## VRAM Detection (model_management.py)
The DirectML code path had two hardcoded `1024 * 1024 * 1024 #TODO` values
in `get_total_memory()` and `get_free_memory()`, causing ComfyUI to report
only 1 GB of VRAM on any AMD/Intel GPU using the DirectML backend — regardless
of actual hardware. This forced NORMAL_VRAM or LOW_VRAM calculations to be
wildly wrong.
Fix for `get_total_memory`:
- On Windows, reads `HardwareInformation.qwMemorySize` from the GPU driver
registry key via `winreg`. This is the 64-bit accurate value (unlike
`Win32_VideoController.AdapterRAM` which overflows at 4 GB).
- Allows override via `COMFYUI_DIRECTML_VRAM_MB` env var.
- Falls back to 6 GB if registry query fails (safe default for modern dGPUs).
Fix for `get_free_memory`:
- Uses `torch_directml.gpu_memory(0)` to get per-tile usage fractions and
derives free memory as `total * (1 - max_usage_fraction)`.
## torchaudio: optional import on AMD/DirectML
torchaudio has a DLL incompatibility with torch-directml (which ships its own
torch runtime). The following files had bare `import torchaudio` at module
level, crashing ComfyUI startup entirely when torchaudio was absent:
- comfy/ldm/lightricks/vae/audio_vae.py
- comfy/audio_encoders/whisper.py
- comfy/audio_encoders/audio_encoders.py
- comfy_extras/nodes_audio.py
- comfy_extras/nodes_lt.py
- comfy_extras/nodes_wandancer.py
Each import is wrapped in `try/except (ImportError, OSError): torchaudio = None`,
matching the pattern already used in comfy/ldm/mmaudio/vae/autoencoder.py and
comfy/ldm/ace/vae/music_dcae_pipeline.py. Audio nodes will degrade gracefully
rather than preventing ComfyUI from starting.
Tested on: AMD Radeon RX 5600 XT (6 GB VRAM, gfx1010, Windows 10)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
For LTX Audio VAE, remove normalization of audio during MEL spectrogram creation.
This aligs inference with training and prevents loud audio from being attenuated.
