Fix intermittent "addmm_cuda not implemented for 'Char'" on partial unload (#76)

[Dawgmastah]

Summary

Fixes the intermittent "addmm_cuda" not implemented for 'Char' crash reported in #76, which persisted even after the load()-only cleanup in c0cc280.

Root cause

c0cc280 strips ComfyUI's stock LowVramPatch and swaps in the scale-aware INT8LowVramPatch, but only inside INT8ModelPatcher.load().

ComfyUI core also installs a stock LowVramPatch on offloaded INT8 layers inside partially_unload() (comfy/model_patcher.py), and that path does not route through load(). The stock patch's __call__ runs:

calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype)

For an INT8 weight the intermediate dtype is torch.int8, so the LoRA matmul hits addmm_cuda not implemented for 'Char'.

Because partially_unload() only fires under memory pressure — and only touches a memory-dependent subset of layers — the failure looked random and appeared to "fix itself" on reload. That matches the reports of the same LoRA working on some runs and failing on others.

Fix

• Extract the LowVramPatch → INT8LowVramPatch swap out of load() into a shared _reinstall_int8_lowvram_patches() helper (behavior unchanged — all three branches preserved).
• Add a partially_unload() override that calls super() then re-runs the helper, so the scale-aware patch is restored after every core path that can install the stock one.

Net change: +30/-1, contained entirely within INT8ModelPatcher.

Notes

I intentionally did not take the alternative "force intermediate_dtype to float32 for non-float weights" monkey-patch suggested in the issue. It would stop the traceback, but the stock LowVramPatch would then return a float weight with no re-quantization against the per-row scale — silently corrupting those layers instead of crashing.

🤖 Generated with Claude Code