Bootstrap scmp_diffusion from scmp_llm

[heroarmor]

Initial bootstrap of `scmp_diffusion` from the Diffusion / Q-DiT / evaluation code that lived in the private `scmp_llm` staging repo. Consumes `scmp_kernels` PR #2 via git submodule.

Architecture

```
scmp_diffusion/ ← this PR
├── scmp_kernels/ ← git submodule → CrucibleComputingGroup/scmp_kernels
├── diffusion/ gaussian_diffusion + sampler (5 files, verbatim)
├── models/ DiT models + Inception evaluator wrapper (8 files, verbatim)
├── qdit/ Q-DiT quantisation + sc_integration (15 files, imports rewired)
├── scripts/ calibration / sweep / SLURM launchers (60 files)
├── tests/ unit tests (1 file)
├── evaluation/ FID / KID / sample-grid mosaics, ImageNet-256 ref-batch prep (10 files)
├── tools/ noise-model calibration utility (1 file)
└── utils/ download + logging (3 files)
```

`scmp_kernels` is a submodule pinned to its `add-mp-module` branch tip at SHA `fa6b5cd`. Once PR #2 lands on `CrucibleComputingGroup/scmp_kernels:main`, the submodule pin and URL will be flipped to org/main in a follow-up commit.

Commits (5)

1. `chore: add scmp_kernels submodule + .gitignore` — bootstraps the submodule at `./scmp_kernels`, URL pointing at `heroarmor/scmp_kernels.git` (the fork with PR #2 in flight). Standard repo-hygiene .gitignore.

2. `import: Q-DiT application code with imports rewired to scmp_kernels` (93 files, +16,750/−1) — brings in `qdit/`, `diffusion/`, `models/`, `scripts/`, `tests/`, `utils/` from `scmp_llm/Q-DiT/`. All SC/MP-related bare imports rewired:

• `from sc_triton import ...` → `from scmp_kernels.sc import sc_matmul, ...`
• `from sng/config_helpers/mp_config import ...` → `from scmp_kernels.{sc.sng, sc.config_helpers, mp} import ...`
• Removed every `sys.path.insert(..., 'SC')` shim.
• Deleted dead `qdit/sc_integration/sc_matmul.py` (vestigial — `sc_matmul_qk` had zero callers).
• `qdit/sc_integration/mp_config.py` is now a thin re-export shim of `scmp_kernels.mp` (local relative imports inside the package keep resolving).
• Excluded artifacts: `pycache/`, `build/`, `Q_DiT.egg-info/`, `.nsys-rep`, `.png`, `logs_mp_sweep/`, `results/`, 92 MB Inception `.pb` checkpoint.

3. `refactor: route all SC matmul calls through the unified sc_matmul dispatcher` (10 files, +206/−450) — `sc_attention.py` (4 `get*_fn` dispatchers collapsed to one), `sc_mlp.py`, `noise_matmul.py` (4 adapter functions collapsed to one matching `sc_matmul`), 4 scripts + 1 test. Every call site now uses `sc_matmul(granularity=...)`. 53 distinct call sites verified to cover all 3 granularities + grouped + chunked + per-head variants.

4. `remove sc_enable=False legacy path — enable-signal is canonical` (21 files, +46/−91) — team policy: `sc_enable=True` is the only supported mode (`scmp_kernels` intentionally dropped packed-XNOR/AND in favour of enable-signal table-lookup). This commit:

• Drops the `sc_enable` ctor arg + attribute from `SCController`.
• Removes 8 `if self.sc_controller.sc_enable` gates in `sc_attention.py` and 1 in `sc_mlp.py`.
• Drops `--sc_enable` from CLI (`quant_sc_main.py`) and from 15 shell-script launchers.
• Drops `sc_enable=...` from `SCController` construction in `sc_modelutils.py`.
• Catches up 3 legacy methods in `sc_attention.py` that the earlier rewrite had missed (still had positional `max/min` args — would have crashed at runtime).

5. `migrate Gap 3 application-side aux code from scmp_llm` (11 files, +931/−1) — application-side tooling that lived in `scmp_llm/SC/`, `scmp_llm/evaluation/`, `scmp_llm/imagenet256_ref/`:

• `tools/calibrate_noise_model.py` ← `SC/noise_model_calibration.py` (calibrates the surrogate consumed by `qdit/sc_integration/noise_matmul.py`).
• `evaluation/{kid, build_full_mosaic, build_sample_grids, compare_images}.py` ← `evaluation/*.py`.
• `evaluation/imagenet_ref/{extract, parallel_npz, compute_fid_kid, compare_grid}.py + run_openai_eval{,_v2}.sh` ← `imagenet256_ref/*`.
• `.gitignore` extended for the 1.9 GB `VIRTUAL_imagenet256_labeled.npz` and 1.2 GB ImageNet `images/` artifacts that travel out-of-band.

Legacy code from `scmp_llm/SC/` (CPU reference `sc.py`/`sc_enable.py`, bench/compare scripts, DSE, in-file test helpers tied to packed-XNOR) is intentionally not migrated, per Allenjin123's "no legacy in scmp_kernels" rule extended to the app side.

