{
  "schema": "anvil-serving.candidate-load-failure/v1",
  "observed_at": "2026-09-03T18:58:09Z",
  "candidate": "gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090-SGLang-DSpark-165169",
  "host_label": "Primary Node",
  "gpu": "NVIDIA GeForce RTX 5090",
  "runtime": {
    "engine": "sglang",
    "image": "lmsysorg/sglang:v0.5.18@sha256:bde16a8447b19e89056b9eea06c72be6c02801dc89d528c9ea90c53368fd74bf",
    "target_revision": "b8ca3826548c9a7735642feb05c3c473f1fede1f",
    "draft_revision": "eba1ac5a66c74902eaa95a4000a7c5eda96d8e95",
    "context_tokens": 165169,
    "max_running_requests": 2,
    "kv_cache_dtype": "fp8_e4m3",
    "mem_fraction_static": 0.9
  },
  "load_progress": {
    "target_weight_gb_reported": 17.10,
    "draft_weight_gb_reported": 1.44,
    "target_kv_tokens_allocated": 142219,
    "draft_kv_tokens_allocated": 142219,
    "target_kv_k_gb": 2.17,
    "target_kv_v_gb": 2.17,
    "draft_kv_k_gb": 0.68,
    "draft_kv_v_gb": 0.68,
    "reached_readiness": false
  },
  "failure": {
    "stage": "DSpark draft CUDA-graph capture and FlashInfer autotune",
    "exception": "RuntimeError",
    "message": "mat1 and mat2 shapes cannot be multiplied (14x5120 and 2560x248320)",
    "classification": "target-drafter-runtime hidden-width incompatibility",
    "memory_failure": false,
    "repeat_without_change": false
  },
  "decision": "rejected",
  "decision_note": "The advertised speculative lane cannot serve with the pinned final target, drafter, and SGLang v0.5.18 combination. Lowering memory or context would not repair the matrix-width mismatch."
}