{
 "meta": {
  "generated_by": "vllm/build-study-dataset.py",
  "campaign": "2026-08-15 .. 2026-08-19",
  "rig": {
   "gpus": "2x modded RTX 4090 48GB (96GB total)",
   "compute_capability": 89,
   "gpu0_link": "PCIe x16 (display card)",
   "gpu1_link": "PCIe x4",
   "system_ram_gb": 128,
   "os": "Windows 11 + Docker Desktop (WSL2)",
   "shared_memory_per_block_bytes": 101376,
   "power_policy": "330W cap + core clock locked 0-2200MHz (stability requirement on modded cards)"
  },
  "engine": {
   "image": "vllm/vllm-openai@sha256:ffb2d59b1c059a5bd8d781320c9f5189de8293693b7d95da54befddaa54abf52",
   "vllm": "0.26.0",
   "notes": [
    "pinned by digest - :latest self-upgraded 0.25.1->0.26.0 once and invalidated a week of data",
    "system_fingerprint varies with serving config; it is NOT engine identity",
    "MTP speculative decoding deadlocks on long cold prefill on 0.26.0 while /health returns 200"
   ]
  },
  "vram_column_guidance": "Use weights_gb as the footprint column, NOT vram_while_serving_mb. At --gpu-memory-utilization 0.92 vLLM allocates 92% of each card BY DESIGN (weights + preallocated KV fill to the target), so measured serving VRAM is ~45GB/card for every model regardless of size - it measures the setting, not the model. This is the same instrument error the first MXFP4 verdict tool made (see instrument_failures). weights_gb is validated 8/8 exact against shard indexes. gpt-oss's vram_while_serving_mb is retained as the MXFP4-existence-proof exhibit, not as a comparable column.",
  "units": {
   "decode_tok_s": "tokens/second single stream",
   "load_s": "container start to /v1/models ready",
   "ctx": "measured usage.prompt_tokens unless stated",
   "draw_w": "sum across cards the model occupies"
  }
 },
 "models": {
  "Devstral-Small-2-24B-AWQ": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T07:44:49",
   "weights_gb": 32.1,
   "architecture": "Mistral3ForConditionalGeneration",
   "native_ctx": 393216,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 699
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": null,
   "thinking_kwarg": null,
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "GLM-4.7-Flash-GPTQ": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T07:56:36",
   "weights_gb": 16.6,
   "architecture": "Glm4MoeLiteForCausalLM",
   "native_ctx": 202752,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "auto",
    "gpu_util": "0.92",
    "load_s": 441
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 1.0
   },
   "thinking_kwarg": "enable_thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": 4,
   "reason_excluded": null
  },
  "LFM2.5-1.2B-Instruct": {
   "status": "VERIFIED",
   "onboarded": "2026-08-16T19:31:41",
   "weights_gb": 2.3,
   "architecture": "Lfm2ForCausalLM",
   "native_ctx": 128000,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 122
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": null,
   "thinking_kwarg": "thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "Llama-3.3-70B-AWQ": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T08:13:56",
   "weights_gb": 39.8,
   "architecture": "LlamaForCausalLM",
   "native_ctx": 131072,
   "verified": {
    "max_model_len": 21110,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 940
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 0.6,
    "top_p": 0.9
   },
   "thinking_kwarg": null,
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "Llama-3.3-70B-FP8": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T11:15:05",
   "weights_gb": 72.7,
   "architecture": "LlamaForCausalLM",
   "native_ctx": 131072,
   "verified": {
    "max_model_len": 32768,
    "tp": 2,
    "gpus": "0,1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1536
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 0.6,
    "top_p": 0.9
   },
   "thinking_kwarg": null,
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "Llama-4-Scout-INT4": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T05:27:49",
   "weights_gb": 64.9,
   "architecture": "Llama4ForConditionalGeneration",
   "native_ctx": 10485760,
   "verified": {
    "max_model_len": 32768,
    "tp": 2,
    "gpus": "0,1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1291
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 0.6,
    "top_p": 0.9
   },
   "thinking_kwarg": null,
   "template_handles_tools": true,
   "moe_experts_per_tok": 1,
   "reason_excluded": null
  },
  "NVIDIA-Nemotron-3-Super-120B-A12B-AWQ-4bit": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T12:58:07",
   "weights_gb": 80.7,
   "architecture": "NemotronHForCausalLM",
   "native_ctx": 262144,
   "verified": {
    "max_model_len": 32768,
    "tp": 2,
    "gpus": "0,1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1816
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 1.0,
    "top_p": 0.95
   },
   "thinking_kwarg": "enable_thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": 22,
   "reason_excluded": null
  },
  "Nemotron-3.5-Lightning-30B-FP8": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T03:34:38",
   "weights_gb": 32.3,
   "architecture": "NemotronHForCausalLM",
   "native_ctx": 262144,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1002
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 1.0,
    "top_p": 0.95
   },
   "thinking_kwarg": "enable_thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": 6,
   "reason_excluded": null
  },
  "Qwen3.8-27B-AWQ": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T10:18:04",
   "weights_gb": 27.7,
   "architecture": "Qwen3_5ForConditionalGeneration",
   "native_ctx": 262144,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1470
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20
   },
   "thinking_kwarg": "enable_thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "Qwen3.8-27B-FP8": {
   "status": "VERIFIED",
   "onboarded": "2026-08-17T10:42:43",
   "weights_gb": 30.9,
   "architecture": "Qwen3_5ForConditionalGeneration",
   "native_ctx": 262144,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1522
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20
   },
   "thinking_kwarg": "enable_thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "Ternary-Bonsai-27B-AWQ": {
   "status": "VERIFIED",
   "onboarded": "2026-08-16T23:27:43",
   "weights_gb": 18.7,
   "architecture": "Qwen3_5ForConditionalGeneration",
   "native_ctx": 262144,
   "verified": {
    "max_model_len": 32768,
    "tp": 1,
    "gpus": "1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 608
   },
   "vram_while_serving_mb": null,
   "vendor_sampling": {
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20
   },
   "thinking_kwarg": "enable_thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": null,
   "reason_excluded": null
  },
  "gpt-oss-120b-official": {
   "status": "VERIFIED",
   "onboarded": "2026-08-18T09:20:39",
   "weights_gb": 65.2,
   "architecture": "GptOssForCausalLM",
   "native_ctx": 131072,
   "verified": {
    "max_model_len": 131072,
    "tp": 2,
    "gpus": "0,1",
    "kv_cache_dtype": "fp8",
    "gpu_util": "0.92",
    "load_s": 1334
   },
   "vram_while_serving_mb": {
    "gpu0_mb": 46821,
    "gpu1_mb": 45837
   },
   "vendor_sampling": null,
   "thinking_kwarg": "thinking",
   "template_handles_tools": true,
   "moe_experts_per_tok": 4,
   "reason_excluded": null
  }
 },
 "sweep": [
  {
   "model": "lfm2.5-1.2b",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 21,
   "decode_tok_s": 308.3,
   "ttft_s": 0.018,
   "load_s": 136,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "lfm2.5-1.2b",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 8286,
   "decode_tok_s": 303.5,
   "ttft_s": 0.367,
   "load_s": 136,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "lfm2.5-1.2b",
   "error": "HTTP Error 400: Bad Request"
  },
  {
   "model": "qwen3.8-27b",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 24,
   "decode_tok_s": 35.8,
   "ttft_s": 0.096,
   "load_s": 783,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.8-27b",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 7432,
   "decode_tok_s": 35.3,
   "ttft_s": 3.126,
   "load_s": 783,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.8-27b",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 29714,
   "decode_tok_s": 34.3,
   "ttft_s": 11.211,
   "load_s": 783,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "llama-3.3-70b",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 47,
   "decode_tok_s": 23.2,
   "ttft_s": 0.1,
   "load_s": 995,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "llama-3.3-70b",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 7455,
   "decode_tok_s": 22.2,
   "ttft_s": 8.514,
   "load_s": 995,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "llama-3.3-70b",
   "error": "HTTP Error 400: Bad Request"
  },
  {
   "model": "nemotron3-super-120b",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 28,
   "decode_tok_s": 63.9,
   "ttft_s": 0.085,
   "load_s": 1883,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "nemotron3-super-120b",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 7722,
   "decode_tok_s": 64.0,
   "ttft_s": 2.695,
   "load_s": 1883,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "nemotron3-super-120b",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 30861,
   "decode_tok_s": 63.6,
   "ttft_s": 10.852,
   "load_s": 1883,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "devstral-small-2-24b-awq",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 15,
   "decode_tok_s": 60.6,
   "ttft_s": 0.042,
   "load_s": 501,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "devstral-small-2-24b-awq",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3713,
   "decode_tok_s": 59.2,
   "ttft_s": 1.387,
   "load_s": 501,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "devstral-small-2-24b-awq",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 14891,
   "decode_tok_s": 55.8,
   "ttft_s": 4.454,
   "load_s": 501,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "llama-3.3-70b-fp8",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 47,
   "decode_tok_s": 19.1,
   "ttft_s": 0.093,
   "load_s": 1610,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "llama-3.3-70b-fp8",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3737,
   "decode_tok_s": 19.0,
   "ttft_s": 4.046,
   "load_s": 1610,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "llama-3.3-70b-fp8",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 14891,
   "decode_tok_s": 18.6,
   "ttft_s": 12.802,
   "load_s": 1610,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "llama-4-scout-int4",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 21,
   "decode_tok_s": 53.7,
   "ttft_s": 0.1,
   "load_s": 1305,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "llama-4-scout-int4",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3712,
