{
  "device": "NVIDIA GB10",
  "compute_capability": "SM121",
  "workload": {
    "resolution": [1344, 768],
    "frames": 124,
    "packed_tokens": 37810,
    "steps": 12,
    "seed": 440420,
    "attention": "sage2",
    "production_flags": {
      "H3_NVFP4_SCALE_BACKEND": "vortex",
      "H3_NVFP4_SCALE_VERSION": "1",
      "H3_FUSED_ELEMENTWISE": "1",
      "H3_SAGE_QKV_LAYOUT": "strided_nhd",
      "H3_CUTE_QKV_RING": "0"
    }
  },
  "canonical": {
    "elapsed_seconds": 288.9257714879932,
    "video_checksum": -91481.7890625,
    "audio_checksum": 678.401611328125,
    "peak_allocated_bytes": 16036566016,
    "peak_reserved_bytes": 19484639232
  },
  "block_24_p50_ms": {
    "module_forward": 468.21730851661414,
    "attention": 258.47244350006804,
    "qk_rmsnorm_rope": 12.200642988318577,
    "qkv_projection": 29.554627515608445,
    "attention_output_projection": 12.935250502778217,
    "mlp_fc1": 38.37749600643292,
    "mlp_swiglu": 23.899421998066828,
    "mlp_fc2": 66.63533198297955,
    "layout_materialization": 0.015351508744061
  },
  "sage2_nhd_p50_ms": {
    "total": 260.0231445103418,
    "qk_quantize": 7.684325042646378,
    "v_quantize": 10.653970995917916,
    "attention_mainloop": 238.8111045002006,
    "framework_residual_estimate": 2.873743971576914
  },
  "active_mainloop": "sageattention_sm89::qk_int8_sv_f8_accum_f16_fuse_v_scale_attn_inst_buf",
  "hopper_wgmma_retarget": {
    "target": "sm_121a",
    "result": "compile rejected",
    "reason": "SM121 ptxas does not support Hopper wgmma fence, mma_async, commit_group, or wait_group instructions"
  },
  "blackwell_umma_constraint": "CUTLASS SM120/121 supports F8/F6/F4 UMMA but not the INT8 QK operation required for exact Sage2 parity",
  "decision": "Prototype exact AdaLN modulation directly into NVFP4 scale/pack for QKV and FC1. Retain Comfy GEMMs. Keep native Sage2 attention research paused unless the classic MMA schedule can be retuned without changing arithmetic."
}