Repository navigation

vla.cpp performance review - #32

Merged

khanhnd61-vr merged 61 commits into

mainfrom

b11223-numerics-perf

Sep 30, 2026

Commits on Sep 28, 2026
- bump llama.cpp to b11223
  anindex
  committed
  209cf8c
- bump sentencepiece to v0.2.1
  anindex
  committed
  3faea0a
- bump OpenVINO, macOS runner and python deps
  anindex
  committed
  eff454e
- harden vla-server and vlm-server
  anindex
  committed
  79533b2
- fix loader fuse, option precedence and C API races
  anindex
  committed
  66ce698
- size graphs from node counts, validate pi0/smolvla/oft input
  anindex
  committed
  3a07dfc
- validate GR00T, Evo-1, VLA-JEPA and TurboVLA metadata
  anindex
  committed
  b7a4e7d
- fix Octo crashes on bad tokens and stats
  anindex
  committed
  d32b1ac
- fix BitVLA uploads, device choice and kernel races
  anindex
  committed
  9bf951d
- fix BF16 CUDA hook overflow and races
  anindex
  committed
  aa40537
- fix quantizer, converters and eval client
  anindex
  committed
  06969f6
- harden CI, add CUDA, OpenCL and Metal build checks
  anindex
  committed
  c1eaec5
- update changelog
  anindex
  committed
  d90aee9
- pair LIBERO runs with seeded noise, add octo, turbovla, vla_jepa
  anindex
  committed
  4a7a423
- move GR00T and VLA-JEPA onto shared modules
  anindex
  committed
  e020e49
- share the PaliGemma stack between pi0 and pi05
  anindex
  committed
  abf3203
- share the dual tower and RoPE helpers between OFT and VLA-Adapter
  anindex
  committed
  5343639
- put SmolVLA on gguf_reader and one graph builder
  anindex
  committed
  82833df
- use shared layers in Octo
  anindex
  committed
  a85458a
- share image preprocessing with Evo-1 and BitVLA
  anindex
  committed
  aa14ea7
- wire --flash-attn into GR00T N1.5/N1.6 and multi-view encoders
  anindex
  committed
  271ca9c
- vla-cli takes runtime flags and --config
  anindex
  committed
  61115ca
- -hf picks files and tags, --text uses per-arch prompts
  anindex
  committed
  f041e07
- add install rules and a pip wheel for the bindings
  anindex
  committed
  9bb1423
- fix release packaging, multi-arch multi-stage Docker image
  anindex
  committed
  2d289db
- bake pi05 adaRMS at load, GEGLU and on-device vision copy for pi0/pi05
  anindex
  committed
  c749729
- bake DiT conditioning at load, keep one GR00T embodiment resident
  anindex
  committed
  7e6ece2
- SmolVLA: widen weights at load, in-graph pixel shuffle, embeddings off device
  anindex
  committed
  dcf9641
- cache TurboVLA text encoding and Octo constants
  anindex
  committed
  6b7039b
- upload Evo-1, OFT and VLA-Adapter constants once, keep vision on device
  anindex
  committed
  0e73bbf
- drop the BitVLA ViT pad copy
  anindex
  committed
  9a36ab6
- pi0: feed raw projector output to the LM
  anindex
  committed
  f2f5c02
- pi05: drop the image scale round-trip
  anindex
  committed
  25523b6
- apply DINOv3 final norm in TurboVLA
  anindex
  committed
  cf2dc60
- octo: match the JAX reference numerics
  anindex
  committed
  6bd49f4
- use erf GELU in Qwen3-VL patch mergers
  anindex
  committed
  db4a3ed
- smolvla: keep cross-attn k/v projections in F32
  anindex
  committed
  67693db
- evo1: match the reference prompt and constant state dims
  anindex
  committed
  a8ecdf3
- bitvla: round to bf16 to nearest even, recover legacy scale
  anindex
  committed
  dbf422b
- tap VLA-Adapter head blocks like the reference
  anindex
  committed
  4e14852
- eval client: normalize VLA-Adapter proprio, zero constant GR00T dims
  anindex
  committed
  4f85ac4
- changelog for the math fixes and speedups
  anindex
  committed
  2a69d3b
- tokenizer in the GGUF, Octo always built, VLA_SPM option
  anindex
  committed
  c753c25
- add --num-steps as a load-time option
  anindex
  committed
  55b9f01
- drop dead preprocessing and norm helpers
  anindex
  committed
  e2af0b2
- update docs for this release
  anindex
  committed
  335a87f
- changelog for tokenizer and num-steps
  anindex
  committed
  585b465
- bitvla: convert bf16 uploads with ggml's row helper
  anindex
  committed
  0548f1f
- bound Octo token count, quiet safetensors probe, refuse unused act-dtype
  anindex
  committed
  ad56e9b
- pi05 converter QUANTILES only, vla_jepa venv override, learned noise length
  anindex
  committed
  33e90ba
- self-contained wheel, macOS tarball without servers, portable PTX, licenses
  anindex
  committed
  a8698f7
- fix stale docs and changelog claims
  anindex
  committed
  a427801
Commits on Sep 29, 2026
Commits on Sep 30, 2026
- fix(openvino): key the RoPE sin/cos cache on the position input
  khanhnd61-vr
  committed
  3ff447c
- docs(turbovla): drop the re-convert notes
  khanhnd61-vr
  committed
  450992c
- release: v0.4.0 changelog, version bump, split README into docs
  khanhnd61-vr
  committed
  07196d0
- add report libero and Intel UPX board
  khanhnd61-vr
  committed
  c7e1eb2
- update CHANGELOG for v0.4.0
  khanhnd61-vr
  committed
  096dc61

Commits

Commits on Sep 28, 2026

Commits on Sep 29, 2026

Commits on Sep 30, 2026