   "decode_tok_s": 53.0,
   "ttft_s": 1.624,
   "load_s": 1305,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "llama-4-scout-int4",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 14866,
   "decode_tok_s": 52.0,
   "ttft_s": 5.103,
   "load_s": 1305,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "nemotron-3.5-lightning-30b-fp8",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 28,
   "decode_tok_s": 158.4,
   "ttft_s": 0.07,
   "load_s": 1140,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "nemotron-3.5-lightning-30b-fp8",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3861,
   "decode_tok_s": 158.5,
   "ttft_s": 0.74,
   "load_s": 1140,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "nemotron-3.5-lightning-30b-fp8",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 15444,
   "decode_tok_s": 156.8,
   "ttft_s": 0.983,
   "load_s": 1140,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.8-27b-fp8",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 24,
   "decode_tok_s": 16.6,
   "ttft_s": 0.127,
   "load_s": 1087,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.8-27b-fp8",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3870,
   "decode_tok_s": 16.5,
   "ttft_s": 1.047,
   "load_s": 1087,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.8-27b-fp8",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 15414,
   "decode_tok_s": 16.4,
   "ttft_s": 3.642,
   "load_s": 1087,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "ternary-bonsai-27b-awq",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 24,
   "decode_tok_s": 50.7,
   "ttft_s": 0.07,
   "load_s": 735,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "ternary-bonsai-27b-awq",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3870,
   "decode_tok_s": 50.2,
   "ttft_s": 1.625,
   "load_s": 735,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "ternary-bonsai-27b-awq",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 15414,
   "decode_tok_s": 49.1,
   "ttft_s": 5.452,
   "load_s": 735,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "gpt-oss-120b-official",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 78,
   "decode_tok_s": 153.4,
   "ttft_s": 1.282,
   "load_s": 1334,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "gpt-oss-120b-official",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3911,
   "decode_tok_s": 149.1,
   "ttft_s": 1.672,
   "load_s": 1334,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "gpt-oss-120b-official",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 15494,
   "decode_tok_s": 140.7,
   "ttft_s": 3.324,
   "load_s": 1334,
   "gpus": "0,1",
   "spec_decode": false
  },
  {
   "model": "glm-4.7-flash-gptq",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 17,
   "decode_tok_s": 134.5,
   "ttft_s": 0.031,
   "load_s": 516,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "glm-4.7-flash-gptq",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3863,
   "decode_tok_s": 127.8,
   "ttft_s": 0.374,
   "load_s": 516,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "glm-4.7-flash-gptq",
   "error": "no tokens generated"
  },
  {
   "model": "qwen3.6-27b (MTP off)",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 24,
   "decode_tok_s": 47.4,
   "ttft_s": 0.087,
   "load_s": 266,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.6-27b (MTP off)",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3870,
   "decode_tok_s": 46.9,
   "ttft_s": 1.656,
   "load_s": 266,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.6-27b (MTP off)",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 15414,
   "decode_tok_s": 45.9,
   "ttft_s": 5.505,
   "load_s": 266,
   "gpus": "1",
   "spec_decode": false
  },
  {
   "model": "qwen3.6-27b (MTP on)",
   "requested_ctx_label": 0,
   "measured_prompt_tokens": 24,
   "decode_tok_s": 51.1,
   "ttft_s": 0.157,
   "load_s": 336,
   "gpus": "1",
   "spec_decode": true
  },
  {
   "model": "qwen3.6-27b (MTP on)",
   "requested_ctx_label": 4000,
   "measured_prompt_tokens": 3870,
   "decode_tok_s": 44.6,
   "ttft_s": 1.969,
   "load_s": 336,
   "gpus": "1",
   "spec_decode": true
  },
  {
   "model": "qwen3.6-27b (MTP on)",
   "requested_ctx_label": 16000,
   "measured_prompt_tokens": 15414,
   "decode_tok_s": 47.8,
   "ttft_s": 6.473,
   "load_s": 336,
   "gpus": "1",
   "spec_decode": true
  }
 ],
 "power_clock": {
  "conditions": {
   "A": "330W cap + clock locked 0-2200 (production setting)",
   "B": "450W cap + clock locked 0-2200 (isolates power limit)",
   "C": "330W cap + clocks unlocked (isolates clock lock)"
  },
  "valid_rows": [
   {
    "timestamp": "2026-08-17T15:15:23",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.097,
    "decode_tok_s": 48.5,
    "peak_clock": 2190,
    "mean_draw": 60.5,
    "peak_draw": 178.4,
    "peak_temp": 49
   },
   {
    "timestamp": "2026-08-17T15:15:37",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 1.179,
    "decode_tok_s": 31.9,
    "peak_clock": 2190,
    "mean_draw": 78.8,
    "peak_draw": 257.8,
    "peak_temp": 51
   },
   {
    "timestamp": "2026-08-17T15:15:50",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.181,
    "decode_tok_s": 30.8,
    "peak_clock": 2190,
    "mean_draw": 64.5,
    "peak_draw": 230.8,
    "peak_temp": 51
   },
   {
    "timestamp": "2026-08-17T15:15:58",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.249,
