tools: Omega-QVLA dit_svdquant_v1 pack → E0M3/UE4M3 converter + format doc + synthetic fixture

[Lmy271828]

Scope

Offline tooling for converting Omega-QVLA dit_svdquant_v1 rank-0 packs into the packed E0M3 weight and tile-interleaved UE4M3 SFB format consumed by FlashRT's existing SM110 GEMM.

This PR does not add a frontend, pipeline, runtime consumer, server, CUDA graph, CMake source, pybind symbol, or public Python API. It has no effect on existing model inference or default builds.

Included:

• tools/convert_omega_pack_e0m3.py: schema validation and offline conversion.
• tools/check_omega_e0m3_layer.py: CPU emulation plus direct converted-artifact GEMM verification on Thor.
• tools/gen_omega_pack_fixture.py: deterministic four-record synthetic fixture.
• docs/omega_pack_e0m3.md: Milestone 1 format and reproduction contract.

Audit follow-up

Commit 7355c3c6 addresses the merge review:

• Removes documentation for runtime/serving/graph files that are not present in this PR. External runtime experiments are no longer represented as repository functionality.
• Validates every selected record before CUDA work: format, rank=0, W4A4 metadata, N/K metadata, low-rank shapes, positive finite scale tables, rotation shapes, finite tensors, and a real int64 permutation.
• Fails instead of skipping unsupported records. An unfiltered complete pack requires 252 records by default; fixtures and filtered subsets require complete selected coverage.
• Adds CPU-only --validate-only for schema and coverage checks.
• Records schema version, selected layers, and source/selected counts in the artifact. weights, aux, and selected_layers must cover the same set.
• Writes through a unique sibling temporary file and atomically replaces the destination only after complete conversion. The source pack cannot be used as --out.
• Adds --artifact to the checker. This path quantizes only the activation and feeds the artifact's stored packed and sfb tensors directly to cutlass_fp4_gemm_e0m3w; it fails when artifact-vs-fp16 cosine is below --min-artifact-cos.
• Narrows .gitignore from global *.pt to /artifacts/omega_e0m3/*.pt.
• Adds 13 CPU contract tests.

Reproduce

CPU validation and emulation:

python tools/gen_omega_pack_fixture.py --out /tmp/fixture_pack.pt
python tools/convert_omega_pack_e0m3.py \
    --pack /tmp/fixture_pack.pt --out /tmp/unused.pt --validate-only
python tools/check_omega_e0m3_layer.py \
    --pack /tmp/fixture_pack.pt --mode emulate

Thor direct artifact round-trip:

python tools/convert_omega_pack_e0m3.py \
    --pack /tmp/fixture_pack.pt --out /tmp/fixture_e0m3.pt --fold none
python tools/check_omega_e0m3_layer.py \
    --pack /tmp/fixture_pack.pt --artifact /tmp/fixture_e0m3.pt \
    --mode kernel --min-artifact-cos 0.98

Full pack conversion defaults to a strict 252/252 coverage gate. Use --expected-records only for a deliberately different complete pack.

Validation

• Tool contract tests: 13 passed.
• Changed Python scripts: py_compile passed.
• git diff --check: passed.
• Default fixture generation: passed; four schema records load with weights_only=True.
• Fixture CPU validation: 4/4 passed.
• Fixture CPU emulation: passed.
• Private path scan: clean.

The Thor host was offline during this audit follow-up, and the real 252-record pack was not available locally. Before merge, rerun the two hardware gates introduced here: fixture packed/SFB direct GEMM round-trip and real-pack 252/252 conversion plus representative direct artifact GEMM checks. Cross-platform model E2E is not required because this PR has no runtime integration.

[LiangSu8899]

Audit fixes are pushed in 7355c3c6.

The PR is now limited to its actual Milestone 1 surface. The converter validates all selected records and fails on unsupported format/rank/shape/rotation/permutation/scale metadata; full-pack conversion defaults to a strict 252-record gate. Output is atomic and records explicit schema/coverage metadata.

The checker now has a real artifact round-trip: --artifact loads the converted packed and sfb tensors and feeds them directly to the SM110 GEMM. It does not regenerate the weight from weight_res_q, and it returns failure below --min-artifact-cos.

Validation completed locally:

• 13 tool contract tests passed.
• Default four-record fixture generated and loaded with weights_only=True.
• CPU validation reported 4/4.
• CPU emulate checker passed.
• py_compile and git diff --check passed.
• Merge simulation against current main is clean.

The Thor host was offline and the real 252-record pack was not present locally. The PR body now lists the remaining pre-merge hardware gates precisely: direct fixture packed/SFB GEMM round-trip, then real-pack 252/252 conversion and representative direct artifact GEMM checks. No model E2E is needed for this offline-only PR.