3.2 KiB

Raw Blame History

M01 CI Run 3 — Stub Repositories (Iterative)

Date: 2026-03-08
Branch: m01-ci-truthfulness
Trigger: Stub repository approach (ac965561 → f013e553)

1. Workflow Identity

Workflow	Latest Run	Status
Linter	22814396752	✓ success
Tests	22814850488	Partial (server ✓, img2img/txt2img 500)

2. Approach

Stub repositories instead of cache/submodules:

scripts/dev/create_stub_repos.py creates minimal repositories/ layout
Satisfies paths.py assertion and import chain
Deterministic, no network, no cloning

3. Stub Progression (by run)

Blocker	Fix
`paths.py` assert (ddpm.py)	Add ddpm.py stub
`LatentDiffusion` import	Add LatentDiffusion, LatentDepth2ImageDiffusion
`ldm.util`	Add ldm.util.default
`ldm.modules.midas`	Add ldm.modules.midas
`ldm.modules.distributions`	Add distributions, DiagonalGaussianDistribution
`ldm.modules.diffusionmodules.openaimodel`	Add openaimodel, explicit init imports
`sgm.models`, `sgm.modules.*`	Add sgm stubs (attention, diffusionmodules, encoders, conditioner)
`k_diffusion.external`, `k_diffusion.utils`	Add external, utils, sampling stubs
`k_diffusion.sampling.get_sigmas_*`	Add get_sigmas_karras, exponential, polyexponential

4. Current Stub Layout

repositories/
  stable-diffusion-stability-ai/
    ldm/
      models/diffusion/ddpm.py
      util.py
      modules/
        __init__.py (imports distributions, diffusionmodules)
        encoders/modules.py
        attention/, diffusionmodules/ (model, openaimodel)
        distributions/distributions.py (DiagonalGaussianDistribution)
        midas/
  generative-models/
    sgm/
      models/diffusion/ (DiffusionEngine)
      modules/ (attention, diffusionmodules, encoders, conditioner)
  k-diffusion/
    k_diffusion/
      __init__.py (imports utils, sampling, external)
      utils.py, external.py, sampling.py
  BLIP/
  stable-diffusion-webui-assets/

5. Dynamic Stub Approach (Commit `5a76c617`)

Replaced manual file-by-file stubs with dynamic stub modules:

_StubFinder (MetaPathFinder): catches any ldm.* or sgm.* import the default finder misses
_StubLoader: creates _StubModule with __path__, __file__ for inspect
_StubModule: resolves attributes as submodules, stub classes, or dicts (UPPER_CASE)
Stub classes: metaclass returns no-op for missing attrs, {} for ATTENTION_MODES
UPPER_CASE on modules → dict (e.g. midas.api.ISL_PATHS)
__file__ = <stub:name> so inspect.getfile works
Keeps ddpm.py for paths.py assertion; DDPM, LatentDiffusion with __init__(*a,**k)
Keeps k_diffusion file-based (needs real get_sigmas_*, torch, etc.)

Eliminates whack-a-mole import chain.

6. Run 4 — Closeout Verification

Trigger: Milestone closeout commit 0bd566f5 (M01_summary, M01_audit, M02_plan, ledger update).

Workflow	Run ID	Status
Linter	22815773475	✓ success
Tests	22815773481	Partial (17 pass, 18 fail)

Closeout verification run. No functional changes. CI remains consistent with Run 3.

3.2 KiB Raw Blame History