SpecPrune-VLA reproduction
Collection
Independent reproduction of SpecPrune-VLA (ICML 2026, arXiv 2509.05614): logbook, bundle, base models. • 6 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
gpu: string
n_images: int64
models: struct<vit_so400m_patch14_siglip_224.webli: struct<params_B: double, input_px: int64, n_patch_tokens (... 162 chars omitted)
child 0, vit_so400m_patch14_siglip_224.webli: struct<params_B: double, input_px: int64, n_patch_tokens: int64, median_ms_2img: double>
child 0, params_B: double
child 1, input_px: int64
child 2, n_patch_tokens: int64
child 3, median_ms_2img: double
child 1, vit_large_patch14_reg4_dinov2.lvd142m: struct<params_B: double, input_px: int64, n_patch_tokens: int64, median_ms_2img: double>
child 0, params_B: double
child 1, input_px: int64
child 2, n_patch_tokens: int64
child 3, median_ms_2img: double
vision_total_ms: double
action_head_ms_assumed: double
vision_plus_action_ms: double
schedules: struct<full: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: doubl (... 561 chars omitted)
child 0, full: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
child 1, static_arxiv: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10
...
chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
child 3, static_claim: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
child 4, final_claim: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
iters: int64
dims: struct<D: int64, H: int64, M: int64, heads: int64, L_total: int64, Lr_arxiv: int64, Lr_claim: int64>
child 0, D: int64
child 1, H: int64
child 2, M: int64
child 3, heads: int64
child 4, L_total: int64
child 5, Lr_arxiv: int64
child 6, Lr_claim: int64
speedups: struct<arxiv_static_only: double, arxiv_two_level(final): double, claim_static_only: double, claim_t (... 24 chars omitted)
child 0, arxiv_static_only: double
child 1, arxiv_two_level(final): double
child 2, claim_static_only: double
child 3, claim_two_level(final): double
torch: string
dtype: string
to
{'gpu': Value('string'), 'torch': Value('string'), 'dtype': Value('string'), 'dims': {'D': Value('int64'), 'H': Value('int64'), 'M': Value('int64'), 'heads': Value('int64'), 'L_total': Value('int64'), 'Lr_arxiv': Value('int64'), 'Lr_claim': Value('int64')}, 'iters': Value('int64'), 'schedules': {'full': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'static_arxiv': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'final_arxiv': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'static_claim': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'final_claim': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}}, 'speedups': {'arxiv_static_only': Value('float64'), 'arxiv_two_level(final)': Value('float64'), 'claim_static_only': Value('float64'), 'claim_two_level(final)': Value('float64')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 478, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
gpu: string
n_images: int64
models: struct<vit_so400m_patch14_siglip_224.webli: struct<params_B: double, input_px: int64, n_patch_tokens (... 162 chars omitted)
child 0, vit_so400m_patch14_siglip_224.webli: struct<params_B: double, input_px: int64, n_patch_tokens: int64, median_ms_2img: double>
child 0, params_B: double
child 1, input_px: int64
child 2, n_patch_tokens: int64
child 3, median_ms_2img: double
child 1, vit_large_patch14_reg4_dinov2.lvd142m: struct<params_B: double, input_px: int64, n_patch_tokens: int64, median_ms_2img: double>
child 0, params_B: double
child 1, input_px: int64
child 2, n_patch_tokens: int64
child 3, median_ms_2img: double
vision_total_ms: double
action_head_ms_assumed: double
vision_plus_action_ms: double
schedules: struct<full: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: doubl (... 561 chars omitted)
child 0, full: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
child 1, static_arxiv: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10
...
chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
child 3, static_claim: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
child 4, final_claim: struct<median_ms: double, min_ms: double, p10_ms: double, std_ms: double, TFLOPs: double, flops_frac (... 17 chars omitted)
child 0, median_ms: double
child 1, min_ms: double
child 2, p10_ms: double
child 3, std_ms: double
child 4, TFLOPs: double
child 5, flops_frac_of_full: double
iters: int64
dims: struct<D: int64, H: int64, M: int64, heads: int64, L_total: int64, Lr_arxiv: int64, Lr_claim: int64>
child 0, D: int64
child 1, H: int64
child 2, M: int64
child 3, heads: int64
child 4, L_total: int64
child 5, Lr_arxiv: int64
child 6, Lr_claim: int64
speedups: struct<arxiv_static_only: double, arxiv_two_level(final): double, claim_static_only: double, claim_t (... 24 chars omitted)
child 0, arxiv_static_only: double
child 1, arxiv_two_level(final): double
child 2, claim_static_only: double
child 3, claim_two_level(final): double
torch: string
dtype: string
to
{'gpu': Value('string'), 'torch': Value('string'), 'dtype': Value('string'), 'dims': {'D': Value('int64'), 'H': Value('int64'), 'M': Value('int64'), 'heads': Value('int64'), 'L_total': Value('int64'), 'Lr_arxiv': Value('int64'), 'Lr_claim': Value('int64')}, 'iters': Value('int64'), 'schedules': {'full': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'static_arxiv': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'final_arxiv': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'static_claim': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}, 'final_claim': {'median_ms': Value('float64'), 'min_ms': Value('float64'), 'p10_ms': Value('float64'), 'std_ms': Value('float64'), 'TFLOPs': Value('float64'), 'flops_frac_of_full': Value('float64')}}, 'speedups': {'arxiv_static_only': Value('float64'), 'arxiv_two_level(final)': Value('float64'), 'claim_static_only': Value('float64'), 'claim_two_level(final)': Value('float64')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SpecPrune-VLA reproduction bundle
Independent reproduction of SpecPrune-VLA: Accelerating Vision-Language-Action Models
via Action-Aware Self-Speculative Pruning (ICML 2026, OpenReview MjE62ZnpRH,
arXiv 2509.05614).
Base model: OpenVLA-OFT (Llama-2-7B backbone) —
moojink/openvla-7b-oft-finetuned-libero-spatial
and the object/goal/10 variants · code scaffold: https://github.com/moojink/openvla-oft ·
benchmark: https://github.com/Lifelong-Robot-Learning/LIBERO
What's here
src/
flops_model.py # Claim 4: analytical FLOPs / token-retention model (paper A.3)
bench_llm.py # Claim 1: OpenVLA-OFT LLM-backbone latency under pruning schedules
bench_vision.py # vision-encoder (SigLIP+DINOv2) latency for end-to-end
analyze_speedup.py # combine LLM+vision -> end-to-end speedups
make_results.py # speedup comparison table + plotly figure
make_flops_figure.py # FLOPs-model validation figure
make_poster.py # builds poster.png + poster_embed.html (Playwright render + hotspots)
remote_setup.sh # deps on the GPU box
outputs/
flops_model.json # FLOPs model results (validates paper A.3: 0.44/0.37, 63% reduction)
bench_llm_results.json / bench_vision_results.json / speedup_summary.json
speedup_table.{csv,md} / speedup_figure.html / flops_figure.html
poster_source.html / poster.png / poster_embed.html
paper.txt # extracted arXiv text (evidence for Claim 5 quotes)
Key results
| Claim | Verdict | Evidence |
|---|---|---|
| 1 — 1.46×/1.57× speedup, <0.7% SR drop | speedup reproduced | LLM 1.998× on real RTX 3090 vs 2.09× reported; e2e 1.49–1.75× brackets 1.57× |
| 2 — per-suite speedup + SR | partial | speedup band reproduced; SR needs unreleased code |
| 3 — ablation 1.54×/1.51× | partial | speed↔accuracy mechanism + speedups reproduced; SR out of scope |
| 4 — ~48% retention → ~37% FLOPs | reproduced (analytical) | model matches paper A.3: static 0.43/0.44, final 0.36/0.37, 63.8% LLM reduction, Table-8 exact |
| 5 — LIBERO-only, no real robot | refuted | arXiv 2509.05614 adds a real-robot section (Flexiv arm, 1.70×) |
How to rerun
# 1. Analytical FLOPs model (Claim 4) — local, CPU, seconds
python3 src/flops_model.py --out outputs/flops_model.json
# 2. Speedup benchmark (Claim 1) — needs a CUDA GPU (paper used RTX 3090 / A800)
pip install torch timm
python3 src/bench_llm.py --iters 60 --warmup 15 --out outputs/bench_llm_results.json
python3 src/bench_vision.py --iters 60 --warmup 15 --out outputs/bench_vision_results.json
python3 src/analyze_speedup.py --out outputs/speedup_summary.json
# 3. Figures + poster
python3 src/make_results.py && python3 src/make_flops_figure.py
python3 src/make_poster.py # needs playwright + chromium
# paper text (evidence): curl -sL https://arxiv.org/pdf/2509.05614 | pdftotext -layout - paper.txt
GPU used: 1× NVIDIA RTX 3090 (24 GB) on Vast.ai, ~25 min, ~$0.08. (HF Jobs were unavailable — account returned HTTP 402, insufficient credits.)
No official SpecPrune-VLA code exists, so the speedup is reproduced at the mechanism level (real OpenVLA-OFT backbone at the paper's token-retention schedule); the success-rate numbers require the unreleased token-selection policy + full LIBERO rollouts.
- Downloads last month
- 41