Skip to content

Enable the shared-memory MPMD e2e test. #844

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
copybara-service wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Enable the shared-memory MPMD e2e test. #844

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 3 additions & 15 deletions tpu_sync/api/torch/kv_cache_store_mpmd_e2e_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,6 @@
import subprocess
import sys
import time
import unittest

_LOG_DIR = os.environ.get("TEST_TMPDIR", os.environ.get("TMPDIR", "/tmp"))
os.environ.setdefault("TPU_LOG_DIR", _LOG_DIR)
Expand Down Expand Up @@ -865,20 +864,9 @@ def test_mpmd_8rank_e2e_save_and_load_with_slices(self):
self.fail("One or more workers failed!")

# The same 8-rank save/load/compare run with the KV pools (and metadata
# tables) placed in shared memory. Every rank runs in its own process on the
# same host, so this is the configuration that needs per-rank segment names
# and one real segment per allocation; today the ranks collide on one name
# and the pools are additionally left unregistered with the DMA engine on
# multi-process TPUv7 (the DmaMap workaround). Enable only after the
# shared-memory allocator rework AND the libtpu release that fixes
# multi-process DmaMap (with the workaround removed) — the run is not
# production-representative before both.
@unittest.skip(
"Shared-memory segments collide across same-host MPMD ranks and the"
" allocator aliases every allocation into one segment; also requires"
" the libtpu multi-process DmaMap fix. Enable after the allocator"
" rework and the libtpu pin bump."
)
# tables) placed in shared memory. Every rank runs in its own process on
# the same host, so this exercises per-rank segment names, one region per
# allocation within a segment, and true multi-process DMA mapping.
def test_mpmd_8rank_e2e_save_and_load_in_shared_memory(self):
world_size = 8
prepare_tpu_environment(world_size)
Expand Down
Loading