    "decode_tok_s": 49.3,
    "peak_clock": 2190,
    "mean_draw": 69.1,
    "peak_draw": 235.1,
    "peak_temp": 51
   },
   {
    "timestamp": "2026-08-17T15:16:00",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.182,
    "decode_tok_s": 288.3,
    "peak_clock": 2190,
    "mean_draw": 100.8,
    "peak_draw": 192,
    "peak_temp": 51
   },
   {
    "timestamp": "2026-08-17T15:16:01",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.069,
    "decode_tok_s": 287.1,
    "peak_clock": 2190,
    "mean_draw": 216.1,
    "peak_draw": 216.8,
    "peak_temp": 52
   },
   {
    "timestamp": "2026-08-17T15:16:06",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.066,
    "decode_tok_s": 300.8,
    "peak_clock": 2190,
    "mean_draw": 125.7,
    "peak_draw": 228.1,
    "peak_temp": 52
   },
   {
    "timestamp": "2026-08-17T15:16:07",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.068,
    "decode_tok_s": 285,
    "peak_clock": 2190,
    "mean_draw": 225.5,
    "peak_draw": 235.7,
    "peak_temp": 52
   },
   {
    "timestamp": "2026-08-17T15:16:09",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.061,
    "decode_tok_s": 302.8,
    "peak_clock": 2190,
    "mean_draw": 226,
    "peak_draw": 235.7,
    "peak_temp": 53
   },
   {
    "timestamp": "2026-08-17T15:16:13",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.067,
    "decode_tok_s": 99,
    "peak_clock": 2190,
    "mean_draw": 210.7,
    "peak_draw": 231.7,
    "peak_temp": 53
   },
   {
    "timestamp": "2026-08-17T15:16:20",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.137,
    "decode_tok_s": 55.7,
    "peak_clock": 2190,
    "mean_draw": 39.2,
    "peak_draw": 78.8,
    "peak_temp": 50
   },
   {
    "timestamp": "2026-08-17T15:16:22",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.249,
    "decode_tok_s": 258.7,
    "peak_clock": 2190,
    "mean_draw": 120.8,
    "peak_draw": 236.4,
    "peak_temp": 53
   },
   {
    "timestamp": "2026-08-17T15:16:31",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.06,
    "decode_tok_s": 94.6,
    "peak_clock": 2655,
    "mean_draw": 87.9,
    "peak_draw": 200.5,
    "peak_temp": 54
   },
   {
    "timestamp": "2026-08-17T15:16:44",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.151,
    "decode_tok_s": 32,
    "peak_clock": 2655,
    "mean_draw": 96.7,
    "peak_draw": 301.9,
    "peak_temp": 56
   },
   {
    "timestamp": "2026-08-17T15:17:06",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.182,
    "decode_tok_s": 19.2,
    "peak_clock": 2655,
    "mean_draw": 52.5,
    "peak_draw": 301.2,
    "peak_temp": 52
   },
   {
    "timestamp": "2026-08-17T15:17:22",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.251,
    "decode_tok_s": 25.5,
    "peak_clock": 2655,
    "mean_draw": 43.1,
    "peak_draw": 121.4,
    "peak_temp": 49
   },
   {
    "timestamp": "2026-08-17T15:17:25",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.182,
    "decode_tok_s": 109.8,
    "peak_clock": 2655,
    "mean_draw": 121.2,
    "peak_draw": 298.9,
    "peak_temp": 50
   },
   {
    "timestamp": "2026-08-17T15:17:27",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.076,
    "decode_tok_s": 265.5,
    "peak_clock": 2655,
    "mean_draw": 124.2,
    "peak_draw": 283.5,
    "peak_temp": 54
   },
   {
    "timestamp": "2026-08-17T16:01:49",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.019,
    "decode_tok_s": 301.5,
    "peak_clock": 2190.0,
    "mean_draw": 122.3,
    "peak_draw": 212.2,
    "peak_temp": 50.0
   },
   {
    "timestamp": "2026-08-17T16:01:51",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.997,
    "decode_tok_s": 283.9,
    "peak_clock": 2190.0,
    "mean_draw": 253.1,
    "peak_draw": 267.7,
    "peak_temp": 56.0
   },
   {
    "timestamp": "2026-08-17T16:01:53",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.016,
    "decode_tok_s": 302.4,
    "peak_clock": 2190.0,
    "mean_draw": 233.3,
    "peak_draw": 235.9,
    "peak_temp": 51.0
   },
   {
    "timestamp": "2026-08-17T16:01:54",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.064,
    "decode_tok_s": 288.2,
    "peak_clock": 2190.0,
    "mean_draw": 228.4,
    "peak_draw": 232.0,
    "peak_temp": 52.0
   },
   {
    "timestamp": "2026-08-17T16:01:55",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.015,
    "decode_tok_s": 308.1,
    "peak_clock": 2190.0,
    "mean_draw": 234.8,
    "peak_draw": 237.5,
    "peak_temp": 52.0
   },
   {
    "timestamp": "2026-08-17T16:01:57",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.066,
    "decode_tok_s": 285.6,
    "peak_clock": 2190.0,
    "mean_draw": 220.0,
    "peak_draw": 227.5,
    "peak_temp": 53.0
   },
   {
    "timestamp": "2026-08-17T16:02:02",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.026,
    "decode_tok_s": 303.7,
    "peak_clock": 2190.0,
    "mean_draw": 125.2,
    "peak_draw": 232.0,
    "peak_temp": 53.0
   },
   {
    "timestamp": "2026-08-17T16:02:03",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.066,
    "decode_tok_s": 285.5,
    "peak_clock": 2190.0,
    "mean_draw": 221.7,
    "peak_draw": 236.6,
    "peak_temp": 53.0
   },
   {
    "timestamp": "2026-08-17T16:02:04",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.015,
    "decode_tok_s": 304.3,
    "peak_clock": 2190.0,
    "mean_draw": 233.5,
    "peak_draw": 236.6,
    "peak_temp": 53.0
   },
   {
    "timestamp": "2026-08-17T16:02:06",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.065,
