{
 "title": "GLM-5.3-Flash on two DGX Sparks: four serving stacks",
 "recorded": "2026-10-01 to 2026-10-03",
 "hardware": "2x DGX Spark (GB10 128 GB), TP2 over 200 Gb RoCE",
 "thinking": "on for all speed rows, island and volcano; thinking-off row uses Kindling's own gate decode test",
 "stacks": {
  "kindling-nvfp4": "kindlingai/glm-5.3-flash-gx10 @ d5abfe1, nvidia/GLM-5.3-Flash-NVFP4, DFlash2 drafter; vLLM nightly fork; smoketest 8/8 (first boot; second boot's smoketest hit a test-script temp-file bug, 0/8 reported, responses correct)",
  "mia-exl3": "MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks @ 6278ecb, published image exl3-instanttensor, default profile (MAX_NUM_SEQS=4), DFlash2 @ dc77ff1; port 8889 (8888 on spark-78f1 belongs to an unrelated F*EX service); source rebuild at 6278ecb failed Mia's own scheduler self-test, so the published GHCR image was used",
  "tf-exl3": "TensorFold 0.6.2 @ 56e2e3e, Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw @ 25a44fd, incoai DFlash2 @ bf582e4 (auto policy), one request at a time",
  "tf-mlx4": "TensorFold 0.6.2 @ 56e2e3e, TensorFold/GLM-5.3-Flash-MLX-4bit-MTP @ d2b17f5, incoai DFlash2 @ bf582e4, one request at a time"
 },
 "drafter_license": "incoai/GLM-5.3-Flash-DFlash2 is CC BY-NC-ND",
 "speed": {
  "kindling-nvfp4": {
   "c1_agg": 52.6,
   "c1_ttft": 0.382,
   "c4_agg": 66.9,
   "c4_ttft": 0.838,
   "c8_agg": 86.8,
   "c8_ttft": 1.151,
   "16k_prompt": 15987,
   "16k_ttft": 12.0174,
   "16k_prefill": 1330.3,
   "16k_decode": 41.06,
   "64k_prompt": 63994,
   "64k_ttft": 31.3019,
   "64k_prefill": 2044.4,
   "64k_decode": 39.65,
   "128k_prompt": 128005,
   "128k_ttft": 81.1203,
   "128k_prefill": 1578.0,
   "128k_decode": 40.38,
   "64k_replay_ttft": 2.4585,
   "4x32k_wall": 69.3,
   "exact_rr": 8,
   "exact_conc": 0,
   "gate_thinkoff": {
    "structured": 93.3,
    "code": 70.4,
    "prose": 39.6
   }
  },
  "mia-exl3": {
   "c1_agg": 41.9,
   "c1_ttft": 0.318,
   "c4_agg": 56.3,
   "c4_ttft": 0.652,
   "c8_agg": 59.7,
   "c8_ttft": 29.757,
   "16k_prompt": 15990,
   "16k_ttft": 10.4245,
   "16k_prefill": 1533.9,
   "16k_decode": 31.22,
   "64k_prompt": 63994,
   "64k_ttft": 41.0275,
   "64k_prefill": 1559.8,
   "64k_decode": 31.61,
   "128k_prompt": 128004,
   "128k_ttft": 83.0265,
   "128k_prefill": 1541.7,
   "128k_decode": 30.8,
   "64k_replay_ttft": 2.3651,
   "4x32k_wall": 114.0,
   "exact_rr": 0,
   "exact_conc": 0,
   "gate_thinkoff": {
    "structured": 70.9,
    "code": 52.4,
    "prose": 23.6
   }
  },
  "tf-exl3": {
   "c1_agg": 43.1,
   "c1_ttft": 0.25,
   "c4_agg": 36.5,
   "c4_ttft": 24.12,
   "c8_agg": 37.1,
   "c8_ttft": 66.816,
   "16k_prompt": 15986,
   "16k_ttft": 28.9358,
   "16k_prefill": 552.5,
   "16k_decode": 31.09,
   "64k_prompt": 63995,
   "64k_ttft": 102.0584,
   "64k_prefill": 627.0,
   "64k_decode": 33.79,
   "128k_prompt": 128003,
   "128k_ttft": 209.7061,
   "128k_prefill": 610.4,
   "128k_decode": 36.04,
   "64k_replay_ttft": 0.1466,
   "4x32k_wall": 233.3,
   "exact_rr": 8,
   "exact_conc": 8,
   "gate_thinkoff": null
  },
  "tf-mlx4": {
   "c1_agg": 57.5,
   "c1_ttft": 0.228,
   "c4_agg": 53.4,
   "c4_ttft": 19.916,
   "c8_agg": 53.2,
   "c8_ttft": 30.172,
   "16k_prompt": 15989,
   "16k_ttft": 13.9433,
   "16k_prefill": 1146.7,
   "16k_decode": 50.85,
   "64k_prompt": 63995,
   "64k_ttft": 51.0851,
   "64k_prefill": 1252.7,
   "64k_decode": 43.16,
   "128k_prompt": 128001,
   "128k_ttft": 108.6047,
   "128k_prefill": 1178.6,
   "128k_decode": 48.34,
   "64k_replay_ttft": 0.1307,
   "4x32k_wall": 123.0,
   "exact_rr": 8,
   "exact_conc": 8,
   "gate_thinkoff": {
    "structured": 83.8,
    "code": 68.9,
    "prose": 40.2
   }
  }
 },
 "volcano": {
  "prompt": "Create a single self-contained HTML file using Three.js that renders a visually spectacular, bright, colorful, and highly detailed 3D animated volcanic eruption scene with glowing lava, thick ash, flying ejecta, and dramatic environmental effects.\n\nRequirements:\n\n* Use Three.js loaded from a CDN.\n\n* Everything must be contained in one HTML file and work directly when opened in a modern browser.\n\n* Fill the entire browser viewport edge to edge with the 3D scene. Do not leave margins, borders, blank space, or unused areas.\n\n* Create a bright, polished, AAA-quality stylized 3D visual style with vibrant colors, detailed materials, cinematic lighting, soft shadows, atmospheric depth, and excellent visibility.\n\n* Create a large detailed volcanic island landscape surrounded by colorful terrain, a tropical coastline, turquoise sea, black sand beaches, jungle vegetation, a small village with huts and a wooden pier near the shore, and distant smaller islands.\n\n* Make the volcano the clear central focus of the scene, large enough to see important details such as the crater rim, layered rock strata, old solidified lava channels, fumaroles, and vegetation climbing the lower slopes.\n\n* Position the volcano in the center of the island at the beginning of the animation, calm and dormant, with a thin wisp of steam rising from the crater.\n\n* Include a small research station or observation post on the lower slope with instruments, antennas, and warning lights.\n\n* Start with a brief calm pre-eruption moment with subtle atmospheric movement such as light steam drifting, birds circling, gentle waves on the shore, and warning lights beginning to blink.\n\n* Show the ground beginning to tremble, with small rocks tumbling down the slopes and dust shaking loose from the crater rim.\n\n* Show the crater gradually beginning to glow from within, brightening from dull red to intense orange-white as pressure builds.\n\n* Create bright, powerful animated lava fountains erupting from the crater with multiple layers, including a bright white-yellow core, glowing orange lava blobs, and long arcing trails of molten rock.\n\n* Add intense heat glow and warm light from the lava reflecting onto the crater walls, the ash cloud, the nearby slopes, and the research station.\n\n* Create large volumes of thick animated ash and steam billowing upward from the crater into a towering eruption column, expanding outward at the top into a broad anvil shape.\n\n* Make the ash cloud roll, expand, swirl, and rise naturally instead of appearing as simple static clouds, with the underside lit orange by the lava.\n\n* Add glowing volcanic bombs and ejecta launched in ballistic arcs from the crater, leaving glowing trails and landing on the slopes with bursts of sparks and dust.\n\n* Add smaller particles, sparks, glowing embers, ash flakes, and drifting cinders throughout the eruption area.\n\n* Show rivers of glowing lava breaking over the crater rim and flowing down the flanks in branching channels, with a bright liquid core and a cooling darker crust forming on the surface as it advances.\n\n* Make nearby trees, grass, and vegetation react as the lava and hot ash approach, bending, catching fire, and burning.\n\n* Show steam explosions where lava reaches the sea, with white steam plumes and hissing spray.\n\n* Add visible atmospheric distortion or heat shimmer above the lava and crater if possible.\n\n* Add occasional volcanic lightning flickering inside the ash column.\n\n* Include subtle camera vibration and a brief cinematic camera shake during the initial tremor and the main eruption without making the scene difficult to watch.\n\n* Use a smooth cinematic camera that begins close to the coastline and village, then slowly pulls back and rises to reveal the full eruption, the lava rivers, and the towering ash column, while keeping the volcano and major effects clearly visible.\n\n* As the camera rises, smoothly reveal more of the colorful island, the coastline, the sea, clouds, and the distant horizon.\n\n* Keep the volcano large and prominent throughout the sequence.\n\n* Add layered clouds at different heights, with the eruption column eventually climbing toward and punching through them.\n\n* Use a bright blue daytime sky with colorful white clouds and strong sunlight so the lava, ash, glow, and environment remain clearly visible.\n\n* Avoid a dark night scene. The result should be bright, vibrant, high-contrast, and visually clear.\n\n* Use realistic-looking shadows and dynamic lighting, including the changing influence of the bright lava glow.\n\n* Keep the viewport visually dense and impressive throughout the animation, with no excessive empty sky or unused space.\n\n* Use layered terrain, sea, vegetation, ash, clouds, and atmospheric effects to maintain visual interest as the camera pulls back.\n\n* The animation should follow a clear sequence: calm volcano with steam wisp → warning lights blink → ground trembles and rocks tumble → crater glows brighter → lava fountain erupts → ash column rises → volcanic bombs arc outward → lava rivers flow down the flanks → vegetation burns → lava reaches the sea and steam explodes → ash column punches through the clouds with lightning → camera pulls back to reveal the whole island.\n\n* Use smooth animation for the lava, ash, ejecta, sparks, particles, vegetation, sea, steam, clouds, lighting, shadows, and camera movement.\n\n* Do not add overlay text, titles, labels, HUD, captions, menus, buttons, borders, or UI elements over the scene.\n\n* The browser viewport should contain only the animated 3D volcanic eruption scene, filling the entire screen edge to edge.\n\n* Start the animation automatically when the page loads.\n\n* Add smooth camera controls only if they do not obstruct the scene or interfere with the cinematic composition.\n\n* Optimize for smooth performance in a modern browser.\n\n* No build process is allowed, and there must be no external dependencies other than the Three.js CDN.",
  "sampling": {
   "temperature": 1.0,
   "top_p": 0.95
  },
  "max_tokens": "serving window",
  "recording": "headless Chrome, fixed-step 24 fps virtual clock, 75 s from the first frame, no interaction; 1920x1080 scene (landscape shows more of each scene with no code change; vertical 1080x1200 captures also kept)",
  "arms": {
   "tf-exl3": {
    "visual": "volcano",
    "label": "tf-exl3",
    "base": "http://10.0.0.109:8080/v1",
    "model": "GLM-5.3-Flash-EXL3-TR3-4bpw",
    "prompt_tokens": 1172,
    "completion_tokens": 91944,
    "reasoning_chars": 225836,
    "content_chars": 69073,
    "wall_s": 3113.9,
    "ttft_s": 3.88,
    "decode_tok_s": 29.6,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 66818,
    "html_lines": 1550,
    "max_tokens": 162584,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-02 23:24:59",
    "original": "/lab/glm5.3local/volcano-four-stacks/tf-exl3/",
    "original_sha256": "a67fdfbc296b59943ef314d184e20abb94280f4fe1c36dc86cfed4684feb091b",
    "repairs": 1
   },
   "kindling-nvfp4": {
    "visual": "volcano",
    "label": "kindling-nvfp4",
    "base": "http://10.0.0.109:8002/v1",
    "model": "glm53",
    "prompt_tokens": 1172,
    "completion_tokens": 62104,
    "reasoning_chars": 139129,
    "content_chars": 55442,
    "wall_s": 1561.2,
    "ttft_s": 2.24,
    "decode_tok_s": 39.8,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 52841,
    "html_lines": 1229,
    "max_tokens": 162584,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-03 00:50:47",
    "original": "/lab/glm5.3local/volcano-four-stacks/kindling-nvfp4/",
    "original_sha256": "17f3a5d2084bbd638dd6bde3bf65fa482ca4c797f78285b3082659d92e48b424",
    "repairs": 0
   },
   "tf-mlx4": {
    "visual": "volcano",
    "label": "tf-mlx4",
    "base": "http://10.0.0.109:8080/v1",
    "model": "TensorFold-GLM-5.3-Flash-MLX-4bit-MTP",
    "prompt_tokens": 1172,
    "completion_tokens": 112773,
    "reasoning_chars": 287588,
    "content_chars": 73303,
    "wall_s": 2506.4,
    "ttft_s": 1.58,
    "decode_tok_s": 45.0,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 71447,
    "html_lines": 1590,
    "max_tokens": 162584,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-03 00:09:32",
    "original": "/lab/glm5.3local/volcano-four-stacks/tf-mlx4/",
    "original_sha256": "57e59cd5e69a288691f80cd4e090f4713df7b4d1e8f747f832573689d46ca98d",
    "repairs": 0
   },
   "mia-exl3": {
    "visual": "volcano",
    "label": "mia-exl3",
    "base": "http://10.0.0.109:8889/v1",
    "model": "GLM-5.3-Flash-EXL3",
    "prompt_tokens": 1172,
    "completion_tokens": 65912,
    "reasoning_chars": 146079,
    "content_chars": 64300,
    "wall_s": 2560.7,
    "ttft_s": 1.33,
    "decode_tok_s": 25.8,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 61842,
    "html_lines": 1312,
    "max_tokens": 162584,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-02 22:11:19",
    "original": "/lab/glm5.3local/volcano-four-stacks/mia-exl3/",
    "original_sha256": "775b6123a2904e6ffb16f13552affcfbe8e4f39285a6b0093fdcb1e97483d324",
    "repairs": 0
   }
  },
  "ranking_by_eye": [
   "tf-exl3",
   "kindling-nvfp4",
   "tf-mlx4",
   "mia-exl3"
  ]
 },
 "island": {
  "prompt": "Create an interactive tropical island weather simulator in Three.js. Let me switch between sunshine, rain, thunderstorms with lightning, and a hurricane. Make the island and surrounding ocean react to the weather, with smooth transitions and convincing motion. Include orbit and zoom controls. Make it mobile-friendly with touch controls, easy-to-tap weather buttons, a responsive layout, and smooth performance on phones. Make it visually impressive, and choose your own style and details.\n\nDeliver everything in one HTML file, with no external libraries besides Three.js and its official addons.",
  "recording": "fixed-step 24 fps, 30 s, weather buttons clicked at 0/6/13/20/27 s (sun, rain, storm, hurricane, sun)",
  "arms": {
   "tf-exl3": {
    "label": "tf-exl3",
    "base": "http://10.0.0.109:8080/v1",
    "model": "GLM-5.3-Flash-EXL3-TR3-4bpw",
    "prompt_tokens": 122,
    "completion_tokens": 72014,
    "reasoning_chars": 175797,
    "content_chars": 53327,
    "wall_s": 2324.9,
    "ttft_s": 0.7,
    "decode_tok_s": 31.0,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 50612,
    "html_lines": 1158,
    "max_tokens": 163634,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-02 12:18:12",
    "repairs": 0,
    "original": "/lab/glm5.3local/island-weather-four-stacks/tf-exl3/"
   },
   "kindling-nvfp4": {
    "label": "kindling-nvfp4",
    "base": "http://10.0.0.109:8002/v1",
    "model": "glm53",
    "prompt_tokens": 122,
    "completion_tokens": 60832,
    "reasoning_chars": 147185,
    "content_chars": 42491,
    "wall_s": 1474.0,
    "ttft_s": 0.38,
    "decode_tok_s": 41.3,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 39996,
    "html_lines": 832,
    "max_tokens": 163634,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-02 10:51:39",
    "repairs": 2,
    "original": "/lab/glm5.3local/island-weather-four-stacks/kindling-nvfp4/"
   },
   "tf-mlx4": {
    "label": "tf-mlx4",
    "base": "http://10.0.0.109:8080/v1",
    "model": "TensorFold-GLM-5.3-Flash-MLX-4bit-MTP",
    "prompt_tokens": 122,
    "completion_tokens": 98082,
    "reasoning_chars": 262769,
    "content_chars": 55441,
    "wall_s": 2168.4,
    "ttft_s": 0.62,
    "decode_tok_s": 45.2,
    "finish_reason": "stop",
    "error": null,
    "html_bytes": 53281,
    "html_lines": 1185,
    "max_tokens": 163634,
    "sampling": {
     "temperature": 1.0,
     "top_p": 0.95
    },
    "thinking": "on",
    "when": "2026-10-02 13:11:52",
    "repairs": 2,
    "original": "/lab/glm5.3local/island-weather-four-stacks/tf-mlx4/"
   }
  },
  "not_run": {
   "mia-exl3": "Mia's kit joined after the island round; it ran the volcano only"
  }
 },
 "incidents": [
  "A first Mia speed run was invalidated: two copies of the harness started together after a relaunch and shared the server. All Mia numbers here come from the single clean rerun started 2026-10-02 21:11.",
  "TensorFold ignores Kindling's chat_template_kwargs thinking:false key; the first TF thinking-off pass was thinking. The published TF row uses enable_thinking:false and was checked to emit zero reasoning tokens.",
  "TensorFold EXL3 was not run on the thinking-off gate test."
 ],
 "caveats": [
  "One generation per stack per prompt at temperature 1.0: the island and volcano ranks describe these runs, not the quantization formats.",
  "TensorFold 0.6.2 serves GLM one request at a time, so its 4- and 8-at-once rows are queueing.",
  "Mia's default profile runs four requests at a time (MAX_NUM_SEQS=4); requests 5-8 queue.",
  "Kindling vs Mia is NVFP4 vs EXL3 on vLLM, but two different vLLM builds with their own patches, and Mia's kit pins a newer DFlash2 revision (dc77ff1 vs bf582e4)."
 ]
}