{
 "device": {
  "jax": "0.11.2",
  "backend": "gpu",
  "devices": [
   "cuda:0"
  ],
  "python": "3.14.8",
  "device_kind": "NVIDIA GeForce RTX 3090",
  "gpu": "NVIDIA GeForce RTX 3090",
  "driver": "615.78.08",
  "gpu_memory_gb": 24.0,
  "torch": "2.14.1+cu130",
  "torch_cuda": "13.0"
 },
 "config": {
  "note": "Locked A/B of compute_mamba_scan on GPU before/after 9386bfb (state-size-dependent tile configs), 3 interleaved rounds per side, medians in ms; transcribed from the A/B log. kernel_mamba_gpu.json covers state_dim=16 only.",
  "commit": "9386bfb",
  "rounds": 3
 },
 "results": [
  {
   "batch": 8,
   "seq": 1024,
   "dim": 128,
   "state": 16,
   "mode": "fwd",
   "before_ms": 0.23,
   "after_ms": 0.21,
   "iqr_overlap": true
  },
  {
   "batch": 8,
   "seq": 1024,
   "dim": 128,
   "state": 16,
   "mode": "bwd",
   "before_ms": 0.57,
   "after_ms": 0.57,
   "iqr_overlap": true
  },
  {
   "batch": 8,
   "seq": 1024,
   "dim": 128,
   "state": 64,
   "mode": "fwd",
   "before_ms": 6.9,
   "after_ms": 0.44,
   "iqr_overlap": false
  },
  {
   "batch": 8,
   "seq": 1024,
   "dim": 128,
   "state": 64,
   "mode": "bwd",
   "before_ms": 8.68,
   "after_ms": 1.27,
   "iqr_overlap": false
  },
  {
   "batch": 8,
   "seq": 1024,
   "dim": 128,
   "state": 128,
   "mode": "fwd",
   "before_ms": 8.21,
   "after_ms": 0.92,
   "iqr_overlap": false
  },
  {
   "batch": 8,
   "seq": 1024,
   "dim": 128,
   "state": 128,
   "mode": "bwd",
   "before_ms": 41.61,
   "after_ms": 6.52,
   "iqr_overlap": false
  },
  {
   "batch": 4,
   "seq": 2048,
   "dim": 512,
   "state": 64,
   "mode": "fwd",
   "before_ms": 26.73,
   "after_ms": 0.98,
   "iqr_overlap": false
  },
  {
   "batch": 4,
   "seq": 2048,
   "dim": 512,
   "state": 64,
   "mode": "bwd",
   "before_ms": 33.63,
   "after_ms": 3.96,
   "iqr_overlap": false
  },
  {
   "batch": 4,
   "seq": 2048,
   "dim": 512,
   "state": 128,
   "mode": "fwd",
   "before_ms": 28.33,
   "after_ms": 3.22,
   "iqr_overlap": false
  },
  {
   "batch": 4,
   "seq": 2048,
   "dim": 512,
   "state": 128,
   "mode": "bwd",
   "before_ms": 174.36,
   "after_ms": 26.63,
   "iqr_overlap": false
  },
  {
   "batch": 2,
   "seq": 4096,
   "dim": 256,
   "state": 32,
   "mode": "fwd",
   "before_ms": 0.48,
   "after_ms": 0.46,
   "iqr_overlap": true
  },
  {
   "batch": 2,
   "seq": 4096,
   "dim": 256,
   "state": 32,
   "mode": "bwd",
   "before_ms": 1.6,
   "after_ms": 1.54,
   "iqr_overlap": false
  }
 ]
}