docker-validation.json

{
  "purpose": "Fresh-container capacity and regression checks, not replacement headline benchmarks",
  "date": "2026-09-16",
  "hardware": {
    "gpus": "2x RTX 3090 24 GB",
    "ram": "128 GB, 8-channel DDR4-3200"
  },
  "runtime": {
    "build_source_commit": "c2529343cbee0ac9a897ae5a38baf3982ba5220f",
    "deployment": "Public Dockerfile built from an empty Docker image cache; public docker_serve.sh; read-only checkpoint mount; no native runtime mounts",
    "base_image_digest": "sha256:fc120ece0a388cc0aa1caad4a9f1cd92113484ab7ec2fd0efadd62585be05bf8",
    "local_image_index_digest": "sha256:feedb4420154bda05f69cc39f55633e7e7f0b317c5a8b20675c2b5a3c5e28221",
    "local_image_amd64_manifest_digest": "sha256:98f6c694c8781fe9b9a5940fd3976ab82ef6280f931fc6fe2d5b4f77ddbb656f",
    "checkpoint_revision": "ef554143369a706525336f6b42a09094835dc077",
    "checkpoint_check": "Target/MTP indices, shard headers, tensor counts and payload byte totals passed; tensor payload hashes were not repeated in this run",
    "os": "Ubuntu 26.04 LTS",
    "driver": "595.84",
    "docker": "29.1.3",
    "nvidia_container_toolkit": "1.20.0",
    "torch": "2.13.0+cu130",
    "cuda": "13.0",
    "vllm": "0.1.dev20073+g8e685d198",
    "humming_kernels": "0.1.12",
    "overlay_files_verified_inside_image": 29,
    "overlay_manifest_sha256": "55c9bde05764d2156e6167095a76a07c60f02d7baef11ec01e78e81cfbaf2921",
    "gpu_free_tests_passed_inside_image": 32,
    "cuda_peer_access": "Both directions passed capability checks and tensor-copy equality checks inside the image",
    "custom_all_reduce": false,
    "allocator": "expandable_segments:True",
    "kv_dtype": "bfloat16",
    "qsa": "approximate, with the 64 MiB workspace fix",
    "max_parallel_loading_workers": "1 passed by launcher; ignored by pinned runtime",
    "swap": "8 GiB original plus 56 GiB temporary loading swap; temporary file removed after each server"
  },
  "two_clients": {
    "settings": {
      "MAX_NUM_SEQS": 2,
      "MAX_MODEL_LEN": 131072,
      "MAX_NUM_BATCHED_TOKENS": 2048,
      "KV_CACHE_MEMORY_BYTES": 4697620480,
      "VLLM_WNA16_STATIC_HOT_CACHE_SIZE": 80,
      "MTP_DEPTH": 3
    },
    "gpu_kv_tokens": 263416,
    "startup_seconds": 474.078613759,
    "graph_capture_sizes": "Default, not manually restricted",
    "request_recipe": "repo-chat; second client rotates the public source corpus by half its characters; unique cache salt per request; greedy forced-length outputs",
    "warmup_policy": "One short pair followed by one full-context pair; no repeated statistical runs",
    "short": {
      "requests": 2,
      "input_tokens_each": 1024,
      "output_tokens_each": 2048,
      "exact_counts_passed": true,
      "ttft_seconds": [5.595631230, 5.595673189],
      "individual_post_first_chunk_tps": [45.838999912, 49.348504339],
      "overlapping_decode_seconds": 41.480472803,
      "overlapping_output_tokens": [1812, 2047],
      "overlapping_aggregate_tps": 93.031726478
    },
    "full128k": {
      "requests": 2,
      "input_tokens_each": 129024,
      "output_tokens_each": 2048,
      "exact_counts_passed": true,
      "wall_seconds": 404.180576801,
      "ttft_seconds": [180.075548741, 356.006889247],
      "individual_post_first_chunk_tps_including_other_client_prefill": [9.283074617, 42.509164195],
      "overlapping_decode_seconds": 44.583164454,
      "overlapping_output_tokens": [1830, 1821],
      "overlapping_aggregate_tps": 81.891898988,
      "peak_kv_usage_fraction": 0.9708737864,
      "peak_running_requests": 2,
      "preemptions": 0
    },
    "isolation_checks": {
      "sequential": "2/2 passed",
      "concurrent": "2/2 passed"
    },
    "mtp_counters_whole_suite": {
      "drafted_tokens": 9474,
      "accepted_tokens": 5141
    },
    "allocation_retries_loading": 2,
    "allocation_retries_inference": 0,
    "peaks": {
      "cpu_c": 82.25,
      "gpu_c": [70, 75],
      "gpu_w": [283.06, 294.54],
      "gpu_mib": [23211, 23643],
      "min_host_available_gib": 14.195163727,
      "max_host_swap_used_gib": 52.853134155
    },
    "cleanup": "Server stopped normally; temporary swap removed; original swap retained"
  },
  "single_request_hot84": {
    "settings": {
      "MAX_NUM_SEQS": 1,
      "MAX_MODEL_LEN": 262144,
      "MAX_NUM_BATCHED_TOKENS": 4096,
      "KV_CACHE_MEMORY_BYTES": 4429185024,
      "VLLM_WNA16_STATIC_HOT_CACHE_SIZE": 84,
      "MTP_DEPTH": 3
    },
    "settings_note": "Image defaults, with no hot-cache or capacity overrides",
    "gpu_kv_tokens": 276313,
    "startup_seconds": 467.701298013,
    "request_recipe": "repo-chat, unique cache salt per request, greedy forced-length outputs",
    "probe_order": [[262016, 128], [1024, 128], [128, 1024]],
    "first_user_request_is_full_context": true,
    "all_exact_counts_passed": true,
    "ttft_seconds_in_probe_order": [229.853154168, 2.534141732, 1.864057507],
    "request_duration_seconds_in_probe_order": [231.622007849, 4.137608428, 16.643088296],
    "post_first_chunk_tps_in_probe_order": [71.803638258, 79.210467235, 69.220256268],
    "preemptions": 0,
    "mtp_counters_whole_suite": {
      "drafted_tokens": 1527,
      "accepted_tokens": 768
    },
    "allocation_retries_loading": 2,
    "allocation_retries_inference": 0,
    "peaks": {
      "cpu_c": 82.375,
      "gpu_c": [70, 74],
      "gpu_w": [285.17, 292.28],
      "gpu_mib": [23955, 23953],
      "min_host_available_gib": 24.039257050,
      "max_host_swap_used_gib": 53.309982300
    },
    "cleanup": "Server stopped normally; temporary swap removed; original swap retained",
    "limits": "Fresh server, not cold filesystem. First-use JIT warnings occurred during inference. No explicit warmup or repeats; short outputs are capacity evidence, not sustained-decode benchmarks."
  },
  "limits": [
    "One server launch per profile and one run per shape on one host; no universal driver or display-memory guarantee.",
    "Decode overlap is the intersection of both streams' generation intervals; do not sum rates whose intervals differ.",
    "The first long client's generation overlaps the second client's prefill. Its whole-generation rate is not isolated decode throughput.",
    "Forced output lengths and small isolation checks do not constitute a model-quality evaluation.",
    "Fresh container startup includes kernel compilation. The existing checkpoint and host filesystem cache were reused.",
    "Host swap was used and sample vmstat checks observed page-ins; these are not fully RAM-resident performance claims."
  ]
}