[Bugfix] LoRA for DeepSeek V3.2 (#35077)

Signed-off-by: Hollow Man <hollowman@opensuse.org>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟
2026-04-22 19:33:50 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Jee Jee Li
parent 04eac6ba24
commit a250f1bd5f
13 changed files with 614 additions and 37 deletions
+7 -12
View File
@@ -1620,18 +1620,13 @@ def _parallel_worker(
else:
print("F", end="")
finally:
# Note: for some reason DeepEP buffers don't seem to be
# entirely reusable on B200. In order to work around this
# we clear the all2all manager's cache after each testpoint.
cap = current_platform.get_device_capability()
if (
cap is not None
and cap.major == 10
and (
test_config.backend == "deepep_low_latency"
or test_config.backend == "deepep_high_throughput"
)
):
# DeepEP managers are not reliably reusable across many subtests in
# a single worker process. Tear them down after each DeepEP case so
# later subtests do not inherit stale communication state.
if test_config.backend in {
"deepep_low_latency",
"deepep_high_throughput",
}:
torch.accelerator.synchronize()
all2all_manager = get_ep_group().device_communicator.all2all_manager
if all2all_manager is not None: