Skip to content

Ko3n1g/pranav/ray k8s issue debug #562

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

Closed
ko3n1g wants to merge 5 commits into from
Closed

Ko3n1g/pranav/ray k8s issue debug #562

Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
bfc9c0c
Temp fix for k8s issue
pthombre Jan 14, 2026
e3b0018
fix: Pin master addr
ko3n1g Jan 19, 2026
51af26a
add delay between workers
ko3n1g Jan 19, 2026
4b86332
socket
ko3n1g Jan 19, 2026
7ed9663
cleanup
ko3n1g Jan 22, 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
25 changes: 22 additions & 3 deletions nemo_deploy/llm/megatronllm_deployable_ray.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@
import logging
import os
import random
import socket
import time
from typing import Any, Dict, Optional

Expand Down Expand Up @@ -66,6 +67,12 @@ def __init__(
**model_config_kwargs,
):
# Use replica-specific environment variables to avoid conflicts
master_addr = "127.0.0.1"
Copy link
Copy Markdown
Contributor

@pthombre pthombre Jan 20, 2026

Choose a reason for hiding this comment

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

What is the need to hard code the master address and port here?

with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.bind(("", 0)) # Bind to port 0 lets OS pick a free port
new_port = s.getsockname()[1]
self.master_port = str(new_port)

os.environ["MASTER_PORT"] = master_port
# All ranks must use the SAME MASTER_ADDR (rank 0 node IP)
os.environ["MASTER_ADDR"] = master_addr if master_addr else ray._private.services.get_node_ip_address()
Expand All @@ -82,6 +89,11 @@ def __init__(
LOGGER.info(f"Replica {replica_id} - MASTER_PORT: {os.environ['MASTER_PORT']}")
LOGGER.info(f"Replica {replica_id} - MASTER_ADDR: {os.environ['MASTER_ADDR']}")

if rank != 0:
sleep_time = 5 + rank # Rank 1 waits 6s, Rank 2 waits 7s, etc.
LOGGER.info(f"Worker {rank}: Sleeping {sleep_time}s to avoid JIT lock contention...")
time.sleep(sleep_time)

try:
self.model = MegatronLLMDeployableNemo2(
nemo_checkpoint_filepath=nemo_checkpoint_filepath,
Expand Down Expand Up @@ -192,9 +204,8 @@ def __init__(
deployment_node_id = node.get("NodeID")
break

rank_0_worker = ModelWorker.options(
scheduling_strategy=NodeAffinitySchedulingStrategy(node_id=deployment_node_id, soft=False)
).remote(
# Common arguments for rank 0 worker
rank_0_kwargs = dict(
nemo_checkpoint_filepath=nemo_checkpoint_filepath,
rank=0,
world_size=num_gpus,
Expand All @@ -216,6 +227,14 @@ def __init__(
micro_batch_size=micro_batch_size,
**model_config_kwargs,
)

# Use node affinity if we found a matching node, otherwise use default scheduling
if deployment_node_id is not None:
rank_0_worker = ModelWorker.options(
scheduling_strategy=NodeAffinitySchedulingStrategy(node_id=deployment_node_id, soft=True)
).remote(**rank_0_kwargs)
else:
rank_0_worker = ModelWorker.remote(**rank_0_kwargs)
worker_futures.append(rank_0_worker)

# Wait for rank 0 to start before creating other workers
Expand Down