Skip to content

[Wave 3] Medium — Module sweep: Training orchestration, multi-stage pipelines, resume & no-train modes (part 2/2) #572

Open

Description

@cemililik

Wave 3 · 2 finding(s) in 1 fix package(s). Fix these whenever a pull request touches this module ("clean the module you touch").

Findings

  • CR-A04-050 (code, Low) — Dry-run 'dora' field reads the deprecated use_dora flag and reports false for lora.method: dora
    severity Low · reachability default · intent contract-violation · confidence likely · independently confirmed · reproduced end to end · already tracked: KNOWN-OPEN C31-RUN-OUTCOME — docs/roadmap/phase-16-trust-surface-hardening.md:522-525 (S8 open)
    Failure scenario. An operator verifying that DoRA is enabled via dry-run (text or JSON envelope) is told it is not.
    Where. forgelm/cli/_dry_run.py:53 (confirmed), forgelm/config.py:390-439, forgelm/model.py:238-240, docs/usermanuals/en/reference/json-output.md:100
    Recommended fix. Use model._resolve_peft_flags (or report lora.method).

  • ↳ same root cause: DOC-D3B03-011 (docs, Medium) — docs/usermanuals/en/reference/json-output.md: wrong detail — dora is the effective value of the corresponding config flag.
    severity Medium · already tracked: KNOWN-OPEN C31-RUN-OUTCOME docs/roadmap/phase-16-trust-surface-hardening.md:522-525
    Where. docs/usermanuals/en/reference/json-output.md:100 (Turkish mirror: docs/usermanuals/tr/reference/json-output.md:100) · type: wrong detail
    The document says. dora is the effective value of the corresponding config flag.
    What the code does. The dry-run reads only the deprecated lora.use_dora. With the canonical lora.method: "dora", the dry-run printed "dora": false (I ran it), while training applies DoRA (use_dora = config.lora.use_dora or peft_method == 'dora').
    Code references. forgelm/cli/_dry_run.py:53, forgelm/model.py:238
    Proposed correction. Fix the code: "dora": config.lora.use_dora or config.lora.method == "dora". Optionally also expose lora.method in the envelope.
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

Done when

  • Every finding above is fixed or closed with a reason (tick it here).
  • Code fixes carry a test; doc fixes update the Turkish mirror in the same PR.

Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.

Activity

  1. added
    bugSomething isn't working
    documentationImprovements or additions to documentation
    backlog: knownAlready tracked in docs/roadmap
    severity: mediumReal defect with limited reach or a workaround
    source: review-2026-10Found by the October 2026 code and documentation review
    wave: 3Roadmap wave 3: medium severity, module sweeps
    on Oct 9, 2026
  2. cemililik commented on Oct 11, 2026

    @cemililik
    CollaboratorAuthor

    Where to start

    Thanks for taking a look. Please confirm the issue is unassigned and comment before starting so work does not overlap.

    Start with these files on development:

    Make dry-run dora agree with the effective lora.method/use_dora resolution used by model loading. Add a parametrized regression covering method=dora, ordinary LoRA and the supported legacy flag; assert the JSON result without loading a model.

    Run the focused tests:

    pytest tests/test_cli.py --no-cov

    Run these guards/checks as applicable to the changed files:

    ruff format --check forgelm/cli/_dry_run.py forgelm/model.py tests/test_cli.py
    ruff check forgelm/cli/_dry_run.py forgelm/model.py tests/test_cli.py

    If user-facing documentation changes, update the existing Turkish mirror in the same PR; keep commands, code and terminal output in English and structurally aligned. English-only contributor standards need no new mirror.

    Open the PR against development with Fixes #572 once every finding is resolved or explicitly closed with a reason; use Refs #572 for partial work and list the resolved finding IDs.

  3. cemililik commented on Oct 15, 2026

    @cemililik
    CollaboratorAuthor

    I'm removing good first issue here: this finding is tracked by Phase 16 step S8 (#894), which lands as one root-cause family with its own review rounds (see "Roadmap steps" in CONTRIBUTING.md), so a separate fix could collide with it. Please don't start on it without coordinating on #894 first.

  4. added
    blockedWaits for another issue (see Blocked by); set and removed automatically, never by hand
    on Oct 20, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: pipelineCode area: pipelinebacklog: knownAlready tracked in docs/roadmapblockedWaits for another issue (see Blocked by); set and removed automatically, never by handbugSomething isn't workingdocumentationImprovements or additions to documentationseverity: mediumReal defect with limited reach or a workaroundsource: review-2026-10Found by the October 2026 code and documentation reviewwave: 3Roadmap wave 3: medium severity, module sweeps

    Type

    No type

    Projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions