Refactor - #21

Open

khanhnd61-vr wants to merge 21 commits into

mainfrom

refactor

Commits on Aug 13, 2026

give the cuda extension first refusal on fused add/mul so bf16 activations stop aborting
khanhnd61-vr
committed
ad9a56b
address bf16 elementwise rows by block index instead of a per-element divide
khanhnd61-vr
committed
0daf584
make the llama.cpp tag and server binary overridable so regressions can be bisected across builds
khanhnd61-vr
committed
3e802de
read boolean env switches by value so setting one to 0 turns it off
khanhnd61-vr
committed
1efb7a3
widen the bf16 elementwise path to eight elements per thread on contiguous rows
khanhnd61-vr
committed
4d29e95
split the model tree into layer, module and model levels
khanhnd61-vr
committed
87a7e8f
make precision and attention runtime options instead of per-arch env switches
khanhnd61-vr
committed
b5fb71a
fold the qwen3-vl tower's private rope and attention onto the layer primitives
khanhnd61-vr
committed
f787c6e
put gr00t n1.7 and vla-jepa on the shared tower, backbone and dit modules
khanhnd61-vr
committed
69e908c
share the siglip tower and gemma stack between pi0 and pi0.5
khanhnd61-vr
committed
3db71d6
give the dinov2+siglip dual tower a module that both archs declare through
khanhnd61-vr
committed
2a41e1a
route evo1 and bitvla weight loading through the shared loader
khanhnd61-vr
committed
c3b8982
share the siglip block weight struct with smolvla
khanhnd61-vr
committed
0b46154
drop the weight structs gr00t n1.7 no longer defines locally
khanhnd61-vr
committed
dfdb585
one statement per line across src
khanhnd61-vr
committed
8f70e60
tighten spacing around arithmetic operators across src
khanhnd61-vr
committed
9678dc4
tighten for-header spacing across src
khanhnd61-vr
committed
c45fc17
move the eval and ci harnesses onto the runtime option flags
khanhnd61-vr
committed
04df961

Navigation Menu

Uh oh!

Refactor - #21

Refactor#21
khanhnd61-vr wants to merge 21 commits into
mainfrom
refactor

Commits on Aug 13, 2026

give the cuda extension first refusal on fused add/mul so bf16 activations stop aborting

address bf16 elementwise rows by block index instead of a per-element divide

make the llama.cpp tag and server binary overridable so regressions can be bisected across builds

read boolean env switches by value so setting one to 0 turns it off

widen the bf16 elementwise path to eight elements per thread on contiguous rows

split the model tree into layer, module and model levels

make precision and attention runtime options instead of per-arch env switches

fold the qwen3-vl tower's private rope and attention onto the layer primitives

put gr00t n1.7 and vla-jepa on the shared tower, backbone and dit modules

share the siglip tower and gemma stack between pi0 and pi0.5

give the dinov2+siglip dual tower a module that both archs declare through

route evo1 and bitvla weight loading through the shared loader

share the siglip block weight struct with smolvla

drop the weight structs gr00t n1.7 no longer defines locally

one statement per line across src

tighten spacing around arithmetic operators across src

tighten for-header spacing across src

move the eval and ci harnesses onto the runtime option flags

Commits on Aug 14, 2026

revert the harness commit that swept untracked docs and eval trees into git

move the tracked eval and ci harnesses onto the runtime option flags

keep model.h free of ggml so the pure unit tests build standalone