{
  "kind": "qwen3_next_norm_reuse_check",
  "rfc": "RL-Align/RL-Kernel#428",
  "op": "qwen3_next_rms_norm",
  "rl_kernel_commit": "88f59f7d06741ba533a7b9b718b331ee885b93b7",
  "tracked_tree_dirty": false,
  "environment": {
    "gpu": "NVIDIA B200",
    "capability": [
      10,
      0
    ],
    "cuda": "13.0",
    "libraries": {
      "torch": "2.13.0",
      "triton": "3.7.1",
      "transformers": "5.17.0",
      "vllm": "0.30.0",
      "flashinfer-python": "0.6.18.post1",
      "liger-kernel": "0.8.4",
      "fla-core": "0.5.2",
      "transformer_engine": "2.20.2",
      "megatron-core": null
    },
    "megatron_source_commit": "52fbcbc"
  },
  "quick": false,
  "checks": [
    "gates"
  ],
  "results": {
    "rl_kernel": {
      "source": "rl-kernel Qwen3NextRMSNormCudaOp",
      "backward": "full",
      "contract_gates": {
        "forward": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [],
          "stderr_tail": []
        },
        "gradient": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/full') tensor=dweight max_abs=0.00000000e+00 passed=True",
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/chunked') tensor=dweight max_abs=0.00000000e+00 passed=True"
          ],
          "stderr_tail": []
        }
      }
    },
    "reference": {
      "source": "rl-kernel PyTorch reference",
      "backward": "full"
    },
    "transformers": {
      "source": "transformers 5.17.0 Qwen3NextRMSNorm",
      "backward": "full",
      "contract_gates": {
        "forward": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [],
          "stderr_tail": []
        },
        "gradient": {
          "returncode": 1,
          "passed": false,
          "failed_lines": [
            "op=qwen3_next_rms_norm profile=cuda_bf16 candidate=__main__._gate_child.<locals>.Swapped::cuda passed=False",
            "invariance pair=('BN/full', 'BN/chunked') tensor=dx transform=chunk max_abs=1.52587891e-05 passed=False",
            "invariance pair=('BN/full', 'B1-singleton_aggregate/full/s2') tensor=dx transform=batch_size max_abs=1.52587891e-05 passed=False",
            "invariance pair=('BN/full', 'B1-singleton_aggregate/chunked/s2') tensor=dx transform=chunk max_abs=1.52587891e-05 passed=False"
          ],
          "singleton_aggregate": [
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/full') tensor=dweight max_abs=0.00000000e+00 passed=True",
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/chunked') tensor=dweight max_abs=0.00000000e+00 passed=True"
          ],
          "stderr_tail": []
        }
      }
    },
    "liger": {
      "source": "Liger 0.8.4 RMSNorm, offset 1, gemma",
      "backward": "full",
      "contract_gates": {
        "forward": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [],
          "stderr_tail": []
        },
        "gradient": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/full') tensor=dweight max_abs=0.00000000e+00 passed=True",
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/chunked') tensor=dweight max_abs=0.00000000e+00 passed=True"
          ],
          "stderr_tail": []
        }
      }
    },
    "fla": {
      "source": "FLA 0.5.2 rms_norm, weight passed as 1 + w",
      "backward": "full",
      "contract_gates": {
        "forward": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [],
          "stderr_tail": []
        },
        "gradient": {
          "returncode": 0,
          "passed": true,
          "failed_lines": [],
          "singleton_aggregate": [
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/full') tensor=dweight max_abs=0.00000000e+00 passed=True",
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/chunked') tensor=dweight max_abs=0.00000000e+00 passed=True"
          ],
          "stderr_tail": []
        }
      }
    },
    "transformer_engine": {
      "source": "TE 2.20.2 RMSNorm(zero_centered_gamma)",
      "backward": "x-only"
    },
    "megatron": {
      "source": "Megatron BatchInvariantRMSNormFn(zero_centered_gamma=True)",
      "backward": "full",
      "contract_gates": {
        "forward": {
          "returncode": 1,
          "passed": false,
          "failed_lines": [
            "op=qwen3_next_rms_norm profile=cuda_bf16 candidate=__main__._gate_child.<locals>.Swapped::cuda passed=False",
            "accuracy config=BN/full max_abs=1.72778440e+00 max_rel=1.14331472e+00 passed=False",
            "accuracy config=BN/chunked max_abs=1.72778440e+00 max_rel=1.14331472e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/full/s0 max_abs=1.72659349e+00 max_rel=1.14331472e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/full/s1 max_abs=1.72467911e+00 max_rel=1.14325333e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/full/s2 max_abs=1.72347617e+00 max_rel=1.14318168e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/full/s3 max_abs=1.72778440e+00 max_rel=1.14321959e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/chunked/s0 max_abs=1.72659349e+00 max_rel=1.14331472e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/chunked/s1 max_abs=1.72467911e+00 max_rel=1.14325333e+00 passed=False",
            "accuracy config=B1-singleton_aggregate/chunked/s2 max_abs=1.72347617e+00 max_rel=1.14318168e+00 passed=False"
          ],
          "singleton_aggregate": [],
          "stderr_tail": []
        },
        "gradient": {
          "returncode": 1,
          "passed": false,
          "failed_lines": [
            "op=qwen3_next_rms_norm profile=cuda_bf16 candidate=__main__._gate_child.<locals>.Swapped::cuda passed=False",
            "invariance pair=('BN/full', 'BN/chunked') tensor=dx transform=chunk max_abs=1.90734863e-06 passed=False",
            "invariance pair=('BN/full', 'B1-singleton_aggregate/chunked/s3') tensor=dx transform=chunk max_abs=1.90734863e-06 passed=False"
          ],
          "singleton_aggregate": [
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/full') tensor=dweight max_abs=0.00000000e+00 passed=True",
            "singleton_aggregate pair=('BN/full', 'B1-singleton_aggregate/chunked') tensor=dweight max_abs=0.00000000e+00 passed=True"
          ],
          "stderr_tail": []
        }
      }
    },
    "flashinfer": {
      "source": "FlashInfer 0.6.18.post1 gemma_rmsnorm",
      "backward": "none"
    },
    "vllm": {
      "source": "vLLM 0.30.0 GemmaRMSNorm.forward_cuda",
      "backward": "none"
    }
  }
}