Robotics
Transformers
Safetensors
English
molmoact2
image-text-to-text
OpenRAL
rskill
vision-language-action
nf4
4-bit precision
so100_follower
so101_follower
vla
so101
so100
manipulation
custom_code
8-bit precision
Instructions to use OpenRAL/rskill-molmoact2-multi-so101-nf4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use OpenRAL/rskill-molmoact2-multi-so101-nf4 with Transformers:
# Load model directly from transformers import AutoModelForImageTextToText model = AutoModelForImageTextToText.from_pretrained("OpenRAL/rskill-molmoact2-multi-so101-nf4", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
docs: HF model card for OpenRAL/rskill-molmoact2-multi-so101-nf4 v0.1.0
Browse files
README.md
CHANGED
|
@@ -23,7 +23,7 @@ base_model_relation: quantized
|
|
| 23 |
inference: false
|
| 24 |
---
|
| 25 |
|
| 26 |
-
# rskill-molmoact2-so101-nf4
|
| 27 |
|
| 28 |
> **OpenRAL rSkill** — MolmoAct2 (Ai2's open action reasoning model: a
|
| 29 |
> Molmo2-ER embodied-reasoning VLM backbone with a flow-matching
|
|
@@ -33,7 +33,7 @@ inference: false
|
|
| 33 |
> Robots: SO-100 and SO-101 follower arms. **Apache-2.0 weights** — commercial
|
| 34 |
> use permitted.
|
| 35 |
|
| 36 |
-
This package wraps `hf://OpenRAL/rskill-molmoact2-so101-nf4` (an
|
| 37 |
NF4-quantized mirror of `allenai/MolmoAct2-SO100_101`) with a `rskill.yaml`
|
| 38 |
manifest that adds capability checking, license surfacing, latency budgets,
|
| 39 |
and local registry integration. It does **not** copy model weights — they
|
|
@@ -128,13 +128,13 @@ it does **not** retrain or copy the full-precision weights.
|
|
| 128 |
|
| 129 |
| Field | Value |
|
| 130 |
| --- | --- |
|
| 131 |
-
| `name` | `OpenRAL/rskill-molmoact2-so101-nf4` |
|
| 132 |
| `version` | `0.1.0` |
|
| 133 |
| `license` | `apache-2.0` |
|
| 134 |
| `role` | `s1` |
|
| 135 |
| `embodiment_tags` | `["so100_follower", "so101_follower"]` |
|
| 136 |
| `runtime` / `quantization.dtype` | `pytorch` / `int4` (NF4) |
|
| 137 |
-
| `weights_uri` | `hf://OpenRAL/rskill-molmoact2-so101-nf4` |
|
| 138 |
| `chunk_size` / `n_action_steps` | 10 / 10 (full chunk replay) |
|
| 139 |
| `latency_budget.per_chunk_ms` | 1000 ms |
|
| 140 |
| `commercial_use_allowed` | `true` (Apache-2.0) |
|
|
@@ -154,7 +154,7 @@ print(pkg.manifest.name, pkg.manifest.version)
|
|
| 154 |
|
| 155 |
```bash
|
| 156 |
# CLI:
|
| 157 |
-
uv run openral rskill install OpenRAL/rskill-molmoact2-so101-nf4
|
| 158 |
uv run openral rskill check # does this host meet the requirements?
|
| 159 |
```
|
| 160 |
|
|
@@ -201,7 +201,7 @@ host):
|
|
| 201 |
```bash
|
| 202 |
HF_TOKEN=<write-token> uv run python tools/quantize_rskill.py \
|
| 203 |
--source allenai/MolmoAct2-SO100_101 \
|
| 204 |
-
--target OpenRAL/rskill-molmoact2-so101-nf4 \
|
| 205 |
--loader transformers --trust-remote-code
|
| 206 |
```
|
| 207 |
|
|
@@ -215,7 +215,7 @@ numbers are available yet. Run the reproduction command in
|
|
| 215 |
|
| 216 |
This rSkill package (`rskill.yaml`, `README.md`, `eval/so101.json`) is
|
| 217 |
**Apache-2.0**. The wrapped weights at
|
| 218 |
-
`hf://OpenRAL/rskill-molmoact2-so101-nf4` (NF4 mirror of
|
| 219 |
`allenai/MolmoAct2-SO100_101`) are also released under **Apache-2.0** by Ai2 —
|
| 220 |
commercial use is permitted; review the upstream LICENSE before deployment.
