Skip to content

Bump sglang to v0.5.15 - #1655

Merged
guapisolo merged 11 commits into
mainfrom
bump_sglang_v0.5.15
Jul 16, 2026
Merged

Bump sglang to v0.5.15#1655
guapisolo merged 11 commits into
mainfrom
bump_sglang_v0.5.15

Conversation

@yueming-yuan

@yueming-yuan yueming-yuan commented Jul 13, 2026

Copy link
Copy Markdown
Collaborator

ci-image-tag: bump-sglang-v0.5.15-test
ci-sglang-pr: sglang-miles-v0.5.15

Summary

Bump sglang-miles from v0.5.14 → v0.5.15.

  • docker/Dockerfile: SGLANG_IMAGE_TAG v0.5.14 → v0.5.15, SGLANG_BRANCH sglang-miles → sglang-miles-v0.5.15
  • docker/build.py (cu12-x86): SGLANG_IMAGE_TAG v0.5.12-cu129 → v0.5.15-cu129

sglang branch

sglang-miles-v0.5.15 (24 commits rebased onto tag v0.5.15). Notable: dropped #26980 (already upstream in v0.5.15); folded #27750 into the weight-update-sessions commit; kept v0.5.15's upstreamed ComparableWeight comparator incl. the #29623 ue8m0 scale-unpack fix.

Testing

Docker build via docker-build.yml workflow_dispatch (variant cu13-x86): https://github.com/radixark/miles/actions/runs/29290296331

Notes

  • miles wheels (yueming-yuan/miles-wheels) left at v0.5.12; building as-is — will revisit if ABI-incompatible with v0.5.15.
  • Post-bump CI failures (transformers/torch/CUDA deltas) expected; will fix as follow-up PRs onto this branch.

- Dockerfile: SGLANG_IMAGE_TAG v0.5.14 -> v0.5.15, SGLANG_BRANCH
  sglang-miles -> sglang-miles-v0.5.15
- build.py cu12-x86: SGLANG_IMAGE_TAG v0.5.12-cu129 -> v0.5.15-cu129
gemini-code-assist[bot]

This comment was marked as resolved.

…_update, pull_weights, load_lora_adapter_from_distributed endpoints (v0.5.15 FastAPI)
…in/End/PullWeightsReqOutput) in weight_updater
The final-rollout forced save (should_run_periodic_action) writes a
~220GB megatron dist checkpoint that dominates the 2250s CI timeout
(write workers throttled in balance_dirty_pages on overlayfs). Nothing
in the smoke test consumes the checkpoint; save/load coverage lives in
dedicated tests. save_interval=None disables all saves incl. the forced
final one.

Also: fixture Pinaster/GLM-5.2_5layer config.json fixed upstream (HF
commit 1c74913) — mlp_layer_types/indexer_types trimmed to 5 entries for
transformers>=5.12 validate_layer_type.
…obal

sglang >= 0.5.15 moved server-args storage into runtime_context;
get_global_server_args() no longer reads the legacy module-private
_global_server_args that p2p._create_cpu_replica was writing. The
trainer-side CPU replica build (get_model -> get_rope ->
is_true_on_policy_enabled) then raised 'Global server args is not set
yet!'. set_global_server_args_for_scheduler() is the public setter on
both v0.5.14 (writes the module global) and v0.5.15 (publishes into
runtime_context).

@guapisolo guapisolo left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

tiny changes. LGTM

@guapisolo
guapisolo merged commit eb17279 into main Jul 16, 2026
36 checks passed
@guapisolo
guapisolo deleted the bump_sglang_v0.5.15 branch July 16, 2026 18:55
@yueming-yuan yueming-yuan mentioned this pull request Jul 25, 2026
2 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants