Serving systems choose a weight format for each model and deployment. A low-cost screen predicts the resulting KL divergence from the unquantized model's Fisher information using $\tfrac12Δ^\top FΔ$, the expansion underlying Hessian-aware and second-order post-training quantizers. We assess the reliability of this screen in GPU batches with channel-wise int8, per-tensor fp8 and group-wise int4 on Qwen3-8B, Qwen3-4B, OLMo-2-7B, Qwen3-32B and Qwen3-30B-A3B. Relative predictions hold across a 16-fold range of perturbation scales, int4 group sizes and rounding modes, while absolute predictions are heavy-tailed. A trimmed average brings 10 of 12 cells on Qwen3-4B and OLMo-2 within 15% of measured KL. The MoE model exhibits nonquadratic response at small perturbations, with int8 KL scaling as $α^{0.63}$. Computing per-sequence gradients in bf16 can cause fourfold prediction errors. We trace these errors to arithmetic in the lower half of the backward pass and eliminate them by using fp32 there. Batch-invariant kernels run at 0.55 times the throughput, while changes in reduction order introduce a floor of 0.17-0.22 nats per sequence. We build a screener with one streamed fp32 Fisher pass, trimmed tensor-block predictions with a band, a flag for aligned tensors and per-layer format allocation. On the unseen Mistral-7B-v0.3 family, all six predictions fall within the band, all three KL budgets are met, and vLLM throughput is predicted within 0.4%.
Looped language models such as Ouro reuse one block of layers several times per token. A rollout engine can batch the deeper loops of several response positions and run them only when a deep readout is due. This deferred schedule is fast, but the learner of an on-policy RL algorithm recomputes the sampled tokens' log-probabilities with a dense pass, and the two computations differ in floating point: in FP32 at 4096 tokens about 18% of the sampled tokens miss a $10^{-5}$ same-weight gate, and with bf16 weights nearly every token does. We use a hybrid learner log-probability, $\ell=\ell_{\mathrm{dense}}+\mathrm{sg}(\ell_{\mathrm{sched}}-\ell_{\mathrm{dense}})$, where $\mathrm{sg}$ denotes stop-gradient. Its value is a no-grad replay through the rollout's own schedule, and its gradient is the dense learner's. The replay reproduces the rollout bit for bit in all 167 FP32 and 28 bf16 real-weight cells we ran on three platforms, including rollouts whose batch shrinks at stop tokens, and the gradient has the same FP32 error as other dense implementations in all 102 pre-registered comparisons from 96 to 4096 tokens; the full GRPO loss of a training step matches a float64 reference to $3.1\times10^{-4}$. On one GPU, deferral raises rollout throughput by $1.47$ to $1.56\times$, and by $1.18$ to $1.54\times$ once every step runs from CUDA graphs, bit for bit. In a two-node GRPO loop, one-wave KV admission and CUDA graphs shorten an update by $2.71\times$ with the gate bitwise at every update, and four nodes keep the update time within 2% at the same per-GPU load. In 100-update runs held-out GSM8K accuracy rose from 23% to 82% with the mixed-depth readout that deferral relies on and from 34% to 90% with the native readout, which reached a higher accuracy in less time. For looped models beyond one node we analyse a ring mapping of layers to GPUs and measure its communication paths.