    "decode_tok_s": 277.3,
    "peak_clock": 2190.0,
    "mean_draw": 226.0,
    "peak_draw": 228.6,
    "peak_temp": 54.0
   },
   {
    "timestamp": "2026-08-17T16:02:07",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.015,
    "decode_tok_s": 304.7,
    "peak_clock": 2190.0,
    "mean_draw": 233.2,
    "peak_draw": 234.5,
    "peak_temp": 54.0
   },
   {
    "timestamp": "2026-08-17T16:02:09",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.065,
    "decode_tok_s": 284.2,
    "peak_clock": 2190.0,
    "mean_draw": 229.8,
    "peak_draw": 238.4,
    "peak_temp": 55.0
   },
   {
    "timestamp": "2026-08-17T16:02:14",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.015,
    "decode_tok_s": 317.2,
    "peak_clock": 2640.0,
    "mean_draw": 139.1,
    "peak_draw": 252.1,
    "peak_temp": 56.0
   },
   {
    "timestamp": "2026-08-17T16:02:15",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.066,
    "decode_tok_s": 297.2,
    "peak_clock": 2640.0,
    "mean_draw": 291.2,
    "peak_draw": 299.6,
    "peak_temp": 57.0
   },
   {
    "timestamp": "2026-08-17T16:02:16",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.015,
    "decode_tok_s": 317.7,
    "peak_clock": 2640.0,
    "mean_draw": 307.9,
    "peak_draw": 311.1,
    "peak_temp": 57.0
   },
   {
    "timestamp": "2026-08-17T16:02:18",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.065,
    "decode_tok_s": 290.9,
    "peak_clock": 2640.0,
    "mean_draw": 297.5,
    "peak_draw": 317.1,
    "peak_temp": 58.0
   },
   {
    "timestamp": "2026-08-17T16:02:19",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.016,
    "decode_tok_s": 311.6,
    "peak_clock": 2640.0,
    "mean_draw": 313.9,
    "peak_draw": 317.1,
    "peak_temp": 58.0
   },
   {
    "timestamp": "2026-08-17T16:02:20",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.064,
    "decode_tok_s": 295.5,
    "peak_clock": 2640.0,
    "mean_draw": 305.7,
    "peak_draw": 310.4,
    "peak_temp": 59.0
   },
   {
    "timestamp": "2026-08-17T22:54:57",
    "model": "qwen3.8-27b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 24,
    "ttft_s": 0.082,
    "decode_tok_s": 35.8,
    "peak_clock": 2190.0,
    "mean_draw": 258.3,
    "peak_draw": 265.4,
    "peak_temp": 54.0
   },
   {
    "timestamp": "2026-08-17T22:55:22",
    "model": "qwen3.8-27b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28622,
    "ttft_s": 13.201,
    "decode_tok_s": 34.3,
    "peak_clock": 2190.0,
    "mean_draw": 287.1,
    "peak_draw": 309.9,
    "peak_temp": 60.0
   },
   {
    "timestamp": "2026-08-17T22:55:34",
    "model": "qwen3.8-27b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 24,
    "ttft_s": 0.079,
    "decode_tok_s": 35.8,
    "peak_clock": 2190.0,
    "mean_draw": 283.0,
    "peak_draw": 285.1,
    "peak_temp": 58.0
   },
   {
    "timestamp": "2026-08-17T22:55:46",
    "model": "qwen3.8-27b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28622,
    "ttft_s": 0.339,
    "decode_tok_s": 34.4,
    "peak_clock": 2190.0,
    "mean_draw": 283.0,
    "peak_draw": 285.4,
    "peak_temp": 59.0
   },
   {
    "timestamp": "2026-08-17T22:55:57",
    "model": "qwen3.8-27b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 24,
    "ttft_s": 0.08,
    "decode_tok_s": 35.8,
    "peak_clock": 2190.0,
    "mean_draw": 283.3,
    "peak_draw": 285.2,
    "peak_temp": 59.0
   },
   {
    "timestamp": "2026-08-17T22:56:09",
    "model": "qwen3.8-27b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28622,
    "ttft_s": 0.344,
    "decode_tok_s": 34.4,
    "peak_clock": 2190.0,
    "mean_draw": 284.2,
    "peak_draw": 286.5,
    "peak_temp": 62.0
   },
   {
    "timestamp": "2026-08-17T22:56:25",
    "model": "qwen3.8-27b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 24,
    "ttft_s": 0.079,
    "decode_tok_s": 35.8,
    "peak_clock": 2190.0,
    "mean_draw": 279.2,
    "peak_draw": 285.6,
    "peak_temp": 60.0
   },
   {
    "timestamp": "2026-08-17T22:56:37",
    "model": "qwen3.8-27b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28622,
    "ttft_s": 0.34,
    "decode_tok_s": 34.4,
    "peak_clock": 2190.0,
    "mean_draw": 284.0,
    "peak_draw": 286.9,
    "peak_temp": 60.0
   },
   {
    "timestamp": "2026-08-17T22:56:48",
    "model": "qwen3.8-27b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 24,
    "ttft_s": 0.082,
    "decode_tok_s": 35.8,
    "peak_clock": 2190.0,
    "mean_draw": 284.9,
    "peak_draw": 286.9,
    "peak_temp": 61.0
   },
   {
    "timestamp": "2026-08-17T22:57:00",
    "model": "qwen3.8-27b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28622,
    "ttft_s": 0.34,
    "decode_tok_s": 34.4,
    "peak_clock": 2190.0,
    "mean_draw": 286.5,
    "peak_draw": 288.8,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T22:57:11",
    "model": "qwen3.8-27b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 24,
    "ttft_s": 0.086,
    "decode_tok_s": 35.8,
    "peak_clock": 2190.0,
    "mean_draw": 286.7,
    "peak_draw": 288.8,
    "peak_temp": 61.0
   },
   {
    "timestamp": "2026-08-17T22:57:23",
    "model": "qwen3.8-27b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28622,
    "ttft_s": 0.338,
    "decode_tok_s": 34.4,
    "peak_clock": 2190.0,
    "mean_draw": 286.5,
    "peak_draw": 290.2,
    "peak_temp": 61.0
   },
   {
    "timestamp": "2026-08-17T22:57:39",
    "model": "qwen3.8-27b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 24,
    "ttft_s": 0.079,
    "decode_tok_s": 36.2,
    "peak_clock": 2595.0,
    "mean_draw": 320.7,
    "peak_draw": 331.4,
    "peak_temp": 62.0
   },
   {
    "timestamp": "2026-08-17T22:57:50",
    "model": "qwen3.8-27b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28622,
    "ttft_s": 0.318,
    "decode_tok_s": 34.9,
    "peak_clock": 2535.0,
    "mean_draw": 328.7,
    "peak_draw": 337.6,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T22:58:01",
    "model": "qwen3.8-27b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 24,
    "ttft_s": 0.08,
    "decode_tok_s": 36.2,
    "peak_clock": 2490.0,
    "mean_draw": 329.0,
    "peak_draw": 329.3,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T22:58:13",
    "model": "qwen3.8-27b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28622,
    "ttft_s": 0.32,
    "decode_tok_s": 34.9,
    "peak_clock": 2535.0,
    "mean_draw": 328.3,
    "peak_draw": 335.5,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T22:58:24",
    "model": "qwen3.8-27b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 24,
    "ttft_s": 0.081,
    "decode_tok_s": 36.2,
    "peak_clock": 2505.0,
    "mean_draw": 329.1,
    "peak_draw": 330.0,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T22:58:36",
    "model": "qwen3.8-27b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28622,
    "ttft_s": 0.318,
    "decode_tok_s": 34.8,
    "peak_clock": 2505.0,
    "mean_draw": 328.2,
    "peak_draw": 335.9,
    "peak_temp": 67.0
   },
   {
    "timestamp": "2026-08-17T23:13:47",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 47,
    "ttft_s": 0.064,
    "decode_tok_s": 36.3,
    "peak_clock": 2655.0,
    "mean_draw": 523.4,
    "peak_draw": 539.5,
    "peak_temp": 58.0
   },
   {
    "timestamp": "2026-08-17T23:14:38",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28645,
    "ttft_s": 38.303,
    "decode_tok_s": 33.0,
    "peak_clock": 2640.0,
    "mean_draw": 561.8,
    "peak_draw": 573.8,
    "peak_temp": 61.0
   },
   {
    "timestamp": "2026-08-17T23:14:49",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 47,
    "ttft_s": 0.054,
    "decode_tok_s": 36.3,
    "peak_clock": 2640.0,
    "mean_draw": 565.1,
    "peak_draw": 573.5,
    "peak_temp": 62.0
   },
   {
    "timestamp": "2026-08-17T23:15:01",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28645,
    "ttft_s": 0.099,
    "decode_tok_s": 33.4,
    "peak_clock": 2640.0,
    "mean_draw": 576.2,
    "peak_draw": 579.0,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T23:15:12",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 47,
    "ttft_s": 0.051,
    "decode_tok_s": 36.8,
    "peak_clock": 2640.0,
    "mean_draw": 570.5,
    "peak_draw": 575.8,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T23:15:24",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28645,
    "ttft_s": 0.096,
    "decode_tok_s": 33.5,
    "peak_clock": 2640.0,
    "mean_draw": 574.6,
    "peak_draw": 582.3,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T23:15:39",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 47,
    "ttft_s": 0.049,
    "decode_tok_s": 36.8,
    "peak_clock": 2640.0,
    "mean_draw": 574.0,
    "peak_draw": 583.4,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T23:15:51",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28645,
    "ttft_s": 0.096,
    "decode_tok_s": 33.6,
    "peak_clock": 2640.0,
    "mean_draw": 588.1,
    "peak_draw": 596.6,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T23:16:02",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 47,
    "ttft_s": 0.052,
    "decode_tok_s": 36.8,
    "peak_clock": 2640.0,
    "mean_draw": 582.5,
    "peak_draw": 588.0,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T23:16:14",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28645,
    "ttft_s": 0.096,
    "decode_tok_s": 33.6,
    "peak_clock": 2640.0,
    "mean_draw": 592.3,
    "peak_draw": 600.3,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T23:16:25",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 47,
    "ttft_s": 0.051,
    "decode_tok_s": 36.8,
    "peak_clock": 2640.0,
    "mean_draw": 585.1,
    "peak_draw": 590.0,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T23:16:37",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28645,
    "ttft_s": 0.095,
    "decode_tok_s": 33.6,
    "peak_clock": 2640.0,
    "mean_draw": 594.8,
    "peak_draw": 601.6,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T23:16:53",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 47,
    "ttft_s": 0.049,
    "decode_tok_s": 36.8,
    "peak_clock": 2670.0,
    "mean_draw": 632.4,
    "peak_draw": 657.2,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T23:17:05",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28645,
    "ttft_s": 0.096,
    "decode_tok_s": 33.5,
    "peak_clock": 2670.0,
    "mean_draw": 656.2,
    "peak_draw": 658.2,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T23:17:16",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 47,
    "ttft_s": 0.05,
    "decode_tok_s": 36.8,
    "peak_clock": 2670.0,
    "mean_draw": 654.9,
    "peak_draw": 658.2,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T23:17:28",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28645,
    "ttft_s": 0.095,
    "decode_tok_s": 33.5,
    "peak_clock": 2655.0,
    "mean_draw": 658.2,
    "peak_draw": 662.3,
    "peak_temp": 67.0
   },
   {
    "timestamp": "2026-08-17T23:17:39",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 47,
    "ttft_s": 0.051,
    "decode_tok_s": 36.8,
    "peak_clock": 2670.0,
    "mean_draw": 654.5,
    "peak_draw": 658.0,
    "peak_temp": 67.0
   },
   {
    "timestamp": "2026-08-17T23:17:51",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28645,
    "ttft_s": 0.096,
    "decode_tok_s": 33.5,
    "peak_clock": 2655.0,
    "mean_draw": 658.2,
    "peak_draw": 664.3,
    "peak_temp": 67.0
   }
  ],
  "quarantined_llama_single_gpu_scope": [
   {
    "timestamp": "2026-08-17T16:32:05",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 47,
    "ttft_s": 0.071,
    "decode_tok_s": 29.3,
    "peak_clock": 2655.0,
    "mean_draw": 251.3,
    "peak_draw": 276.6,
    "peak_temp": 56.0
   },
   {
    "timestamp": "2026-08-17T16:32:57",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28645,
    "ttft_s": 40.014,
    "decode_tok_s": 32.5,
    "peak_clock": 2640.0,
    "mean_draw": 321.2,
    "peak_draw": 328.9,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T16:33:09",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 47,
    "ttft_s": 0.058,
    "decode_tok_s": 35.8,
    "peak_clock": 2640.0,
    "mean_draw": 324.0,
    "peak_draw": 326.9,
    "peak_temp": 62.0
   },
   {
    "timestamp": "2026-08-17T16:33:21",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28645,
    "ttft_s": 0.125,
    "decode_tok_s": 32.6,
    "peak_clock": 2640.0,
    "mean_draw": 328.7,
    "peak_draw": 329.3,
    "peak_temp": 62.0
   },
   {
    "timestamp": "2026-08-17T16:33:33",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 47,
    "ttft_s": 0.056,
    "decode_tok_s": 32.3,
    "peak_clock": 2640.0,
    "mean_draw": 312.5,
    "peak_draw": 328.9,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T16:33:47",
    "model": "llama-3.3-70b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28645,
    "ttft_s": 0.098,
    "decode_tok_s": 30.8,
    "peak_clock": 2640.0,
    "mean_draw": 320.5,
    "peak_draw": 329.0,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T16:34:03",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 47,
    "ttft_s": 0.051,
    "decode_tok_s": 33.7,
    "peak_clock": 2640.0,
    "mean_draw": 312.9,
    "peak_draw": 328.9,
    "peak_temp": 63.0
   },
   {
    "timestamp": "2026-08-17T16:34:16",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28645,
    "ttft_s": 0.101,
    "decode_tok_s": 31.7,
    "peak_clock": 2640.0,
    "mean_draw": 320.3,
    "peak_draw": 335.3,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T16:34:27",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 47,
    "ttft_s": 0.056,
    "decode_tok_s": 35.7,
    "peak_clock": 2640.0,
    "mean_draw": 323.7,
    "peak_draw": 330.5,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T16:34:40",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28645,
    "ttft_s": 0.097,
    "decode_tok_s": 32.6,
    "peak_clock": 2640.0,
    "mean_draw": 331.5,
    "peak_draw": 336.5,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T16:34:51",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 47,
    "ttft_s": 0.055,
    "decode_tok_s": 35.7,
    "peak_clock": 2640.0,
    "mean_draw": 329.5,
    "peak_draw": 336.0,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T16:35:03",
    "model": "llama-3.3-70b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28645,
    "ttft_s": 0.097,
    "decode_tok_s": 32.6,
    "peak_clock": 2640.0,
    "mean_draw": 333.1,
    "peak_draw": 336.2,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T16:35:19",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 47,
    "ttft_s": 0.05,
    "decode_tok_s": 35.7,
    "peak_clock": 2640.0,
    "mean_draw": 311.0,
    "peak_draw": 329.0,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T16:35:32",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 28645,
    "ttft_s": 0.097,
    "decode_tok_s": 31.8,
    "peak_clock": 2640.0,
    "mean_draw": 313.9,
    "peak_draw": 329.3,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T16:35:49",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 47,
    "ttft_s": 0.056,
    "decode_tok_s": 26.2,
    "peak_clock": 2640.0,
    "mean_draw": 259.9,
    "peak_draw": 329.2,
    "peak_temp": 66.0
   },
   {
    "timestamp": "2026-08-17T16:36:04",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 28645,
    "ttft_s": 0.455,
    "decode_tok_s": 32.4,
    "peak_clock": 2640.0,
    "mean_draw": 328.5,
    "peak_draw": 329.4,
    "peak_temp": 64.0
   },
   {
    "timestamp": "2026-08-17T16:36:16",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 47,
    "ttft_s": 0.055,
    "decode_tok_s": 35.8,
    "peak_clock": 2640.0,
    "mean_draw": 326.5,
    "peak_draw": 329.2,
    "peak_temp": 65.0
   },
   {
    "timestamp": "2026-08-17T16:36:31",
    "model": "llama-3.3-70b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 28645,
    "ttft_s": 0.097,
    "decode_tok_s": 28.5,
    "peak_clock": 2640.0,
    "mean_draw": 296.6,
    "peak_draw": 329.1,
    "peak_temp": 65.0
   }
  ],
  "lfm_aug16_partial_separate_session": [
   {
    "timestamp": "2026-08-16T16:57:03",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.063,
    "decode_tok_s": 302.1,
    "peak_clock": 2190,
    "mean_draw": 120.4,
    "peak_draw": 204.9,
    "peak_temp": 48
   },
   {
    "timestamp": "2026-08-16T16:57:06",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.991,
