Views
No views yet
| side | rotation | quantizer | per-step |
|---|---|---|---|
| PaliGemma backbone | DuQuant svd_hadamard | GPTQ | single bucket |
| Gemma Expert (action head) | DuQuant svd_hadamard | RTN residual | yes (act_scale_table, 10 steps) |
duquant_rotation_blocks + duquant_rotation_perm).quantized.pt is a
dict keyed by layer name; every record holds the quantized weight
(weight_res_q / baseline_q), the block-form rotation, and (Expert side) the
per-step act_scale_table. They are loaded at inference time on top of the
original pi0.5 FP checkpoint by Omega-QVLA's GptqLinear. You cannot
from_pretrained them directly.| file | suite | records |
|---|---|---|
pi05_object/quantized.pt | libero_object | 252 (126 PaliGemma + 126 Expert) |
pi05_spatial/quantized.pt | libero_spatial | 252 |
pi05_goal/quantized.pt | libero_goal | 252 |
pi05_long/quantized.pt | libero_10 | 252 |
1# 1. Get the repo + the original pi0.5 PyTorch checkpoint
2git clone https://github.com/UCMP13753/Omega-QVLA && cd Omega-QVLA
3
4SUITE=object
5PACK=/path/to/pi05_${SUITE}/quantized.pt
6INCLUDE_BOTH='.*paligemma_with_expert\.(paligemma\.model\.language_model|gemma_expert\.model)\.layers\.[0-9]+\..*\.(q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj).*'
7
8env CONDA_ROOT=$HOME/miniconda3 METHOD=gptq SUITE=$SUITE WBITS=4 ABITS=4 \
9 GPU_LIST=0,1 PORT_BASE=8600 NUM_TRIALS_PER_TASK=10 GR00T_EVAL_INIT_OFFSET=10 \
10 OPENPI_ROOT=$HOME/openpi OPENPI_PY=$HOME/openpi/.venv/bin/python \
11 OPENPI_CONFIG=pi05_libero OPENPI_CHECKPOINT=/path/to/pi05_libero_pytorch \
12 OPENPI_GPTQ_PATH="$PACK" OPENPI_GPTQ_INCLUDE="$INCLUDE_BOTH" \
13 OUTPUT_ROOT=results/eval/pi05_${SUITE} \
14 bash scripts/run_pi05_libero_benchmark.shGptqLinear); the block-form rotation
and per-step scales are applied automatically.