Abstract CUDA hardcodes into configurable te_device_type / te_platform - #3113

Open

lxd-cumt wants to merge 8 commits into

NVIDIA:release_v2.14from

lxd-cumt:cuda_patch

Commits on Jun 17, 2026

fake: try to patch cuda hardcodes

lxd-cumt
authored and
Xianduo Li
committed
cdefc94
remove vendor related

lxd-cumt
authored and
Xianduo Li
committed
e5e85a9
support apply paatches

lxd-cumt
authored and
Xianduo Li
committed
aff2e8a
fix: use transformer_engine_plugin_fl interface and replace assert with explicit raise
Xianduo Li
committed
561efb3
fix: patch cuda hardcodes for plugin compatibility
Xianduo Li
committed
8ae24c3

Commits on Jul 7, 2026

fix: move torch-dependent globals to pytorch/__init__.py and remove redundant code
Xianduo Li
committed
7529fba
fix: use NVTE_PLUGIN env var instead of hardcoded plugin module name
Xianduo Li
committed
b6f31a8

Commits on Aug 19, 2026

Simplify plugin device/platform interface and fix minor issues
Xianduo Li
committed
bccb702