outputs/nvfp4/e2e_speed.json

{
  "model": "klein-4B, Linears-only NVFP4 W4A4 swap (real Nunchaku kernel), attention/norm/VAE/text-enc stay bf16",
  "512px_4step": {
    "bf16_s": 0.513,
    "bf16_step_ms": 101,
    "bf16_vram_gb": 16.9,
    "nvfp4_r128_s": 0.418,
    "nvfp4_r128_step_ms": 76,
    "nvfp4_r128_vram_gb": 12.0,
    "speedup_r128": 1.24,
    "nvfp4_r64_s": 0.413,
    "speedup_r64": 1.24
  },
  "1024px_4step": {
    "bf16_s": 1.794,
    "bf16_step_ms": 373,
    "bf16_vram_gb": 19.6,
    "nvfp4_r128_s": 1.515,
    "nvfp4_r128_step_ms": 303,
    "nvfp4_r128_vram_gb": 14.7,
    "speedup_r128": 1.18,
    "nvfp4_r64_s": 1.494,
    "speedup_r64": 1.2
  },
  "per_layer_gemm_speedup": "2.49x (r128) / 2.75x (r64) \u2014 isolated GEMM",
  "why_lower_e2e": "attention is bf16+O(N^2) and dominates (esp. at 1024px); VAE+text-encode are fixed bf16 overhead; the Linear swap is unfused (per-Linear act-quant). The 9B Nunchaku FULL pipeline got 2.69x because it also fuses attention/quant and is more GEMM-heavy.",
  "vram": "~28% lower (4-bit Linears): 16.9->12.0 GB @512, 19.6->14.7 GB @1024",
  "rank_tax_e2e": "negligible (~1%) at these sizes \u2014 low-rank branch tiny vs the rest"
}