{
  "model": "sshleifer/tiny-gpt2",
  "device": "cuda",
  "dtype": "torch.float16",
  "other_prompt": "Batching an unrelated request should not change mine.",
  "rows": [
    {
      "index": 1,
      "prompt": "Tell me about Richard Feynman",
      "top1_token_id": 5087,
      "top1_text": " factors",
      "top2_token_id": 13584,
      "top2_text": " mutual",
      "batched_top1_token_id": 5087,
      "batched_top1_text": " factors",
      "margin_single": 0.001708984375,
      "margin_batched": 0.001708984375,
      "epsilon_max_abs_delta": 3.0517578125e-05,
      "two_epsilon": 6.103515625e-05,
      "argmax_stable": true
    },
    {
      "index": 2,
      "prompt": "Explain why floating point addition is not associative",
      "top1_token_id": 5087,
      "top1_text": " factors",
      "top2_token_id": 13584,
      "top2_text": " mutual",
      "batched_top1_token_id": 5087,
      "batched_top1_text": " factors",
      "margin_single": 0.0020751953125,
      "margin_batched": 0.0020751953125,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 3,
      "prompt": "Write one sentence about New York City",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.00140380859375,
      "margin_batched": 0.00140380859375,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 4,
      "prompt": "What is a theorem prover?",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.0009765625,
      "margin_batched": 0.0009765625,
      "epsilon_max_abs_delta": 3.814697265625e-06,
      "two_epsilon": 7.62939453125e-06,
      "argmax_stable": true
    },
    {
      "index": 5,
      "prompt": "Give a short definition of batch inference",
      "top1_token_id": 5087,
      "top1_text": " factors",
      "top2_token_id": 13584,
      "top2_text": " mutual",
      "batched_top1_token_id": 5087,
      "batched_top1_text": " factors",
      "margin_single": 0.0020751953125,
      "margin_batched": 0.0020751953125,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 6,
      "prompt": "Name one reason GPU reductions can be nondeterministic",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.0015869140625,
      "margin_batched": 0.0015869140625,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 7,
      "prompt": "Summarize the idea of speculative decoding",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.001220703125,
      "margin_batched": 0.001220703125,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 8,
      "prompt": "What does a neural network logit represent?",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.0015869140625,
      "margin_batched": 0.0015869140625,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 9,
      "prompt": "Explain RMSNorm in simple terms",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.0015869140625,
      "margin_batched": 0.0015869140625,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 10,
      "prompt": "Why do servers batch LLM requests?",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.0015869140625,
      "margin_batched": 0.0015869140625,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 11,
      "prompt": "Give a concise description of CUDA",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.00103759765625,
      "margin_batched": 0.00103759765625,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    },
    {
      "index": 12,
      "prompt": "What is the role of a verifier?",
      "top1_token_id": 16046,
      "top1_text": " stairs",
      "top2_token_id": 17192,
      "top2_text": " vendors",
      "batched_top1_token_id": 16046,
      "batched_top1_text": " stairs",
      "margin_single": 0.0015869140625,
      "margin_batched": 0.0015869140625,
      "epsilon_max_abs_delta": 0.0,
      "two_epsilon": 0.0,
      "argmax_stable": true
    }
  ],
  "note": "Diagnostic only: local HF single-vs-batched forwards, not a proof and not a hosted Tinker logit trace."
}