Skip to content

fix(grpc): drop stale WatchLoadUpdateReq piggyback, bump sglang floor to 0.5.13 #1850

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

Merged
slin1237 merged 2 commits into from
Jun 28, 2026
Merged

fix(grpc): drop stale WatchLoadUpdateReq piggyback, bump sglang floor to 0.5.13 #1850

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
64230cf
fix(grpc): drop stale WatchLoadUpdateReq piggyback, bump sglang floor…
StarDuster Jun 27, 2026
5a26d32
Merge branch 'main' into fix/grpc-sglang-watch-load-update
StarDuster Jun 27, 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
4 changes: 3 additions & 1 deletion grpc_servicer/pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,9 @@ classifiers = [

[project.optional-dependencies]
vllm = ["vllm>=0.19.0", "pyzmq>=25.0.0", "msgspec>=0.18.0"]
sglang = ["sglang>=0.5.10"]
# 0.5.13 is the first release with the shmem load snapshots that replaced
# the WatchLoadUpdateReq piggyback path (removed upstream in v0.5.14).
sglang = ["sglang>=0.5.13"]

@chatgpt-codex-connector chatgpt-codex-connector Bot Jun 27, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Enforce the SGLang floor outside extras

This dependency floor only applies when users install smg-grpc-servicer[sglang]; the repo's source install paths still install grpc_servicer/ without that extra after SGLang is already present (for example scripts/ci_install_sglang.sh pins sglang[all]==0.5.12.post1, and scripts/installation/install-smg.sh force-reinstalls the servicer source into engine images). In those environments, dp_size>1 with total_tokens/total_requests still needs the WatchLoadUpdateReq feed that was removed, because pre-0.5.13 SGLang's data-parallel controller consumes only that request rather than shmem snapshots, so the new code silently runs with stale DP load state instead of enforcing the unsupported version.

Useful? React with 👍 / 👎.

@junliu-mde junliu-mde Jun 27, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks it's better to bump sglang version in CI?

# smg-grpc-proto>=0.4.7 is the first release that ships mlx_engine_pb2;
# without this floor, installing [mlx] against an older proto build would
# crash at import time when smg_grpc_servicer.mlx.server runs.
Expand Down
12 changes: 0 additions & 12 deletions grpc_servicer/smg_grpc_servicer/sglang/request_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,6 @@
ProfileReqOutput,
TokenizedEmbeddingReqInput,
TokenizedGenerateReqInput,
WatchLoadUpdateReq,
)
from sglang.srt.observability.req_time_stats import (
APIServerReqTimeStats,
Expand Down Expand Up @@ -764,17 +763,6 @@ async def cleanup(request_id):

cleanup_tasks.append(asyncio.create_task(cleanup(rid)))

# Forward load info to DataParallelController for token-aware balancing.
# Mirrors TokenizerManager._handle_batch_output logic: when dp_size > 1,
# each scheduler piggybacks its load (num_reqs, num_tokens) on batch output.
# Without this, DPBudget stays at zero and total_tokens/total_requests
# policies degenerate to always picking rank 0.
if self.server_args.dp_size > 1 and batch_out.load is not None:
try:
await self._send_to_scheduler(WatchLoadUpdateReq(loads=[batch_out.load]))
except Exception as e:
logger.warning("Failed to forward DP load update: %s", e)

# Execute all queue.put() operations in parallel
if put_tasks:
await asyncio.gather(*put_tasks, return_exceptions=True)
Expand Down
Loading