{
  "schema": "anvil-serving.compatibility/v1",
  "captured": "2026-08-01",
  "candidate": "thinkingmachines/Inkling-Small-NVFP4",
  "revision": "b6a99534467840620d411e4cd4ad5819b2610d9c",
  "served_model": "inkling-small-nvfp4-tp2",
  "hardware": "2x NVIDIA RTX PRO 6000 Blackwell Max-Q, exclusive TP=2 over PCIe without NVLink",
  "engine_revision": "sglang-b7252cc6b",
  "base_image": "lmsysorg/sglang@sha256:b90c0d760a65bc4dbbe4520bea966c437cc40391dcb7cca2a74922985dc1abeb",
  "final_image": "anvil-sglang@sha256:6a8afc5ca0036c1be8810443636d6f835702d1e2ae5a1d717990b0baf8e70a2f",
  "failures_and_fixes": [
    {
      "failure": "ModelOpt import required accelerate",
      "fix": "install accelerate==1.14.0 in the pinned derived image"
    },
    {
      "failure": "ModelOpt cache and quantization-config lookups ignored the pinned checkpoint revision",
      "fix": "forward the configured revision through both lookups"
    },
    {
      "failure": "SM120 grouped GEMM requested 110592 bytes of shared memory against a 101376-byte hardware limit",
      "fix": "use the existing two-stage schedule for the affected SM120 grouped GEMMs"
    },
    {
      "failure": "the image did not ship silu_and_mul_interleaved_sm_120.json for the Helion activation path",
      "fix": "select SGLang's existing two-stage Triton SiLU/multiply fallback only for SM120 interleaved activations"
    },
    {
      "failure": "WSL2 symmetric-memory logits gather raised SIGFPE",
      "fix": "disable only the incompatible logits optimization and preserve NCCL fallback"
    }
  ],
  "final_result": {
    "status": "healthy",
    "native_quantization": "ModelOpt NVFP4",
    "weight_load_gib_per_rank": [85.93, 86.90],
    "bf16_kv_swa_gib_per_rank": 3.14,
    "decode_graph_capture": "passed with NCCL fallback",
    "kernel_tune_claim": false
  }
}