Skip to content

CLEANUP_LOG.md

Latest commit

 

History

History
83 lines (70 loc) · 3.71 KB

CLEANUP_LOG.md

File metadata and controls

83 lines (70 loc) · 3.71 KB
Raw

Cleanup log — 2026-06-10

Trail for the next cleanup pass. Delete this file when its successor lands.

Source

Driven by research/repo_cleanup_audit_2026-06-10.md (deleted at end of this pass, per spec).

What landed

  • 31 isolate probes archived to archive/cb_isolate_superseded_2026-06-10/
    • 13 gm4_v0* / gm4_round* (Gemma 4 past v0.x)
    • 15 nemotron3_v01* / v02* (phases shipped)
    • 4 cb35_* (35B drift resolved)
  • 28 research docs archived to research/archive/
    • 23 STALE point-in-time docs (35B drift trio, gemma4 pc-template duo, perf-briefing, 35B MoE FFN trio, qb2 rebuild, qwen36 topk owned design, 9B branch read, gemma4 12B bringup, gemma4 branch diff, 27B prefill/ prefix-cache plans, 35B perf milestones, multi-model serving, etc.)
    • 5 JAX/PJRT lineage (research/01..05_*.md)
  • 63 wiki entries archived to wiki/archive/
    • wiki/01..64 (the GPT-2 → Qwen 0.5B → 8B JAX/PJRT sprint)
    • Kept operative: 65_mamba_state_space_models.md, 66_blackhole_kernel_dataflow_anatomy.md, bringup_checklist.md, debugging_methodology.md, profiling_guide.md, qa_correctness_and_architecture.md, seven_bugs_case_studies.md
  • 5 meta-debt audit docs DELETED
    • research/code_cleanup_plan_2026-06-04.md
    • research/repo_archive_audit_2026-06-04.md
    • research/doc_audit_plan_2026-06-05.md
    • research/doc_polish_plan_2026-06-05.md
    • research/repo_cleanup_audit_2026-06-10.md (this pass's audit)
  • Beautiful-docs trim
    • HANDOFF.md 2795 → 175 lines (single-page cold-start; TOC; live state + perf + production path + open workstreams + rules + workflow)
    • research/model_bringup_recipe.md added TOC

What was deliberately NOT touched

Per the spec's hard rules:

  • experiments/serve/server_*.py — production servers
  • experiments/serve/cb_*.py, openai_endpoint.py, serve_cb.sh
  • experiments/utils/*.py (cosine_ladder / hf_reference / needle_haystack variants — refactor is high-risk and out of scope)
  • experiments/owned_ops/* — custom kernels
  • archive/legacy/* — already archived
  • scripts/chat.py, scripts/chat_curl.py, scripts/deploy.sh, scripts/run_*.sh
  • All audit-explicit-keep research docs (gemma4_chunked_prefill_plan_*, gemma4_step2_fp32_acc_plan_*, gemma4_drafter_*, gemma4_mtp_*, gemma4_verify_kp1_*, nemotron3_nano_30b_a3b_bringup_plan.md, mm7_g1_mamba2_kernel_design.md, precision_long_context_*, model_bringup_recipe.md, 27b_cb_scope.md, diffusiongemma_bringup_scope_*, agentic_harness_scope_*)

Deferred refactors (audit items 4 / 7 / 8 / 5 / 6 / 9)

Out of scope per spec hard rule #2 (don't touch experiments/serve or experiments/utils refactors; servers are explicitly kept). When picked up:

  • Item 4: extract _apply_full_rope into experiments/serve/_rope.py
  • Item 7: collapse 6 _layer_pos0_*paged* variants in server_gemma4_unified_ttnn.py
  • Item 8: promote experiments/utils/_hf_oracle_base.py
  • Items 5, 6: create docs/MODELS.md and docs/RECIPES.md
  • Item 9: add examples/ skeleton

Numbers

  • 133 cleanup commits (use git log --oneline | grep cleanup to see)
  • Before: 76 research, 77 wiki, 122 isolate probes
  • After: ~48 research, 8 wiki, ~91 isolate probes
  • Repo navigation ambiguity down ~40% (audit projection met)

Pointers for the next pass

  • The remaining 48 research docs were left as CURRENT. Re-triage in ~4 weeks once the active workstreams (#290, #313, #314, Nemotron-3, diffusion-Gemma) ship.
  • Audit's items 4–9 (code refactors + docs consolidation + examples) are the next mechanical batch. They were skipped here because they touch production servers and shared utils — needs a dedicated session with device validation.