B200 M-bridge misconfig

[srivatsankrishnan]

Summary

Misconfiguration in Qwen TOML file for B200 system found during VER. This PR fixes it.
RM: https://redmine.mellanox.com/issues/4840286

Test Plan

CI/CD
Internal B200 Cluster

Additional Notes

[coderabbitai]

📝 Walkthrough

Walkthrough

A configuration parameter in the Megatron bridge test setup for Qwen 30B model was updated. The gpus_per_node value increased from 4 to 8, while the total num_gpus count remains unchanged at 8.

Changes

Cohort / File(s)	Change Summary
Megatron Bridge GPU Configuration conf/experimental/megatron_bridge/test/b200/megatron_bridge_qwen_30b.toml	Updated gpus_per_node from 4 to 8 in the [cmd_args] section

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

Poem

Eight GPUs now dance on each node so bright,
The Qwen takes flight with computational might,
From four to eight, a simple reconfig tweak,
This bunny approves of changes so sleek! 🐰✨

🚥 Pre-merge checks | ✅ 2

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'B200 M-bridge misconfig' is related to the changeset and refers to fixing a misconfiguration in the B200 Megatron bridge configuration file, which is the primary change.
Description check	✅ Passed	The description is clearly related to the changeset, explaining that it fixes a misconfiguration in the Qwen TOML file for the B200 system found during VER with a reference to the tracking issue.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[greptile-apps]

Greptile Summary

Fixed misconfiguration in B200 Qwen 30B configuration file where gpus_per_node was incorrectly set to 4 instead of 8.

• Corrected gpus_per_node from 4 to 8 to properly match the B200 system architecture (8 GPUs per node)
• This aligns with num_gpus = 8 (single node with 8 GPUs)
• The fix ensures proper GPU resource allocation for Megatron-Bridge workloads on B200 systems

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The change corrects a clear misconfiguration by updating a single parameter value to match the expected B200 hardware configuration. The fix aligns gpus_per_node with the system's actual 8 GPUs per node and is consistent with the num_gpus setting.
• No files require special attention

Important Files Changed

Filename	Overview
conf/experimental/megatron_bridge/test/b200/megatron_bridge_qwen_30b.toml	Fixed gpus_per_node from 4 to 8 to match B200 system configuration with 8 GPUs per node

Sequence Diagram

sequenceDiagram
    participant User
    participant CloudAI
    participant SlurmSystem
    participant MegatronBridge
    participant B200Cluster
    
    User->>CloudAI: Load config: megatron_bridge_qwen_30b.toml
    CloudAI->>CloudAI: Parse cmd_args
    Note over CloudAI: gpus_per_node = 8 (fixed)<br/>num_gpus = 8<br/>gpu_type = "b200"
    CloudAI->>MegatronBridge: Generate launcher command
    MegatronBridge->>MegatronBridge: Validate GPU configuration
    Note over MegatronBridge: Expects 8 GPUs per node<br/>for B200 systems
    MegatronBridge->>SlurmSystem: Submit batch job (1 node, 8 GPUs)
    SlurmSystem->>B200Cluster: Allocate resources
    Note over B200Cluster: Single node with<br/>8 B200 GPUs
    B200Cluster->>SlurmSystem: Resources allocated
    SlurmSystem->>MegatronBridge: Job ID
    MegatronBridge->>CloudAI: Job submitted successfully
    CloudAI->>User: Training started on B200

Loading