Views
No views yet
| side | rotation | quantizer | per-step |
|---|---|---|---|
| Eagle LLM (backbone) | DuQuant svd_hadamard | GPTQ | — |
| DiT (action head) | DuQuant svd_hadamard | RTN residual | yes (act_scale_table, 8 steps) |
quantized.pt is a dict keyed
by layer name (180 records = 84 Eagle LLM + 96 DiT); records hold the quantized
weight (baseline_q / weight_res_q), the DuQuant rotation, and (DiT) the per-step
act_scale_table. Loaded at inference time on top of the original FP checkpoint
youliangtan/gr00t-n1.5-libero-<suite>-posttrain by Omega-QVLA's GptqLinear.| file | suite |
|---|---|
gr00t_object/quantized.pt | libero_object |
gr00t_spatial/quantized.pt | libero_spatial |
gr00t_goal/quantized.pt | libero_goal |
gr00t_long/quantized.pt | libero_10 |
1git clone https://github.com/UCMP13753/Omega-QVLA && cd Omega-QVLA
2
3SUITE=object
4PACK=/path/to/gr00t_${SUITE}/quantized.pt
5LLM_RE='.*backbone\.eagle_model\.language_model\..*\.(q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj).*'
6DIT_RE='.*action_head\.model\.transformer_blocks\.\d+\.(attn1\.(to_q|to_k|to_v|to_out\.0)|ff\.net\.(0\.proj|2)).*'
7
8env CONDA_ROOT=$HOME/miniconda3 \
9 SUITE=$SUITE WBITS=4 ABITS=4 LLM_QUANT=gptq DIT_QUANT=gptq DIT_ATTN=1 DIT_PERSTEP=1 \
10 GR00T_GPTQ_PATH_OVERRIDE="$PACK" GR00T_GPTQ_INCLUDE_OVERRIDE="(${LLM_RE}|${DIT_RE})" \
11 GR00T_GPTQ_MISSING=fallback \
12 GPU_LIST=0,1,2,3 PORT_BASE=8000 NUM_TRIALS_PER_TASK=10 GR00T_EVAL_INIT_OFFSET=10 \
13 OUTPUT_ROOT=results/eval/${SUITE} \
14 bash scripts/run_groot_benchmark.sh