[0z5a]

Incremental review: Files changed.

Direct dependency and ancestor review: #1. Baseline: cleaned #1 head 555a862a45bf243e4ff21258896875ca65133148. This fork PR shows only this layer.

Add a shared-parameter Ouro Megatron provider using the existing custom-provider hook. It preserves physical parameter layout, four sandwich norms, inter-loop normalization and shared-layer gradients; differentiable SDPA handles training. TP=PP=CP=1; unused gate weights are retained and frozen.

H20 correctness check	Result
Tiny shared gradients / recomputation / cached scores at K=2/3/4	Logit error ≤1.2e-7
Packed-sequence isolation / physical parameter count	Passed
Full BF16 logits vs matched HF implicit-causal SDPA	Exact at K=2/3/4
Full K=4 FP32 logits vs default HF SDPA	Error 0
Full K=4 FP32 cached scores vs dense	Error 6.68e-6
Full BF16 cached scores vs dense	Error 0.0478 / 0.1057 / 0.1046

An initial explicit-mask BF16 comparison failed; matched implicit-causal dispatch explains the logit result. BF16 execution is not generally bit-exact. Tiny two-rank training/save/restart passed. The recipe layer #4 completes six full-model updates and resumes a seventh. Reward convergence and long-context coverage remain outstanding; speed improvement is unmeasured.

This historical provider layer is consolidated with later recipe/conversion work in upstream #459. Test and validation files were removed from this diff.

[0z5a]

Superseded by vllm-project#459, now merged into 0z5a-branch. This historical Ouro layer is included in the consolidated upstream integration; closing the duplicate PR.