Commit History
10 days ago
-
Fix rapid-moss on ComfyUI with comfy-aimdo ("aimdo memory compile error")
New ComfyUI runs the MOSS-Image blocks inside comfy-aimdo's memory compiler, which records each block's allocations and replays them; a tensor allocated in one block and freed in a later one aborts it. The node returned a new residual tensor from every block (discussion #1). The native path now runs the whole MOSS-Image forward as a DIFFUSION_MODEL wrapper: the fp32 residual is allocated before the compiled region and updated in place like the native blocks, the adapter tensors reach the GPU before that region, the LoRA stays unmerged in bf16 (as the training sampler computes it), and the time path, modulation and residual run in fp32 from a float64 sinusoid. The HF-loader path is unchanged. Tested on ComfyUI 8f04fd6e (dynamic VRAM on) and 5cc4b870: no graph breaks or rogues, about 1.1 s per 1024² 4-step image on an RTX PRO 6000. READMEs add the measurements and whether stock nodes can replace the custom node. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>c88f33b
12 days ago
-
Fix rapid-moss compatibility with native ComfyUI MOSS models
Verified signatureSupport fused QKV projections and preserve fp32 modulation/residuals in native blocks. Add native UI/API workflows and regression coverage. Validated 1024x1024, 4-step generation with native and HF loaders; 8 tests pass.1ec327b -
README: cautions (required settings, recommended settings, known weaknesses, untested areas)
Verified signature8ad5e9b -
ComfyUI custom node (unmerged fp32 LoRA + fp32 time path + data-space sampler), example workflows, GPU check
Verified signatureeda6717