EPS
← All tasks·#772Completed

workflow-fix: stream-reduce activation spans, never materialize all N

kind: infra#wf-fix#wf-fix-fp:997f657054ce
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed from the predictor-line (#742/#761/#763) infra-friction retrospective (subagent, 2026-06-30). Related in-flight: #770 (RunPod no-port wedge billing), #681 (data-disk bind cutover), #667 (GCP frozen-phase), #769 (CUDA-frag OOM).

Goal

Add a stream-reduce rule: an analysis reading N multi-GB per-context .pt blobs from HF MUST use the canonical _HfStreamSpanSource (download->process->delete, peak one ctx) with the _assert_attn_stream_equivalence gate, never download all N locally; cite #761's 220 GB refill.

Workflow gap

  • Bug observed: #761 re-downloaded the full 220 GB UltraChat answer_spans (4.4GB x 50) into a local cache and refilled /, despite the canonical streaming helper _HfStreamSpanSource already existing in scripts/issue658_fit_predictors.py (download->process->delete, peak ~1 ctx) — a built-but-stranded helper not reused by the sibling bootstrap.

Scope / surfaces

  • Target: .claude/rules/code-style.md

Constraints / invariants

  • Workflow-surface only; ruff + workflow_lint pass; add a test where practical.

Provenance

  • workflow_fix_target: .claude/rules/code-style.md
  • fingerprint: 997f657054ce

Raised by the predictor-line infra retrospective (2026-06-30).

Activity