{
  "provenance": {
    "author": "flozi00",
    "source": "DEEP-TECHNICAL-REPORT.md",
    "prepared": "2026-10-13",
    "source_sha256": "98a81610455daae34cea933a8f306ebbef97a46ec1be17248d279c85f29ff37b",
    "status": "Selected reported context, not independent measurement verification. Throughput figures in the article come from the separately preserved CSVs."
  },
  "inspected_public_sources": {
    "plugin_commit": "8d09f300f4dc1542081c39abe7521f1e29508f5a",
    "qwen_config_revision": "de4b8e4d43b917e7706784d8bb445c9af86a3540",
    "upstream_padding_pr": "https://github.com/vllm-project/vllm/pull/46201"
  },
  "reported_environment": {
    "amd": {
      "hardware": "1 x AMD Instinct MI350P",
      "tp": 1,
      "checkpoint": "custom MXFP4",
      "checkpoint_size_gib": 115,
      "vllm": "0.31.1rc1.dev88",
      "image_label": "nightly-81198e97",
      "rocm": "7.2.3",
      "aiter": "0.1.24.post1"
    },
    "nvidia": {
      "hardware": "2 x RTX PRO 6000 Blackwell Server Edition",
      "tp": 2,
      "checkpoint": "NVFP4",
      "vllm_context": "0.30.0; complete optimized-deployment manifest not in supplied ledger"
    },
    "warmup": "Report states 20 seconds per concurrency level; archived harness only implements one 32-token warm-up request. Additional manual warm-up not verifiable.",
    "omitted_manifest_fields": [
      "checkpoint hashes",
      "complete launch flags",
      "container image digests",
      "replayed prompt seeds",
      "time-aligned GPU telemetry"
    ]
  },
  "reported_microbenchmarks": {
    "historical_nvidia_stock_aggregate_tps": {
      "values": [198, 711, 854],
      "status": "Older repetition-corpus harness; no matching rows in the supplied October ledger; excluded from comparable-harness figures."
    },
    "gdn_mfma_c32_layer_us": {
      "before": 181,
      "after": 127.6
    },
    "qsa_c32_step_ms": {
      "before": 2.85,
      "after": 0.7
    },
    "moe_route_us": [
      {
        "M": 5,
        "before": 21.7,
        "after": 11.7
      },
      {
        "M": 40,
        "before": 30.3,
        "after": 14.1
      },
      {
        "M": 64,
        "before": 33,
        "after": 14.5
      }
    ],
    "rejected_gdn_chunk_rewrite_c32_us": {
      "candidate": 204.5,
      "reference": 177.5,
      "VGPRs_approx": 180
    },
    "rejected_hc_single_launch_down_us": {
      "candidate_range": [
        24,
        28
      ],
      "reference_range": [
        6.1,
        6.4
      ]
    },
    "status": "Reported kernel timings; individual timing traces not part of the uploaded aggregate CSVs."
  },
  "reported_quality": {
    "dataset": "GSM8K",
    "questions": 1319,
    "nvidia": 0.9674,
    "amd_k38": 0.9689,
    "amd_iteration_range": [
      0.9606,
      0.9689
    ],
    "amd_post_cutover": 0.9583,
    "post_cutover_inside_stated_iteration_range": false,
    "status": "Scores reported in notes, not in CSV. Per-example predictions/evaluation manifests unavailable; no paired significance or equivalence test performed."
  },
  "reported_production": {
    "cutover_date": "2026-10-11",
    "failed_probes": 0,
    "observed_aggregate_tps": [
      229.6,
      999.2,
      2145.7
    ],
    "status": "Reported only. No corresponding post-cutover CSV rows or request-level availability log supplied."
  },
  "reported_thermal_context": {
    "sustained_temperature_junction_c": 100,
    "hbm_temperature_c_range": [
      111,
      112
    ],
    "observed_power_w_range": [
      450,
      520
    ],
    "configured_cap_w": 600,
    "sustained_clock_mhz_range": [
      1420,
      1620
    ],
    "status": "Not present as time-aligned samples in supplied CSVs; cannot establish sustained capacity from these short sweeps."
  },
  "review_corrections": [
    "40 labels include two NVIDIA sweeps and invalid/candidate AMD configurations.",
    "282.0 / 1206.2 / 2405.1 combines k38 C1, k41 C8 and k39 C32.",
    "184.7 / 776.9 / 1627.7 combines two NVIDIA sweeps.",
    "Full target state writes at the declared dimensions are 16.875 GiB at C32, not 4.6 GiB.",
    "181 -> 127.6 us means 29.5% less time or 1.42x isolated throughput.",
    "Split-K 5 changes a 40-block grid to 200 unless tile geometry also changes.",
    "Streaming ITL is a token-normalized chunk-gap proxy.",
    "Zero failed probes is not a verified zero-request-error production cutover."
  ]
}