fix(disagg): support per-GPU JSON mapping in --disaggregation-ib-device#9
Open
DavidBellamy wants to merge 1 commit intollm360-mainfrom
Open
fix(disagg): support per-GPU JSON mapping in --disaggregation-ib-device#9DavidBellamy wants to merge 1 commit intollm360-mainfrom
DavidBellamy wants to merge 1 commit intollm360-mainfrom
Conversation
Port of upstream PR sgl-project#23003 applied directly on llm360-main so the octopus-merge into deploy picks it up for tonight's agentic RL pilots. Blocking fix: sglang PD disaggregation currently rejects the launcher's rail-mapping JSON path. Adds a 2-line early return in _validate_ib_devices() that passes JSON content ({...}) and .json file paths through unchanged; get_ib_devices_for_gpu() in mooncake_transfer_engine.py already handles those formats downstream.
DavidBellamy
added a commit
that referenced
this pull request
Apr 19, 2026
…project#23003 Upstream PR sgl-project#23003 (fix/json-ib-device-passthrough) is superseded by #9 (fix/json-ib-device-passthrough-llm360-main), which mirrors the same 7-line patch on top of llm360-main. The upstream branch was based on sgl-project/sglang:main and line-drifts when merged into llm360-main — was triggering a merge conflict in every octopus run after the PR #10 decouple landed. Port harbor's SKIP_UPSTREAM_BRANCHES pattern: env var + jq `NOT_SKIPPED` filter applied to both UPSTREAM_PR_STATE/BRANCHES (cross-fork PRs on LLM360/sglang:llm360-main) and REAL_UPSTREAM_PR_STATE/BRANCHES (upstream sgl-project/sglang PRs). Removes the skipped branch cleanly from the state fingerprint too, so re-running doesn't see it as a pending change.
DavidBellamy
added a commit
that referenced
this pull request
Apr 19, 2026
…project#23003 (#11) Upstream PR sgl-project#23003 (fix/json-ib-device-passthrough) is superseded by #9 (fix/json-ib-device-passthrough-llm360-main), which mirrors the same 7-line patch on top of llm360-main. The upstream branch was based on sgl-project/sglang:main and line-drifts when merged into llm360-main — was triggering a merge conflict in every octopus run after the PR #10 decouple landed. Port harbor's SKIP_UPSTREAM_BRANCHES pattern: env var + jq `NOT_SKIPPED` filter applied to both UPSTREAM_PR_STATE/BRANCHES (cross-fork PRs on LLM360/sglang:llm360-main) and REAL_UPSTREAM_PR_STATE/BRANCHES (upstream sgl-project/sglang PRs). Removes the skipped branch cleanly from the state fingerprint too, so re-running doesn't see it as a pending change.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Port of upstream PR sgl-project#23003 applied directly on llm360-main so tonight's octopus-merge into deploy picks it up for the agentic RL pilots. Upstream PR has been open since 2026-04-16 with no reviews (Gemini quota).
Blocking fix
Current llm360-main sglang rejects the launcher's rail-mapping JSON path:
```
ValueError: Invalid IB devices specified: ['/mnt/weka/shrd/k2pta/rl360/rail_mapping_JOBID.json'].
Available devices: ['mlx5_0', ..., 'mlx5_7']
```
Observed on pilot job 1565856; will recur on 1565857 and 1565858 with the same config.
Patch
Adds a 2-line passthrough in `_validate_ib_devices()` for JSON content (`{...}`) and `.json` paths. `get_ib_devices_for_gpu()` in `mooncake_transfer_engine.py` already handles these formats downstream — the validator was just rejecting them prematurely.