# Bernini Inference Optimization Benchmarks

Measured on 2026-07-01 with 8x H100, Ulysses 8, CFG batching, PyTorch
2.5.1+cu124, BF16, seed 42, 81 frames, 480x848, 16 fps, 50 diffusion steps,
25 semantic-planning steps, and the repository T2V demo prompt. In this
first-round experiment approximate methods were opt-in and the reference path
was dense BF16/FA3; the current launcher recommendation is documented below.

This file records the first-round module and single-prompt experiments. The
six-prompt result and the final CPU-master stack are documented in
[inference_optimization_summary.md](inference_optimization_summary.md).
Baseline terminology follows
[inference_baseline_naming.md](inference_baseline_naming.md). In the table
below, the dense BF16/FA3 row is the experiment-local `optimized-stack
baseline`; it is not the upstream `pristine baseline`.

## Module Microbenchmarks

The GEMM dimensions are the actual per-rank Bernini hot shapes. FP8 uses
per-tensor E4M3 weights and dynamically quantized E4M3 activations.

| Projection (M x K x N) | BF16 | FP8 dynamic | Speedup |
| --- | ---: | ---: | ---: |
| Q/K/V/out (12522 x 5120 x 5120) | 0.8138 ms | 1.0721 ms | 0.759x |
| FFN up (12522 x 5120 x 13824) | 2.6793 ms | 1.9229 ms | 1.393x |
| FFN down (12522 x 13824 x 5120) | 2.4610 ms | 2.6521 ms | 0.928x |
| Cross KV (512 x 5120 x 5120) | 0.0562 ms | 0.0828 ms | 0.679x |
| Cross KV (1024 x 5120 x 5120) | 0.0852 ms | 0.1237 ms | 0.689x |

Only FFN-up passed the dynamic-quantization gate and is integrated behind
`BERNINI_FP8_FFN_UP=1`. Prequantized FP8 GEMMs were 1.76x-1.96x faster, but
that upper bound excludes activation quantization.

PISA was measured on the packed self-attention shape `[100170, 5, 128]`, split
into three independent 33390-token CFG branches. Timings include layout copies,
routing/top-k, sparse attention, and output concatenation.

| HYD density | Dense FA3 | PISA end-to-end | Speedup | Relative RMSE |
| ---: | ---: | ---: | ---: | ---: |
| 0.10 | 12.1849 ms | 4.5552 ms | 2.675x | 0.5859 |
| 0.15 | 12.1849 ms | 7.9036 ms | 1.542x | 0.6010 |
| 0.30 | 12.1849 ms | 12.7882 ms | 0.953x | 0.6049 |

## Full Video Results

SSIM/PSNR compare decoded frames with the `dense quality reference` generated
from the same prompt and seed. Diffusion time is the renderer's
`video_diffusion` stage; request totals are not compared because preprocessing
and first-use caches vary with matrix order.

| Method | Configuration | Skipped steps | Diffusion | Speedup | PSNR | SSIM |
| --- | --- | ---: | ---: | ---: | ---: | ---: |
| Experiment-local optimized-stack baseline | dense BF16 + FA3 | 0/50 | 106.343 s | 1.000x | inf | 1.000000 |
| PISA HYD | density 0.15, layers 4+ | n/a | 84.586 s | 1.257x | 15.538 | 0.632802 |
| TeaCache | threshold 0.05 | 0/50 | 95.833 s | no cache hit | inf | 1.000000 |
| TeaCache | threshold 0.15 | 18/50 | 80.417 s | 1.322x | 31.624 | 0.953897 |
| EasyCache | threshold 0.05, retain 5+last 1 | 11/50 | 82.264 s | 1.293x | 38.798 | 0.977598 |
| EasyCache Wan2.2 setting | threshold 0.06, retain 7+last 1 | 12/50 | 83.930 s | 1.267x | 37.067 | 0.974997 |

FP8 FFN-up reduced the warm denoising-step median from about 1.15 s to 1.12 s
(about 1.03x), but lazy weight quantization added about 31.5 s across the two
experts and the 10-step video comparison was SSIM 0.927980. It is not a useful
one-shot default.

The PISA module kernel is fast, but it changes the seed-aligned generation
trajectory substantially. TeaCache's Wan2.1-14B polynomial transfers
imperfectly to Bernini's two experts. The later six-prompt sweep selected
EasyCache `0.08/4` as the recommended fidelity tier.

## Reproduction

```bash
python scripts/bernini/bench_fp8_gemm.py

PYTHONPATH=/path/to/compatible/pisa/python-packages:$PWD \
  python scripts/bernini/bench_pisa_attention.py \
  --implementation hyd --density 0.10 --density 0.15 --density 0.30

BERNINI_OPTIMIZATION_PROFILE=recommended_easycache \
bash scripts/bernini/run_t2v.sh
```

The current PISA checkout's README asks for newer Torch/Triton than Bernini's
production environment. The isolated overlay uses Triton 3.5.1 with the
production PyTorch 2.5.1 build because this exact kernel combination passed the
module and full-model tests; it is not an officially supported dependency set.

References: [PISA](https://github.com/xie-lab-ml/piecewise-sparse-attention),
[TeaCache](https://github.com/ali-vilab/TeaCache), and
[EasyCache](https://github.com/H-EmbodVis/EasyCache).