|
| 221 |
|
|
|
|
| 23 |
inference: false
|
| 24 |
---
|
| 25 |
|
| 26 |
+
# rskill-molmoact2-multi-so101-nf4
|
| 27 |
|
| 28 |
> **OpenRAL rSkill** — MolmoAct2 (Ai2's open action reasoning model: a
|
| 29 |
> Molmo2-ER embodied-reasoning VLM backbone with a flow-matching
|
|
|
|
| 33 |
> Robots: SO-100 and SO-101 follower arms. **Apache-2.0 weights** — commercial
|
| 34 |
> use permitted.
|
| 35 |
|
| 36 |
+
This package wraps `hf://OpenRAL/rskill-molmoact2-multi-so101-nf4` (an
|
| 37 |
NF4-quantized mirror of `allenai/MolmoAct2-SO100_101`) with a `rskill.yaml`
|
| 38 |
manifest that adds capability checking, license surfacing, latency budgets,
|
| 39 |
and local registry integration. It does **not** copy model weights — they
|
|
|
|
| 128 |
|
| 129 |
| Field | Value |
|
| 130 |
| --- | --- |
|
| 131 |
+
| `name` | `OpenRAL/rskill-molmoact2-multi-so101-nf4` |
|
| 132 |
| `version` | `0.1.0` |
|
| 133 |
| `license` | `apache-2.0` |
|
| 134 |
| `role` | `s1` |
|
| 135 |
| `embodiment_tags` | `["so100_follower", "so101_follower"]` |
|
| 136 |
| `runtime` / `quantization.dtype` | `pytorch` / `int4` (NF4) |
|
| 137 |
+
| `weights_uri` | `hf://OpenRAL/rskill-molmoact2-multi-so101-nf4` |
|
| 138 |
| `chunk_size` / `n_action_steps` | 10 / 10 (full chunk replay) |
|
| 139 |
| `latency_budget.per_chunk_ms` | 1000 ms |
|
| 140 |
| `commercial_use_allowed` | `true` (Apache-2.0) |
|
|
|
|
| 154 |
|
| 155 |
```bash
|
| 156 |
# CLI:
|
| 157 |
+
uv run openral rskill install OpenRAL/rskill-molmoact2-multi-so101-nf4
|
| 158 |
uv run openral rskill check # does this host meet the requirements?
|
| 159 |
```
|
| 160 |
|
|
|
|
| 201 |
```bash
|
| 202 |
HF_TOKEN=<write-token> uv run python tools/quantize_rskill.py \
|
| 203 |
--source allenai/MolmoAct2-SO100_101 \
|
| 204 |
+
--target OpenRAL/rskill-molmoact2-multi-so101-nf4 \
|
| 205 |
--loader transformers --trust-remote-code
|
| 206 |
```
|
| 207 |
|
|
|
|
| 215 |
|
| 216 |
This rSkill package (`rskill.yaml`, `README.md`, `eval/so101.json`) is
|
| 217 |
**Apache-2.0**. The wrapped weights at
|
| 218 |
+
`hf://OpenRAL/rskill-molmoact2-multi-so101-nf4` (NF4 mirror of
|
| 219 |
`allenai/MolmoAct2-SO100_101`) are also released under **Apache-2.0** by Ai2 —
|
| 220 |
commercial use is permitted; review the upstream LICENSE before deployment.
|
| 221 |
|