mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-06-26 17:59:54 +08:00
70724ddf43
On AMD ROCm (gfx11xx APUs/GPUs), the PiD PidNet forward in fp16/bf16 can return an all-NaN tensor through the ROCm/AOTriton attention path, which then corrupts the decoded image. The default (non-AOTriton) path stays clean, so the bad values come from an AOTriton attention miscompilation on gfx11xx rather than the PiD math itself (see ROCm/triton#909 and ROCm/aotriton#179). Guard the PidNet output on AMD: when it is fp16/bf16 and actually contains NaN/Inf, log a one-time warning that points at --use-split-cross-attention and clamp the values with nan_to_num before decode (the same pattern already used for flux/lumina). Non-AMD devices and fp32 paths are unaffected; finite outputs only pay a single isfinite() check. Fixes #14249 Signed-off-by: liminfei-amd <91481003+liminfei-amd@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| model.py | ||
| modules.py | ||
| pid.py | ||