Skip to content

fix(hicache): support legacy SGLang CP context - #367

Merged
ketor merged 1 commit into
mainfrom
fix/hicache-legacy-attn-cp
Aug 31, 2026
Merged

fix(hicache): support legacy SGLang CP context#367
ketor merged 1 commit into
mainfrom
fix/hicache-legacy-attn-cp

Conversation

@ketor

@ketor ketor commented Aug 31, 2026

Copy link
Copy Markdown
Contributor

Problem

xb01 production rolled back the v2.24.3 client after the first Spinfield-managed GLM-5.2 worker crashed in _resolve_parallel_coordinates: its older SGLang ParallelContext exposes attn_cp_size/rank but no attn_dcp_size/rank.

LWS stopped after the first unready worker, limiting impact to one of 17 workers.

Fix

  • preserve the legacy attn_cp_* axis as PCP
  • treat a completely absent independent DCP axis as (1, 0)
  • do not copy the same legacy CP coordinate into both PCP and DCP, which would duplicate key dimensions and alter replica-writer election
  • retain strict validation when only part of an axis is present

Verification

  • added a legacy ParallelContext regression that lacks both DCP fields
  • python3 integration/hicache/tests/test_sg_width_namespace.py: 17/17 passed

The broader unittest discovery also reaches the same 17 passing tests; its unrelated test_client_stats module cannot import locally because this workstation lacks pytest. CI supplies the pinned test environment.

@ketor
ketor added this pull request to the merge queue Aug 31, 2026
Merged via the queue into main with commit 3d490d7 Aug 31, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant