{
  "device": "NVIDIA GB10",
  "compute_capability": "SM121",
  "torch": "2.9.1+cu130",
  "sageattention": "2.2.0",
  "resolution": [
    1344,
    768
  ],
  "frames": 124,
  "packed_tokens": 37810,
  "seed": 440420,
  "sampling_uninstrumented_two_step_seconds": {
    "sage2": 53.53981032199226,
    "sol_bf16_tau_1_3": 46.18868866603589,
    "sol_int8_qk_tau_1_3": 42.72281823604135
  },
  "sampling_stage_step_2_seconds": {
    "sage2": {
      "packing": 0.015245259972289205,
      "denoiser": 25.396856592036784,
      "output_transform": 0.0006340660038404167,
      "sampler_update": 0.000779024965595454,
      "step_total": 25.413675360032357
    },
    "sol_bf16_tau_1_3": {
      "packing": 0.01417812000727281,
      "denoiser": 21.55711886996869,
      "output_transform": 0.0006771690095774829,
      "sampler_update": 0.0007963530370034277,
      "step_total": 21.572905345004983
    },
    "sol_int8_qk_tau_1_3": {
      "packing": 0.014827414997853339,
      "denoiser": 19.75738707598066,
      "output_transform": 0.0006591529818251729,
      "sampler_update": 0.0006342889973893762,
      "step_total": 19.773632204975
    }
  },
  "representative_block_24_p50_milliseconds": {
    "sage2": {
      "module_forward": 514.917,
      "norm": 7.155,
      "modulate": 21.734,
      "qkv_projection": 30.773,
      "rms_rope": 12.535,
      "qkv_and_output_layout": 20.346,
      "attention": 259.942,
      "output_projection": 13.366,
      "residual_gates": 17.335,
      "mlp_fc1": 39.577,
      "swiglu": 25.821,
      "mlp_fc2": 67.35
    },
    "sol_bf16_tau_1_3": {
      "module_forward": 385.35,
      "norm": 7.131,
      "modulate": 21.318,
      "qkv_projection": 30.116,
      "rms_rope": 12.512,
      "qkv_and_output_layout": 14.7,
      "attention": 141.829,
      "output_projection": 12.893,
      "residual_gates": 17.469,
      "mlp_fc1": 39.171,
      "swiglu": 24.015,
      "mlp_fc2": 64.833
    },
    "sol_int8_qk_tau_1_3": {
      "module_forward": 364.141,
      "norm": 7.049,
      "modulate": 20.848,
      "qkv_projection": 29.719,
      "rms_rope": 12.327,
      "qkv_and_output_layout": 14.849,
      "attention": 120.272,
      "output_projection": 13.165,
      "residual_gates": 17.433,
      "mlp_fc1": 39.411,
      "swiglu": 23.971,
      "mlp_fc2": 65.111
    }
  },
  "attention_internal_p50_milliseconds": {
    "sage2": {
      "qk_quantize": 6.45,
      "v_quantize": 10.185,
      "attention_forward": 238.995,
      "framework_and_k_smoothing_estimate": 2.982,
      "total_excluding_external_layout": 258.613
    },
    "sol_bf16_tau_1_3_exact_conditioning_kv": {
      "k_summary": 2.44,
      "v_summary": 2.411,
      "routing_threshold": 2.423,
      "prepare_total": 7.287,
      "routed_forward_estimate": 143.455,
      "total": 150.741
    },
    "sol_int8_qk_tau_1_3_exact_conditioning_kv": {
      "k_summary_and_quantize": 4.449,
      "v_summary": 2.454,
      "q_quantize_and_threshold": 4.27,
      "prepare_total": 11.533,
      "routed_forward_estimate": 114.882,
      "total": 126.416
    },
    "sol_int8_qk_pv_tau_1_3_exact_conditioning_kv": {
      "prepare_total": 11.714,
      "v_int8_quantize": 3.875,
      "routed_forward_estimate": 126.936,
      "total": 138.651
    }
  },
  "quality_gates": {
    "sol_bf16_tau_1_3_exact_conditioning_kv_relative_l2_vs_sage2": 0.6257168054580688,
    "sol_int8_qk_tau_1_3_exact_conditioning_kv_relative_l2_vs_sage2": 0.6258820295333862,
    "sol_fully_dense_tau_minus_100_relative_l2_vs_sage2": 0.017087,
    "sol_fully_dense_tau_minus_100_p50_milliseconds": 947.901,
    "conclusion": "Current Sol routing has no point that is both faster than Sage2 and numerically Sage-like; sparse timings are optimization bounds, not an accepted backend."
  },
  "sage2_block_share_estimates": {
    "attention": 0.505,
    "nvfp4_projections": 0.293,
    "modulation_gates_and_swiglu": 0.126,
    "layout": 0.04,
    "norm_and_rope": 0.038
  },
  "measurement_policy": {
    "sampling_totals": "uninstrumented wall time",
    "sampling_stages": "CUDA-synchronized stage attribution",
    "block_components": "CUDA-synchronized representative block 24 medians",
    "attention_components": "nested CUDA-synchronized probes; residual values are p50 difference estimates",
    "resident_services": "H3 hot runtime and Qwen vLLM remained loaded and idle; medians are preferred over means"
  },
  "next_targets": [
    "Bit-exact fused H3 modulation and residual gates",
    "SM121-specific NVFP4 projection and epilogue fusion",
    "A dense Sage-compatible kernel that reads strided BSHD QKV without materialization",
    "A quality-gated timestep/block hybrid only after full latent and video evaluation"
  ]
}