[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
+10 -3
View File
@@ -4,7 +4,7 @@
import inspect
import warnings
from contextlib import contextmanager
from contextlib import contextmanager, nullcontext
from typing import Any
import torch
@@ -120,11 +120,18 @@ def process_weights_after_loading(
"does not support pre-processed weights"
)
# When quant methods need to process weights after loading
# (for repacking, quantizing, etc), they expect parameters
# (for repacking, quantizing, etc), they typically expect parameters
# to be on the global target device. This scope is for the
# case where cpu offloading is used, where we will move the
# parameters onto device for processing and back off after.
with device_loading_context(module, target_device):
# Methods that can process weights in place (e.g. PLE scale
# validation) set requires_device_loading=False to skip this move.
loading_context = (
device_loading_context(module, target_device)
if quant_method.requires_device_loading
else nullcontext()
)
with loading_context:
quant_method.process_weights_after_loading(module)
# process_weights_after_loading may swap in freshly-created
# Parameters (e.g. FP8 requantization), which are stamped with the