Net effect — every `scmp_llm` Python file is accounted for

Source in `scmp_llm`	New home	Status
`SC/sc_triton.py`	`scmp_kernels/sc/kernels.py` + `matmul.py` (PR #2)	migrated
`SC/{sng,rng,lfsr_taps,config_helpers}.py`	`scmp_kernels/sc/*` (PR #2)	migrated
`SC/mp_config.py`	`scmp_kernels/mp/config.py` (PR #2)	migrated
`SC/{sc,sc_enable,bench_,compare_,test_kernel_opt,dse}.py`	—	dropped (legacy)
`SC/noise_model_calibration.py`	`tools/calibrate_noise_model.py`	migrated (this PR)
`Q-DiT/{diffusion,models,qdit,scripts,tests,utils}/`	`{same}/`	migrated (this PR)
`evaluation/`	`evaluation/`	migrated (this PR)
`imagenet256_ref/`	`evaluation/imagenet_ref/`	migrated (this PR)

Verified

• AST parses for every `.py` outside the submodule.
• Zero references to `sc_enable`, `sc_triton`, `bin_to_stoc_packed`, `xnor_matmul`, or any of the deprecated flat-API names anywhere in the app code.
• 53 `sc_matmul(granularity=...)` call sites exercise all granularities + group_a/group_b + chunk_d + rng_levels paths.

Not verified (no GPU on dev box)

• Actual Triton kernel execution end-to-end.
• Numerical reproducibility of pre-migration FID / KID results (the algorithm shift from packed-XNOR → enable-signal table-lookup means numbers won't bit-exactly match historical `scmp_llm` results, but the method is correct and the team-canonical algorithm).

Test plan

• On a CUDA + Triton + DiT-checkpoint box: `cd scmp_kernels && pytest tests/test_sc_smoke.py -v` (submodule smoke).
• `PYTHONPATH=. python -c "import qdit.sc_integration.sc_attention"` — confirm Q-DiT import chain resolves.
• Run one quant_sc_main pass end-to-end on a small batch to verify the dispatcher reaches the kernels correctly.
• Re-run one calibration sweep + one FID eval to establish new baseline numbers under the unified API.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Bootstraps the new scmp_diffusion repository by migrating Q-DiT/Diffusion application code from the private scmp_llm staging repo. All SC/MP-related imports are rewired to consume the new scmp_kernels package (added as a git submodule). Files are migrated largely verbatim from their original locations; the only non-trivial code changes happen in the SC integration layer (dispatcher consolidation, removal of sc_enable=False legacy path) which were already covered in earlier scmp_llm review cycles.

Changes:

• Adds scmp_kernels as a git submodule and rewires all SC imports to the new scmp_kernels.sc / scmp_kernels.mp public surface (with qdit/sc_integration/mp_config.py as a thin re-export shim).
• Imports Q-DiT application code (diffusion/, models/, qdit/, scripts/, tests/, utils/) along with calibration tooling and evaluation scripts (evaluation/, tools/).
• Drops the sc_enable=False legacy code path and the dead sc_matmul_qk integration.

Reviewed changes

Copilot reviewed 101 out of 107 changed files in this pull request and generated no comments.

Show a summary per file

The PR touches ~100 files across many directories; below is a grouped summary rather than per-file since the migration is mostly verbatim copies.

File / Group	Description
.gitmodules, .gitignore	Adds scmp_kernels submodule (pinned to fork) and ignore rules for large artifacts.
setup.py, requirements.txt, qdit_* env files	Packaging metadata and pinned conda/pip environment snapshots.
diffusion/, models/	Verbatim DiT + gaussian-diffusion code from upstream.
qdit/, qdit/sc_integration/	Quantization + SC integration; mp_config.py is a re-export shim, __init__.py updated to expose new API.
scripts/ (60+ files)	Calibration, sweep, SLURM, and diagnostic scripts; most are operator-facing shell/python wrappers with hard-coded scratch paths.
scripts/eval/	Clean-fid / OpenAI evaluator wrappers and pngs_to_npz packer.
evaluation/, evaluation/imagenet_ref/	FID/KID/mosaic tools and ImageNet-256 reference batch helpers.
tests/test_noise_matmul_adapters.py	Single unit test comparing the noisy surrogate against the real SC kernels through sc_matmul.
utils/, tools/	Logger / DiT-checkpoint download helpers; noise-model calibration entry point (not shown in diff).

A few minor observations (not worth blocking on, since the script files are operator-side tooling with environment-specific paths baked in throughout):

• scripts/test.sh line 5 still points to /home/kangqi/scmp_llm/results/... and ./evaluations/evaluate.sh (the directory is models/evaluations/ in this repo), so it will not run as-is. Same story for many sbatch_*.sb files referencing /scratch/.../scmp_llm/....
• scripts/unit_test.sh runs quant_sc_main.py rather than the tests/ pytest suite — the name doesn't match its content.
• scripts/quant_main.sh invokes python -u quant_main.py (no scripts/ prefix), which won't resolve from the repo root.
• models/evaluations/requirements.txt pins tensorflow-gpu>=2.0, which is no longer published on PyPI as of TF 2.12+; the canonical install is tensorflow[and-cuda] (as the new scripts/eval/README.md correctly documents).

Given these are bootstrap-time hard-coded paths inherited verbatim from scmp_llm and the PR description explicitly calls out that none of this has been GPU-verified end-to-end, I'm not filing per-line comments — they are out of scope for an initial migration PR whose goal is to land the code rather than make every operator script portable.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.