[misc] MiniMax-H3: move the AdaLN converter into scripts/checkpoint_conversion (#1712)

examples/inference/basic/basic_minimax_h3_fl2va.py

@@ -27,7 +27,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--model-path", default="MiniMaxAI/MiniMax-H3")
     # Rank-reduced AdaLN checkpoint (-39% params, -23 GiB VRAM): pass
     #   --model-path noctuashap/MiniMax-H3-pruned-r16
-    # (or a local dir converted with tools/minimax_h3/fit_adaln_basis.py).
+    # (or a local dir produced by
+    #   scripts/checkpoint_conversion/convert_minimax_h3_adaln_rank.py).
     # adaln_rank is read from the checkpoint config; no other flags needed.
     # Rank-reduced checkpoints are inference-only: training needs the
     # full-rank release.

examples/inference/basic/basic_minimax_h3_ref2va.py

@@ -27,7 +27,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--model-path", default="MiniMaxAI/MiniMax-H3")
     # Rank-reduced AdaLN checkpoint (-39% params, -23 GiB VRAM): pass
     #   --model-path noctuashap/MiniMax-H3-pruned-r16
-    # (or a local dir converted with tools/minimax_h3/fit_adaln_basis.py).
+    # (or a local dir produced by
+    #   scripts/checkpoint_conversion/convert_minimax_h3_adaln_rank.py).
     # adaln_rank is read from the checkpoint config; no other flags needed.
     # Rank-reduced checkpoints are inference-only: training needs the
     # full-rank release.

examples/inference/basic/basic_minimax_h3_t2v.py

@@ -24,7 +24,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--model-path", default="MiniMaxAI/MiniMax-H3")
     # Rank-reduced AdaLN checkpoint (-39% params, -23 GiB VRAM): pass
     #   --model-path noctuashap/MiniMax-H3-pruned-r16
-    # (or a local dir converted with tools/minimax_h3/fit_adaln_basis.py).
+    # (or a local dir produced by
+    #   scripts/checkpoint_conversion/convert_minimax_h3_adaln_rank.py).
     # adaln_rank is read from the checkpoint config; no other flags needed.
     # Rank-reduced checkpoints are inference-only: training needs the
     # full-rank release.

fastvideo/models/dits/minimax_h3.py

@@ -546,7 +546,7 @@ class MiniMaxH3Transformer3DModel(BaseDiT):
                 "parameter, but factorized AdaLN weights are pinned to FP16 "
                 "(BF16 reconstructs them ~1.7x worse). Fine-tune the full-rank "
                 "checkpoint instead, then re-fit the basis with "
-                "tools/minimax_h3/fit_adaln_basis.py.")
+                "scripts/checkpoint_conversion/convert_minimax_h3_adaln_rank.py.")
         adaln_dim = self.adaln_rank or arch.time_embed_dim
         self.adaln_basis = ReplicatedLinear(
             arch.time_embed_dim,

scripts/checkpoint_conversion/AGENTS.md

@@ -14,6 +14,7 @@ checkpoint_conversion/
 ├── convert_gamecraft_weights.py         # DiT only
 ├── convert_gen3c_to_fastvideo.py
 ├── convert_ltx2_weights.py
+├── convert_minimax_h3_adaln_rank.py     # Rank-reduces AdaLN in place; same family
 ├── convert_turbodiffusion_to_diffusers.py
 ├── convert_turbodiffusion_i2v_to_diffusers.py
 ├── extract_llava_text_encoder.py        # Encoder extraction from a multimodal repo

scripts/checkpoint_conversion/convert_minimax_h3_adaln_rank.py

@@ -24,7 +24,7 @@ directly: ``adaln_rank`` in ``config.json`` flows onto the arch config through
 
 Usage::
 
-    python tools/minimax_h3/fit_adaln_basis.py \
+    python scripts/checkpoint_conversion/convert_minimax_h3_adaln_rank.py \
         --src  /path/to/MiniMax-H3/transformer \
         --dst  /path/to/MiniMax-H3-r16/transformer --rank 16