Repository navigation

Perf migration - #14

Merged

khanhnd61-vr merged 32 commits into

mainfrom

perf-migration

Jul 6, 2026

Merged

Perf migration#14
khanhnd61-vr merged 32 commits into
mainfrom
perf-migration

Commits on Jul 3, 2026
Commits on Jul 4, 2026
- vla_jepa: reject precomputed_img_emb, guard against null images
  anindex
  committed
  c79a65c
- openvla_oft, vla_adapter: reject precomputed_img_emb, guard null images
  anindex
  committed
  8c6408e
- reject out-of-range tokens and zero padded action dims
  anindex
  committed
  595bcb4
- add vla-cli for one-shot inference without a server
  anindex
  committed
  aae3bf9
- ship vla-cli in the docker image and ci build
  anindex
  committed
  1f32405
- note vla-cli and the token/padding fixes in changelog
  anindex
  committed
  c18f927
- validate vla-cli inputs before load and guard stop_id
  anindex
  committed
  05168b0
- share one gguf reader across the model loaders
  anindex
  committed
  ee2a149
- load quantized weights and add gguf requantizer
  anindex
  committed
  625dcb7
- document quantization and the shared reader
  anindex
  committed
  43a5a0f
- offer only the quant types the writer can pack
  anindex
  committed
  920478b
- release v0.1.1
  anindex
  committed
  bdb9db0
- reject mis-sized images in vla-adapter and openvla-oft
  anindex
  committed
  906030b
- harden evo1 config, bitvla cuda alloc, and server message caps
  anindex
  committed
  16f1d87
- scale cpu threads to core count across the loaders
  anindex
  committed
  15f7c26
- use 64-bit gguf offsets, guard row dtype, cap vlm images
  anindex
  committed
  b703b24
- reuse the prefix cross-attention cache across denoise steps
  anindex
  committed
  543f1d1
- add architecture doc and quickstart, note fixes under 0.1.1
  anindex
  committed
  7f2bf27
Commits on Jul 6, 2026
- add script merge mmproj for pi05, update eval scripts accordingly
  khanhnd61-vr
  committed
  48ab1b8
- fix evo1: use uniform[-1,1) noise
  khanhnd61-vr
  committed
  2efef28
- update converter scripts to bake vision mmproj*.gguf to model.gguf
  khanhnd61-vr
  committed
  b6e58cc
- run regression test and update README.md
  khanhnd61-vr
  committed
  1c22959

Navigation Menu

Repository navigation

Perf migration - #14

Perf migration#14
khanhnd61-vr merged 32 commits into
mainfrom
perf-migration

Commits on Jul 3, 2026

harden gguf arch probe and evo1 input guards

bitvla: absmean ternarize, cache embd reader, fold qkv transpose

vla-server: bound image size and payload, clean shutdown

build pi0/pi0.5/smolvla vision towers in-tree, fetch llama.cpp b9866, drop patch

add dockerfile, github actions build gate, converter deps, tests, docs

evo1: sample base noise from N(0,1), mask padded DiT cross-attention context

smolvla: reject out-of-range lang tokens before the embedding gather

docs: action chunk is padded to max_action_dim; bitvla converter finds latest sidecar

evo1: leave DiT cross-attention unmasked to match the reference

silence -Wall warnings in pi0.5, gr00t-n1.7, vla-adapter

Commits on Jul 4, 2026

vla_jepa: reject precomputed_img_emb, guard against null images

openvla_oft, vla_adapter: reject precomputed_img_emb, guard null images

reject out-of-range tokens and zero padded action dims

add vla-cli for one-shot inference without a server

ship vla-cli in the docker image and ci build

note vla-cli and the token/padding fixes in changelog

validate vla-cli inputs before load and guard stop_id

share one gguf reader across the model loaders

load quantized weights and add gguf requantizer

document quantization and the shared reader

offer only the quant types the writer can pack

release v0.1.1

reject mis-sized images in vla-adapter and openvla-oft

harden evo1 config, bitvla cuda alloc, and server message caps

scale cpu threads to core count across the loaders

use 64-bit gguf offsets, guard row dtype, cap vlm images

reuse the prefix cross-attention cache across denoise steps

add architecture doc and quickstart, note fixes under 0.1.1

Commits on Jul 6, 2026

add script merge mmproj for pi05, update eval scripts accordingly

fix evo1: use uniform[-1,1) noise

update converter scripts to bake vision mmproj*.gguf to model.gguf

run regression test and update README.md