Modelpublic

kamusarj/gr00t_N1.7_checkpoint_ur10e-cup_finetuned_gguf

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes24downloads

Model Card

GR00T N1.7 UR10e Cup Finetuned GGUF

This repository contains an F32 GGUF conversion of `Luke99662244/checkpoint2`, a CKA-pruned GR00T N1.7 checkpoint finetuned for the UR10e pick up the cup task.

The model is intended for the GR00T N1.7 implementation in the `agent/add-turbovla-support` branch of vla.cpp. It is not a general-purpose llama.cpp language model.

Files

File	Description
`checkpoint2-f32.gguf`	Self-contained F32 GGUF model
`config.json`	Source checkpoint architecture and pruning manifest
`pruning_runtime.json`	Resolved pruned runtime structure
`processor/`	GR00T processor, normalization statistics, and embodiment mapping
`evaluation/`	Deterministic GGUF open-loop metrics and trajectory plot

GGUF checksum:

text

c37d3563df050d8a2307740e29a078968e17d4313b8490ed29593eb5ee3999be  checkpoint2-f32.gguf

File size: 8,804,339,712 bytes.

Pruned architecture

Module	Retained depth	Original block indices
Language backbone	8	`0, 1, 2, 3, 8, 9, 10, 15`
Action DiT	16	`0, 1, 2, 3, 4, 5, 6, 8, 21, 23, 24, 25, 26, 27, 28, 31`
VL self-attention	4	`0, 1, 2, 3`

The original block indices are stored in GGUF metadata so vla.cpp preserves the source model's self-attention/cross-attention schedule.

Download

bash

hf download \
    kamusarj/gr00t_N1.7_checkpoint_ur10e-cup_finetuned_gguf \
    checkpoint2-f32.gguf \
    --local-dir ./checkpoints/checkpoint2-gguf

Build vla.cpp

bash

git clone --branch agent/add-turbovla-support \
    https://github.com/kamusarj/vla.cpp.git
cd vla.cpp

cmake -S . -B build -DGGML_CUDA=ON -DVLA_BUILD_SERVERS=OFF
cmake --build build --target vla-cli vla-openloop -j

Use VLA_GR00T_EMBODIMENT=new_embodiment and VLA_GR00T_BF16_WEIGHTS=1 when running the model on a CUDA GPU. The vla-cli interface expects processor-prepared language tokens, normalized state, and the two camera views. See the evaluation script and report linked below for an end-to-end recorded-dataset example.

Open-loop evaluation

The GGUF was evaluated on trajectory 0 of ur10e-cup-eval-v2 using the NVIDIA Isaac-GR00T recorded-dataset protocol:

Setting	Value
Task	`pick up the cup`
Frames	814
Inference requests	51
Execution horizon	16
Active action dimensions	7: 6 arm joints + 1 gripper
Camera views	`side`, `wrist`
Noise seed	`20260813`
GPU	NVIDIA GeForce RTX 5060 Ti 16 GB

Metric	GGUF result
Unnormalized action MSE	0.0009261190
Unnormalized action MAE	0.0113530281
Median total latency	66.56 ms/request
Median vision time	36.59 ms/request
Median action inference time	28.34 ms/request

[image]

The original PyTorch evaluation produced MSE 0.0007250374 and MAE 0.0101807602, but that earlier run did not record the same initial diffusion noise. These full-trajectory values are therefore not a controlled estimate of GGUF conversion error.

A fixed-input step-0 parity test using identical images, tokens, state, and BF16 noise produced cosine similarity 0.99999487 over the seven active action dimensions. This supports that the GGUF conversion closely reproduces the PyTorch checkpoint.

Detailed results, limitations, scripts, and raw logs are available in the PyTorch vs GGUF open-loop report.

Limitations

—The reported evaluation covers one recorded trajectory and is not a closed-loop physical-robot success measurement.
—The runtime is currently provided by the linked vla.cpp branch; generic GGUF tools do not implement this GR00T action model architecture.
—Review the source checkpoint and upstream Isaac-GR00T model terms before redistribution or deployment. The source Hub repository does not declare a license in its model metadata at the revision used for this conversion.