Fix joyfox runtime error: skip its AdaLN deltas

The Comfy-Org checkpoint joyfox was trained against gives adaln_proj an
8-dim input ([96768, 8]); the diffusers transformer projects the full
2688-dim time embedding, so those deltas cannot fold and raised a shape
error at switch time. Attention, MLP and both output heads still fold.

README.md

@@ -61,9 +61,10 @@ reference-tree layout as larry under a `diffusion_model.` prefix, so it folds th
 `H3_REALISM=off` skips loading it.
 
 A fourth set is [`joyfox/MiniMax-H3-Turbo`](https://huggingface.co/joyfox/MiniMax-H3-Turbo) (`joyfox`, 4 steps):
-another ComfyUI-layout turbo LoRA, additionally covering the MLPs, the AdaLN projections, and both output heads
-(`video_out` -> `proj_out`, `audio_out` -> `audio_proj_out`), with a per-key `alpha` folded into `lora_B` at load.
-`H3_JOYFOX=off` skips loading it.
+another ComfyUI-layout turbo LoRA, additionally covering the MLPs and both output heads (`video_out` -> `proj_out`,
+`audio_out` -> `audio_proj_out`), with a per-key `alpha` folded into `lora_B` at load. Its AdaLN deltas are skipped —
+the Comfy-Org checkpoint uses an 8-dim modulation input where diffusers uses the full 2688-dim time embedding, so
+they have no fold target. `H3_JOYFOX=off` skips loading it.
 
 ## AoTI-compiled blocks
 

h3_lora.py

@@ -29,7 +29,10 @@ The two supported LoRAs ship in different layouts:
 * `joyfox` (`joyfox/MiniMax-H3-Turbo`) is another 4-step turbo LoRA, ComfyUI-native like realism but covering more
   of the tree: attention, MLP, the block and final AdaLN projections, and both output heads (`final_layer.video_out`
   -> `proj_out`, `final_layer.audio_out` -> `audio_proj_out`). Ranks are mixed (32 attention/MLP, 8 modulation/heads)
-  with a scalar `alpha` per key, so each `lora_B` is prescaled by `alpha / rank` at load (here always 1.0).
+  with a scalar `alpha` per key, so each `lora_B` is prescaled by `alpha / rank` at load (here always 1.0). Its
+  `adaln_proj` entries are skipped: the Comfy-Org checkpoint it was trained against gives the modulation projection
+  an 8-dimensional input (`[96768, 8]`), while the diffusers transformer projects the full 2688-dim time embedding
+  (`[96768, 2688]`), so those deltas have no counterpart to fold into.
 
 `H3_LORA` selects the larry file (`off` skips loading it), `H3_LIGHTX=off` skips lightx, `H3_REALISM=off` skips
 realism, `H3_JOYFOX=off` skips joyfox, `H3_LORA_DEFAULT` picks which set starts folded, and `H3_LORA_STRENGTH` is
@@ -147,6 +150,8 @@ def _load_joyfox(inner_dim: int) -> dict:
     bases = sorted({key.rsplit(".lora_", 1)[0].removeprefix("diffusion_model.") for key in lora if ".lora_" in key})
     entries = []
     for name in bases:
+        if "adaln_proj" in name:
+            continue  # 8-dim modulation input in the Comfy-Org tree vs 2688-dim in diffusers — cannot fold
         prefixed = f"diffusion_model.{name}"
         a = lora[f"{prefixed}.lora_A.weight"]
         # Per-key alpha (scalar tensor), unlike larry's fixed alpha == rank.