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

[Lmy271828]

What

Offline tooling to consume Omega-QVLA quantized packs (dit_svdquant_v1, W4A4 GPTQ + DuQuant rotations for pi0.5) on FlashRT's SM110 E0M3 path:

• tools/convert_omega_pack_e0m3.py — re-quantizes each record's dequantized fp16 weight into the E0M3 operand layout (packed 4-bit [N, K/2] + tile-interleaved UE4M3 SFB scales) via quantize_e0m3_dynamic_sfa_fp16. Two --fold strategies for the activation-side per-channel calibration table: none (S0) and mean (S1). Aux tensors (DuQuant rotations, permutation, scale tables) are copied through for the runtime consumer.
• tools/check_omega_e0m3_layer.py — single-layer cosine gate: converted E0M3 GEMM vs. fp16 reference and vs. Omega's own fake-quant (pure-torch emulate mode for pre-checks, kernel mode on hardware).
• tools/gen_omega_pack_fixture.py — synthetic miniature pack (schema-identical records, random-orthogonal rotation blocks, outlier-channel weights) so the full path is reviewable without the 4.8 GB real pack.
• docs/omega_pack_e0m3.md — format doc: record schema, the S0/S1 analysis, measured results, roadmap. Purely additive (tools/, docs/); no changes to kernels or frontends. Runtime consumption (flag-gated weight_format wiring) is a follow-up PR.

Key finding: drop the calibration table (S0)

The pack ships a per-step, per-channel activation scale table. Folding its mean into the weights (S1) collapses on real hardware (cos 0.16 on q_proj): the fold presses per-16 block scales down to the UE4M3 subnormal floor (2⁻⁹). Dropping the table entirely (S0) wins — DuQuant's rotation+permutation already whitens channel magnitudes, and per-16 dynamic amax is the stronger quantizer. Measured on Thor (4 layers): S0 vs fp16 = 0.9924–0.9932, ≥ Omega's own fake-quant (0.9921–0.9928). Consequence: no per-step scale dispatch is needed on the runtime path, which simplifies the follow-up wiring PR.

Measured (Thor, pi0.5 LIBERO-10)

• 252/252 records converted; single-layer gates pass (details: §4 of the format doc).
• Downstream, with a runtime consumer wired to the converted artifact: LIBERO-10 ×500 paired success rate 90.4% vs. BF16 baseline 91.6% (McNemar p = 0.53, no significant difference), episode time 58 s vs. 148 s for the PyTorch fake-quant path (2.6×); the consumer's pybind kernels are CUDA-graph capturable (10-step denoise loop in a single graph, 50-episode check 45/50 ≈ eager). Evidence: https://github.com/Lmy271828/FlashRT/blob/libero-eval-fp4/docs/omega_pack_e0m3.md

Reproduce (self-contained, no Omega-QVLA checkout needed)

python tools/gen_omega_pack_fixture.py --out /tmp/fixture_pack.pt   # CPU, seconds                                                       
python tools/convert_omega_pack_e0m3.py --pack /tmp/fixture_pack.pt \                                                                    
   --out /tmp/fixture_e0m3.pt --fold none                           # Thor                                                               
python tools/check_omega_e0m3_layer.py --pack /tmp/fixture_pack.pt \                                                                     
   --mode kernel                                                    # Thor                                                               

Fixture validation on Thor: 4/4 records converted, consumer-level cosine vs fp16 = 0.9887/0.9890 (q_proj/down_proj), E0M3 0.19 ms/layer.

Non-goals
• Runtime weight_format wiring (next PR, per the roadmap in §5).
• SVDQuant low-rank epilogue fusion (rank = 0 in this pack; deferred).
• Per-step weight tables / activation scale dispatch (refuted by S0, §4).