Normalize Motius Pipeline.from_pretrained metadata

README.md

@@ -1,5 +1,5 @@
 ---
-library_name: hftrainer
+library_name: motius
 pipeline_tag: other
 tags:
 - motion-generation
@@ -14,18 +14,18 @@ license: other
 # Go to Zero (MotionMillion)
 
 Million-scale, 3B/7B-parameter autoregressive text-to-motion model ("Go to Zero",
-ICCV 2025 **Highlight**) integrated into the hftrainer Model Zoo. Our
+ICCV 2025 **Highlight**) integrated into the motius Model Zoo. Our
 reproduction is **fully self-contained and independent of the original
 repository** at runtime: the HumanVQVAE (FSQ tokenizer) and the LLaMA
 autoregressive transformer live in
-`hftrainer.models.motion.motionmillion.network`. Only the T2M inference path is
+`motius.models.motion.motionmillion.network`. Only the T2M inference path is
 exercised.
 
 | | |
 |---|---|
 | **Task** | Text-to-Motion (T2M), zero-shot |
 | **Bundle / Pipeline** | `MotionMillionBundle` / `MotionMillionPipeline` |
-| **Processed HF artifact** | [`7B-train`](https://huggingface.co/ZeyuLing/hftrainer-gotozero-7b-train-humanml272), [`3B-train`](https://huggingface.co/ZeyuLing/hftrainer-gotozero-3b-train-humanml272) |
+| **Processed HF artifact** | [`7B-train`](https://huggingface.co/ZeyuLing/Motius-MotionMillion-7B-HumanML272), [`3B-train`](https://huggingface.co/ZeyuLing/Motius-MotionMillion-3B-HumanML272) |
 | **Motion representation** | **humanml3d_272** (272-dim, 30 fps) — *identical layout to MotionStreamer-272* |
 | **Tokenizer** | HumanVQVAE + **FSQ** (levels `[8,8,8,5,5,5]`, codebook 64000) |
 | **AR model** | LLaMA-style 3B / 7B AR transformer (RoPE, length-causal text cross-attn) |
@@ -39,14 +39,14 @@ exercised.
 
 ## Weights
 
-Current hftrainer artifact (diffusers-style `from_pretrained`):
+Current motius artifact (diffusers-style `from_pretrained`):
 
 | Artifact | Location | Contents | Status |
 |---|---|---|---|
-| Go-to-Zero 7B train-only HumanML3D-272 | [`ZeyuLing/hftrainer-gotozero-7b-train-humanml272`](https://huggingface.co/ZeyuLing/hftrainer-gotozero-7b-train-humanml272) | `fsq.safetensors` + `ar.safetensors` + `mm_config.json` + `model_index.json` + `mean.npy` / `std.npy` + `text_encoder/` | public Hub artifact; Flan-T5-XL is packaged as safetensors |
-| Go-to-Zero 3B train-only HumanML3D-272 | [`ZeyuLing/hftrainer-gotozero-3b-train-humanml272`](https://huggingface.co/ZeyuLing/hftrainer-gotozero-3b-train-humanml272) | same layout | public Hub artifact; Flan-T5-XL is packaged as safetensors |
-| local mirror, 7B | `checkpoints/gotozero/hftrainer_7b_train_humanml272` | same layout | optional local cache |
-| local mirror, 3B | `checkpoints/gotozero/hftrainer_3b_train_humanml272` | same layout | optional local cache |
+| Go-to-Zero 7B train-only HumanML3D-272 | [`ZeyuLing/Motius-MotionMillion-7B-HumanML272`](https://huggingface.co/ZeyuLing/Motius-MotionMillion-7B-HumanML272) | `fsq.safetensors` + `ar.safetensors` + `mm_config.json` + `model_index.json` + `mean.npy` / `std.npy` + `text_encoder/` | public Hub artifact; Flan-T5-XL is packaged as safetensors |
+| Go-to-Zero 3B train-only HumanML3D-272 | [`ZeyuLing/Motius-MotionMillion-3B-HumanML272`](https://huggingface.co/ZeyuLing/Motius-MotionMillion-3B-HumanML272) | same layout | public Hub artifact; Flan-T5-XL is packaged as safetensors |
+| local mirror, 7B | `checkpoints/gotozero/motius_7b_train_humanml272` | same layout | optional local cache |
+| local mirror, 3B | `checkpoints/gotozero/motius_3b_train_humanml272` | same layout | optional local cache |
 
 The public leaderboard uses the **train-only** checkpoints for both 7B and 3B.
 The older `t2m_*_all.zip` checkpoints were trained with all HumanML3D splits and
@@ -55,13 +55,13 @@ should not be used for HumanML3D test-set leaderboard rows.
 **Use directly from the Hub (recommended):**
 
 ```python
-from hftrainer.pipelines.motionmillion import MotionMillionPipeline
+from motius.pipelines.motionmillion import MotionMillionPipeline
 
 pipe = MotionMillionPipeline.from_pretrained(
-    "ZeyuLing/hftrainer-gotozero-7b-train-humanml272",
+    "ZeyuLing/Motius-MotionMillion-7B-HumanML272",
     device="cuda",
 )
-# Use "ZeyuLing/hftrainer-gotozero-3b-train-humanml272" for the 3B train-only artifact.
+# Use "ZeyuLing/Motius-MotionMillion-3B-HumanML272" for the 3B train-only artifact.
 # Cast the AR to bf16 for memory-efficient inference:
 import torch
 pipe.bundle.ar.to(dtype=torch.bfloat16)
@@ -79,20 +79,20 @@ bundle = MotionMillionBundle(
 )
 ```
 
-Package the hftrainer artifact from local upstream weights:
+Package the motius artifact from local upstream weights:
 
 ```bash
 python3 scripts/eval/convert_motionmillion_checkpoint.py \
   --ar checkpoints/motionmillion/pretrained_models/motionmillion_7B.pth \
-  --out_dir checkpoints/gotozero/hftrainer_7b_train_humanml272 \
-  --repo_id ZeyuLing/hftrainer-gotozero-7b-train-humanml272 \
+  --out_dir checkpoints/gotozero/motius_7b_train_humanml272 \
+  --repo_id ZeyuLing/Motius-MotionMillion-7B-HumanML272 \
   --text_model_source checkpoints/flan-t5-xl \
   --verify
 ```
 
 Use the same converter with the 3B train-only checkpoint and
-`--out_dir checkpoints/gotozero/hftrainer_3b_train_humanml272 --repo_id
-ZeyuLing/hftrainer-gotozero-3b-train-humanml272` for the 3B artifact.
+`--out_dir checkpoints/gotozero/motius_3b_train_humanml272 --repo_id
+ZeyuLing/Motius-MotionMillion-3B-HumanML272` for the 3B artifact.
 
 ---
 
@@ -112,10 +112,10 @@ text -> Flan-T5-XL -> LLaMA-style AR (greedy, EOS-stopped, ≤150 tokens)
 
 The upstream demo sampler caps generation at 50 motion tokens (about 100 frames
 at 30 fps), which truncates many HumanML3D test clips. The model block size
-supports the full benchmark range, so the hftrainer reproduction uses
+supports the full benchmark range, so the motius reproduction uses
 `max_sample_steps=150` and writes exact-GT-length MS272 predictions.
 
-Convert to HumanML3D-263 with `hftrainer.motion.representation.convert`
+Convert to HumanML3D-263 with `motius.motion.representation.convert`
 (`motion272_to_hml263`).
 
