07_timing.md

07 · Inference timing (Table 7, Appendix G)

Task. Measure single-cell forward-pass latency on A100, batch 1, fp32, and convert to NFE (number of function evaluations per query).

Expected numbers

Method	Median latency / query	NFE per query	GFLOPs
PRESCIENT	194.33 ms	many (SDE solver)	0.07
CellFlow	410.36 ms	many (ODE solver)	—
Chreode	64.79 ms	1 (incl. $-\nabla U$ autograd)	0.90

Prerequisites

python scripts/download_weights.py

CPU-only also works; the relative comparison is what matters.

Run (≈ 2 min)

PYTHONPATH=src python scripts/timing/run_available_timing_benchmark.py \
  --device cuda:0 \
  --n 100 --warmup 10 \
  --output output/reproduce/timing.json

This warms up for 10 forward passes, then times 100 consecutive forward passes (batch 1, fp32) including the autograd backward for $-\nabla_z U_\theta$. The "1 NFE" claim follows because $\hat z_{t+\Delta}$ is produced by a single R_θ call, not by a multi-step ODE/SDE integrator. Read the median from output/reproduce/timing.json.

PRESCIENT / CellFlow rows

The PRESCIENT and CellFlow rows in Table 7 come from running each baseline's public reproduction script on the same A100 with batch 1; we cite each baseline's repo in paper/chreode.pdf Appendix G. We do not ship those baseline timing scripts in this repo to avoid bundling other projects' code — please consult the respective upstreams if you want to re-time them.