 2.12.0.dev+cu128 (nightly), Python 3.12
Windows 11
Symptoms (in order, as we worked through them)

torchcodec missing — TorchCodec is required for load_with_torchcodec. Fixed by installing torchcodec into the Comfy venv. Might be worth adding to requirements.txt.
Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same during VAE encode, on every sample.
After patching (2): CUDA error: invalid argument on every sample, surfaced from inside the Qwen3 LM autoregressive loop at comfy/sd1_clip.py → process_tokens → torch.tensor([[token]], device=device, dtype=torch.long).
Root causes

(2) With the larger 1.5-XL footprint, ComfyUI's load_models_gpu([vae.patcher], force_full_load=False) doesn't fully resident the VAE — vae.first_stage_model keeps weights on CPU while the nodepack's vae_encode_direct bypass moves only the input tensor to CUDA.
(3) ComfyUI logs Model ACE15TEModel_ prepared for dynamic VRAM loading. 9126MB Staged. ... Force pre-loaded 258 weights: 511 KB. The text encoder is streamed CPU→GPU on demand via runtime patch hooks, and those hooks race with the autoregressive LM kernels, poisoning CUDA state — surfaces as a generic "invalid argument" on the next trivial on-device tensor creation. Confirmed with CUDA_LAUNCH_BLOCKING=1.
Fix (code)

Two small patches to the nodepack:

modules/audio_utils.py — in vae_encode_direct, force the VAE onto the target device before the call (the outer load_models_gpu alone is insufficient under XL's memory pressure):

def vae_encode_direct(vae_model, audio_tensor, device, dtype):
vae_model.to(device=device, dtype=dtype) # <-- added
audio_in = audio_tensor.to(device=device, dtype=dtype)
latents = vae_model.encode(audio_in)
return latents.transpose(1, 2).float()
nodes/dataset_preprocess.py — the except Exception as e in the preprocessing loop currently swallows the traceback and only logs str(e). Including traceback.format_exc() is what let us find the real source of the second bug; recommend keeping it:

import traceback
...
except Exception as e:
tb = traceback.format_exc()
error_msg = f"Error processing sample {sample.id}: {str(e)}\n{tb}"
logger.warning(error_msg)
errors.append(error_msg)
Fix (launch flags)

Code patches alone are not enough — ComfyUI must be told not to stream the text encoder:

set CUDA_LAUNCH_BLOCKING=1
venv\Scripts\python.exe main.py --highvram
--highvram disables dynamic VRAM loading and keeps ACE15TEModel_ fully resident. With this, preprocessing runs cleanly sample-by-sample on a 5090. Might be worth a note in the README that ACE-Step 1.5-XL requires --highvram (or equivalently, that dynamic-loaded text encoders are not compatible with the autoregressive sampling path in this pipeline).

Caveat

Only the preprocessing stage is verified so far — training hasn't been tested yet. Will update if anything surfaces there.

[ZenRepublic]

Lora trained well! the only problem was the post processing node.

btw this part in above message: "torchcodec missing — TorchCodec is required for load_with_torchcodec. Fixed by installing torchcodec into the Comfy venv. Might be worth adding to requirements.txt." is irrelevant, i had just forgotten to install the requirements.txt before starting, so claude remembered it as an issue