[Qwen4Exp] Support UVA PLE-offload and Engram tensor parallelism (#54371)

Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Harry Huang
2026-09-09 07:32:34 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 62f3bf58a5
commit 3116c5d06b
16 changed files with 1712 additions and 495 deletions
@@ -47,7 +47,8 @@ class CudaCommunicator(DeviceCommunicatorBase):
global_world_size,
use_all2all=use_all2all,
)
if "tp" not in unique_name:
# Match the group name exactly so ETP does not enable TP-only backends.
if unique_name.split(":")[0] != "tp":
# custom allreduce or torch symm mem can be used only by tp
use_custom_allreduce = False
use_torch_symm_mem = False