 ---
@@ -136,12 +136,12 @@ Reproduce generation with the packaged artifact:
 ```bash
 NGPU=6 TOTAL_SHARDS=6 LIMIT=0 \
 OUT=outputs/evaluation/t2m/humanml3d_official_test/ms272/gotozero_7b_train \
-ARTIFACT=checkpoints/gotozero/hftrainer_7b_train_humanml272 \
+ARTIFACT=checkpoints/gotozero/motius_7b_train_humanml272 \
 STEPS=150 DTYPE=bf16 \
 bash scripts/eval/run_motionmillion_official272_exactlen_genonly.sh
 ```
 
-Swap `OUT` and `ARTIFACT` to the `gotozero_3b_train` / `hftrainer_3b_train_humanml272`
+Swap `OUT` and `ARTIFACT` to the `gotozero_3b_train` / `motius_3b_train_humanml272`
 paths for the 3B train-only row.
 
 ### MotionStreamer-272 evaluator (native space)
@@ -209,7 +209,7 @@ use it as a secondary diagnostic.
 
 ## Implementation notes
 
-- **hftrainer-native runtime**: `hftrainer/models/motion/motionmillion/network/` holds
+- **motius-native runtime**: `motius/models/motion/motionmillion/network/` holds
   `fsq.py` (Finite Scalar Quantization), `resnet.py` / `modules.py` (Haar
   patch/unpatch) / `encdec.py` / `vqvae.py` (HumanVQVAE tokenizer) and
   `llama.py` (LLaMA AR). The `args` namespace was refactored into explicit
@@ -220,8 +220,16 @@ use it as a secondary diagnostic.
   state next to the `trans` weights; a tolerant unpickler stubs the unimportable
   classes so only the tensor `state_dict` is read (repo-independent).
 - **Complete text encoder artifact**: Flan-T5-XL is frozen and stored under
-  `text_encoder/` in the hftrainer artifact. `MotionMillionBundle.from_pretrained`
+  `text_encoder/` in the motius artifact. `MotionMillionBundle.from_pretrained`
   resolves that artifact-local directory automatically.
 - **Representation parity verified**: on smoke pairs the de-normalised prediction
   matches GT per-block scale almost exactly (e.g. position-block std 0.514 vs
   0.514; rot6d-block std 0.495 vs 0.489), confirming MotionMillion-272 ≡ MS-272.
+
+## Direct Loading
+
+```python
+from motius import Pipeline
+
+pipeline = Pipeline.from_pretrained("ZeyuLing/Motius-MotionMillion-3B-HumanML272")
+```

model_index.json

@@ -1,19 +1,40 @@
 {
   "_class_name": "MotionMillionPipeline",
-  "_diffusers_version": "hftrainer",
-  "pipeline": {
-    "library": "hftrainer",
-    "class_name": "MotionMillionPipeline",
-    "module": "hftrainer.pipelines.motionmillion"
-  },
-  "bundle": {
-    "library": "hftrainer",
-    "class_name": "MotionMillionBundle",
-    "module": "hftrainer.models.motion.motionmillion"
-  },
   "weights": {
     "fsq": "fsq.safetensors",
     "ar": "ar.safetensors",
     "text_encoder": "text_encoder"
-  }
-}
+  },
+  "_library_name": "motius",
+  "format_version": 1,
+  "pipeline_class": "motius.pipelines.motionmillion.MotionMillionPipeline",
+  "bundle_class": "motius.models.motionmillion.MotionMillionBundle",
+  "tasks": [
+    "text_to_motion"
+  ],
+  "required_files": [
+    "ar.safetensors",
+    "fsq.safetensors",
+    "mean.npy",
+    "mm_config.json",
+    "std.npy",
+    "text_encoder/.gitattributes",
+    "text_encoder/README.md",
+    "text_encoder/config.json",
+    "text_encoder/generation_config.json",
+    "text_encoder/model-00001-of-00002.safetensors",
+    "text_encoder/model-00002-of-00002.safetensors",
+    "text_encoder/model.safetensors.index.json",
+    "text_encoder/special_tokens_map.json",
+    "text_encoder/spiece.model",
+    "text_encoder/tokenizer.json",
+    "text_encoder/tokenizer_config.json"
+  ],
+  "api": {
+    "loader": "motius.Pipeline.from_pretrained",
+    "task_methods": [
+      "infer_text_to_motion"
+    ]
+  },
+  "artifact_format": "motius-motionmillion-v1"
+}