Skip to content

[PD] Opt-in Core transfer path for NIXL — review MVP #1

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

Draft
samnordmann wants to merge 2 commits into
base: snordmann/transfer-review-base-20260917
Choose a base branch
from
Draft

[PD] Opt-in Core transfer path for NIXL — review MVP #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4b8c71b
[PD] Add PyTorch transfer and lease-lifetime review prototype
samnordmann Sep 17, 2026
db82eef
[PD] Reduce Core transfer experiment to an opt-in hook
samnordmann Sep 17, 2026
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
5 changes: 5 additions & 0 deletions python/sglang/srt/disaggregation/nixl/conn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,6 +450,11 @@ def __init__(
sync_mode=nixl_thread_sync_t.NIXL_THREAD_SYNC_STRICT,
)
self.agent = nixl_agent(str(uuid.uuid4()), agent_config)
if envs.SGLANG_DISAGGREGATION_NIXL_USE_TORCH_TRANSFER.get():
from nixl.torch_transfer import TorchTransferAgent

# The manager's allocation lifetime contract remains unchanged.
self.agent = TorchTransferAgent(self.agent, owner=self)
if num_threads > 0:
# TODO: Remove this once NIXL passes thread parameters from
# nixl_agent_config to explicitly-created backends.
Expand Down
13 changes: 13 additions & 0 deletions python/sglang/srt/disaggregation/nixl/torch_transfer_mvp.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,13 @@
# Core transfer experiment

Set `SGLANG_DISAGGREGATION_NIXL_USE_TORCH_TRANSFER=1` on test workers. Requires
the PyTorch `_transfer` prototype and NIXL `torch_transfer` provider. Default: off.

The NIXL agent is wrapped after creation, retaining its configured backends and
strict thread synchronization. Registration, prepared indexed WRITE, fallback
transfers and notifications route through Core. Bootstrap, allocation ownership,
CUDA ordering, staging and abort protocols are unchanged.

This is a **success-path review prototype**, not a production mode. The manager
must keep raw-pointer allocations stable. Active cancellation and failure recovery
are outside the MVP. GPU serving/staging and throughput remain unvalidated.
1 change: 1 addition & 0 deletions python/sglang/srt/environ.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -668,6 +668,7 @@ class Envs:
# A wedged RDMA stack fails startup here instead of at the scheduler watchdog.
SGLANG_DISAGGREGATION_ENGINE_INIT_TIMEOUT = EnvInt(60)
SGLANG_DISAGGREGATION_NIXL_BACKEND = EnvStr("UCX")
SGLANG_DISAGGREGATION_NIXL_USE_TORCH_TRANSFER = EnvBool(False)
SGLANG_DISAGGREGATION_NIXL_BACKEND_PARAMS = EnvStr("{}")
SGLANG_DISAGG_PREFILL_EARLY_SEND_CACHED_PREFIX = EnvBool(True)
SGLANG_DISAGGREGATION_ZMQ_MAX_SOCKETS = EnvInt(16384)
Expand Down
Loading