Intel GPU backend, shared graph cache, and checkpoint hardening (0.2.0) - #18

Open

khanhnd61-vr wants to merge 34 commits into

mainfrom

Commits on Aug 5, 2026

Support AMD CPU and Intel GPU
khanhnd61-vr
committed
9dff2d9

Commits on Aug 8, 2026

fix backend init edge cases
anindex
committed
ae2882a
fix multisuite lookup in ci matrix
anindex
committed
3105c83
bump llama.cpp to b10326
anindex
committed
3a1b3ff
relax transformers pin, default docker to sm_89
anindex
committed
b2ba0ee
harden servers against malformed requests
anindex
committed
cd246c9
harden model loading against bad checkpoints
anindex
committed
c26c313
fix evo1 padded state dims, document arch quirks, add unit tests
anindex
committed
1ea4c2b
share the dual vision tower between openvla-oft and vla-adapter
anindex
committed
14458b0
share the DiT time embeddings across gr00t and vla-jepa
anindex
committed
ba07aa6
share the CHW image preprocessing across four archs
anindex
committed
b749323
shorten code comments
anindex
committed
87e4aa1
keep the gguf reader open across predict calls
anindex
committed
f42cf5d
raise cpu thread cap to 16, add VLA_N_THREADS
anindex
committed
06a1493
document VLA_N_THREADS and VLA_DEVICE
anindex
committed
06f5cb1
add stable C ABI and libvla
anindex
committed
87a16a1
add python bindings over the C ABI
anindex
committed
be9c314

Commits on Aug 9, 2026

fail pi0 and pi05 load when the checkpoint is missing weights
anindex
committed
c409654
share the qwen3-vl vision tower between gr00t n1.7 and vla-jepa
anindex
committed
843e619
reuse the compute context and graph allocator across predict calls
anindex
committed
d751c8d
add -hf model fetch, vla-bench, release workflow and contributor docs
anindex
committed
e8aaaa0
default the gr00t n1.7 graph cache on, regenerate the benchmark table
anindex
committed
de62b3d
refresh adoption notes, drop em-dashes from contributor docs
anindex
committed
478ac57
build every test target before ctest
anindex
committed
e63396d
validate checkpoint geometry before sizing buffers
anindex
committed
1059543
stop a stalled peer from parking the servers
anindex
committed
9553cdb
drop the write-only backend flags
anindex
committed
3982914
fix the c api load leak and two doc errors
anindex
committed
eca483d
bump llama.cpp to b10331
anindex
committed
c70fddc
release 0.2.0
anindex
committed
7df64e0
share one graph cache across the ggml archs
anindex
committed
1530eb0
share the time embedding, causal mask and patch count
anindex
committed
eea5323
add --text to vla-cli and build aarch64 binaries
anindex
committed
7409154
refresh the benchmark table and add success rates
anindex
committed
a5f3766