{
  "eval_harness": "training/eagle3-v4flash/vllm_eagle_speedup_eval.py",
  "modal_app": "https://modal.com/apps/maniac-inc/main/ap-HZ9jkaoFypw9VsClA36wbY",
  "date": "2026-06-06",
  "gpu": "B200:4",
  "vllm_image": "vllm/vllm-openai:v0.22.0",
  "target": "deepseek-ai/DeepSeek-V4-Flash",
  "draft_checkpoint": "v4-flash-eagle3.1-genv3-blend",
  "num_prompts": 8,
  "max_tokens": 128,
  "num_speculative_tokens": 3,
  "baseline_tok_per_s": 15.619405951323843,
  "spec_tok_per_s": 41.10993617236299,
  "speedup": 2.6319782135426646,
  "acceptance_rate": 0.11038394415357766,
  "mean_accept_len": 1.331151832460733,
  "prompt_exact_match": 2,
  "token_match_rate": 0.4462890625,
  "baseline_self_exact": 1,
  "baseline_self_token_match_rate": 0.359375,
  "notes": "Greedy decode (temperature=0, seed=0). Raw prompts without chat template. Spec vs baseline token_match_rate exceeds baseline_self control — EAGLE verify is lossless within cross-run FP8+EP noise. vLLM requires Maniac deepseek_v4 EAGLE-3 overlay patches."
}