WaveNet.forward() got unexpected keyword argument 'cond' at infer_speedup=1

[MIM18-code]

Bug

When inference is run with infer_speedup=1 (i.e. the full 100-step DDPM path, not the dpm-solver speedup path), the conversion crashes inside the denoising loop:

ERROR: WaveNet.forward() got an unexpected keyword argument 'cond'

The shipped config's default infer_speedup: 10 works fine (10 steps via dpm-solver). infer_speedup: 2 (50 steps via dpm-solver) also works. Only the no-speedup full-step path fails.

Repro

cd ~/HQ-SVC
# Use any source + reference singing wav.
# In configs/hq_svc_infer.yaml, set:
#   infer_speedup: 1
# Then launch gradio_app.py and run any conversion. The diffusion
# loop starts ("sample time step: 0/100") then crashes immediately.

Tested on macOS 15 / Apple Silicon (MPS), PyTorch 2.5, with both infer_method: 'dpm-solver' and infer_method: 'ddim' set in the config — same error in both cases. The dpm-solver/ddim branches in utils/models/diffusion.py are gated by if method is not None and infer_speedup > 1:, so when infer_speedup == 1, control drops into the trailing else: block (around line ~377):

else:
    if use_tqdm:
        for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t):
            x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
    else:
        for i in reversed(range(0, t)):
            x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)

The crash is downstream of self.p_sample(..., cond) — cond is passed positionally here but appears to end up as a keyword somewhere inside, where WaveNet.forward doesn't accept it.

Why this matters

Some inference scenarios benefit from more diffusion steps than the dpm-solver 10/50-step regime, especially when:

• Conditioning audio is noisier than the LibriTTS-style training distribution
• The user wants to A/B test paper-quality numbers (the paper benchmarks at 100 base steps)
• Quality > speed (offline / batch use cases)

Right now the 100-step path is unreachable.

Suggested fix

Likely a one-liner in the no-speedup else: block: change self.p_sample(x, t, cond) to match whatever p_sample's implementation expects (e.g. cond=cond if the kwarg flows through, or remove the cond plumbing if p_sample doesn't need it). Happy to send a PR once I understand the intended call site.

Environment

• macOS 15 / Apple Silicon (MPS)
• PyTorch 2.5.1, transformers 4.46
• HQ-SVC commit 853a188 (current main)
• weights: shawnpi/HQ-SVC (HuggingFace)