{
  "schema": "native-glm-nccl-comparison/v1",
  "cells": [
    {
      "cell": "4k",
      "requests_each": 12,
      "baseline_passed": true,
      "candidate_passed": true,
      "actual_prompt_tokens_p50": 2970.0,
      "metrics": {
        "ttft_p50_ms": {
          "baseline": 544.4157719994109,
          "candidate": 495.45400499937386,
          "percent_change": -8.99345124044093
        },
        "e2e_p50_ms": {
          "baseline": 688.4723299999678,
          "candidate": 637.2978179997517,
          "percent_change": -7.433052828748899
        },
        "decode_tok_s_p50": {
          "baseline": 282.1348827289961,
          "candidate": 288.14985681734345,
          "percent_change": 2.1319498071867393
        }
      }
    },
    {
      "cell": "120k",
      "requests_each": 3,
      "baseline_passed": true,
      "candidate_passed": true,
      "actual_prompt_tokens_p50": 97159.0,
      "metrics": {
        "ttft_p50_ms": {
          "baseline": 16406.720802000564,
          "candidate": 14941.19061299989,
          "percent_change": -8.932499106231962
        },
        "e2e_p50_ms": {
          "baseline": 16569.573079000293,
          "candidate": 15101.824650999333,
          "percent_change": -8.858094418021745
        },
        "decode_tok_s_p50": {
          "baseline": 251.76088780022232,
          "candidate": 255.2379845827617,
          "percent_change": 1.3811107884631024
        }
      }
    }
  ],
  "limitations": [
    "Only 12 short and 3 medium requests per arm; no service-tail or broad inference throughput claim.",
    "Strict count failed baseline 380k cell; no high-context performance headline.",
    "32-word fixed output; marker quality diagnostic only, two repetitions.",
    "Unique prefix requested; cache metadata incomplete, cold cache not established."
  ]
}