    "decode_tok_s": 284.2,
    "peak_clock": 2190,
    "mean_draw": 252,
    "peak_draw": 268.1,
    "peak_temp": 54
   },
   {
    "timestamp": "2026-08-16T16:57:07",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.015,
    "decode_tok_s": 303,
    "peak_clock": 2190,
    "mean_draw": 231.3,
    "peak_draw": 235.9,
    "peak_temp": 50
   },
   {
    "timestamp": "2026-08-16T16:57:08",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.061,
    "decode_tok_s": 288.7,
    "peak_clock": 2190,
    "mean_draw": 226.4,
    "peak_draw": 229.8,
    "peak_temp": 50
   },
   {
    "timestamp": "2026-08-16T16:57:10",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.063,
    "decode_tok_s": 306.5,
    "peak_clock": 2190,
    "mean_draw": 225.7,
    "peak_draw": 234.6,
    "peak_temp": 50
   },
   {
    "timestamp": "2026-08-16T16:57:11",
    "model": "lfm2.5-1.2b",
    "condition": "A",
    "watts": 330,
    "clock": "2200",
    "note": "production: capped + locked",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.063,
    "decode_tok_s": 283.8,
    "peak_clock": 2190,
    "mean_draw": 218.2,
    "peak_draw": 220.6,
    "peak_temp": 51
   },
   {
    "timestamp": "2026-08-16T16:57:16",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.062,
    "decode_tok_s": 307.8,
    "peak_clock": 2190,
    "mean_draw": 117.9,
    "peak_draw": 220.1,
    "peak_temp": 51
   },
   {
    "timestamp": "2026-08-16T16:57:17",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.063,
    "decode_tok_s": 285.1,
    "peak_clock": 2190,
    "mean_draw": 218.5,
    "peak_draw": 234,
    "peak_temp": 52
   },
   {
    "timestamp": "2026-08-16T16:57:19",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.059,
    "decode_tok_s": 304.4,
    "peak_clock": 2190,
    "mean_draw": 226.4,
    "peak_draw": 234,
    "peak_temp": 52
   },
   {
    "timestamp": "2026-08-16T16:57:20",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.061,
    "decode_tok_s": 278.2,
    "peak_clock": 2190,
    "mean_draw": 225.8,
    "peak_draw": 229.9,
    "peak_temp": 53
   },
   {
    "timestamp": "2026-08-16T16:57:22",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.063,
    "decode_tok_s": 304.5,
    "peak_clock": 2190,
    "mean_draw": 226.8,
    "peak_draw": 235,
    "peak_temp": 53
   },
   {
    "timestamp": "2026-08-16T16:57:23",
    "model": "lfm2.5-1.2b",
    "condition": "B",
    "watts": 450,
    "clock": "2200",
    "note": "isolates power limit",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.063,
    "decode_tok_s": 284.5,
    "peak_clock": 2190,
    "mean_draw": 224.6,
    "peak_draw": 227,
    "peak_temp": 53
   },
   {
    "timestamp": "2026-08-16T16:57:28",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 1,
    "prompt_tokens": 21,
    "ttft_s": 0.059,
    "decode_tok_s": 315.6,
    "peak_clock": 2655,
    "mean_draw": 145.5,
    "peak_draw": 290.5,
    "peak_temp": 55
   },
   {
    "timestamp": "2026-08-16T16:57:29",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 1,
    "prompt_tokens": 31921,
    "ttft_s": 0.063,
    "decode_tok_s": 296.1,
    "peak_clock": 2655,
    "mean_draw": 288,
    "peak_draw": 309.4,
    "peak_temp": 56
   },
   {
    "timestamp": "2026-08-16T16:57:31",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 2,
    "prompt_tokens": 21,
    "ttft_s": 0.062,
    "decode_tok_s": 316.4,
    "peak_clock": 2655,
    "mean_draw": 299.1,
    "peak_draw": 309.4,
    "peak_temp": 56
   },
   {
    "timestamp": "2026-08-16T16:57:32",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 2,
    "prompt_tokens": 31921,
    "ttft_s": 0.062,
    "decode_tok_s": 290.3,
    "peak_clock": 2655,
    "mean_draw": 295.1,
    "peak_draw": 305.4,
    "peak_temp": 57
   },
   {
    "timestamp": "2026-08-16T16:57:33",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "short",
    "repeat": 3,
    "prompt_tokens": 21,
    "ttft_s": 0.065,
    "decode_tok_s": 311.5,
    "peak_clock": 2655,
    "mean_draw": 299.6,
    "peak_draw": 303.9,
    "peak_temp": 57
   },
   {
    "timestamp": "2026-08-16T16:57:35",
    "model": "lfm2.5-1.2b",
    "condition": "C",
    "watts": 330,
    "clock": "off",
    "note": "isolates clock lock",
    "ctx": "long",
    "repeat": 3,
    "prompt_tokens": 31921,
    "ttft_s": 0.07,
    "decode_tok_s": 292.9,
    "peak_clock": 2655,
    "mean_draw": 299.3,
    "peak_draw": 314.6,
    "peak_temp": 57
   }
  ],
  "caveats": [
   "lfm rows include a pass taken during 8-way download disk saturation (decode 30-99 vs 277-317 clean) - filter decode<150 to exclude, or keep as the contention exhibit",
   "llama 'locked' conditions recorded peak 2640-2655MHz: -lgc NOT verified holding on GPU 0 under load (command succeeds; enforcement unproven; LOCKCHECK rerun pending). Draw deltas remain valid (A ~570W vs C ~655W at identical decode)",
   "quarantined file: conditions applied and sampled on GPU 1 only while the model spanned both cards"
  ]
 },
 "topology": [
  {
   "ts": "2026-08-17T16:51:54",
   "model": "qwen3.8-27b",
   "topology": "TP1",
   "tp": 1,
   "pp": 1,
   "gpus": "1",
   "ctx": "short",
   "repeat": 1,
   "load_s": 876,
   "prompt_tokens": 24,
   "ttft_s": 0.092,
   "decode_tok_s": 35.8
  },
  {
   "ts": "2026-08-17T16:52:05",
   "model": "qwen3.8-27b",
   "topology": "TP1",
   "tp": 1,
   "pp": 1,
   "gpus": "1",
   "ctx": "short",
   "repeat": 2,
   "load_s": 876,
   "prompt_tokens": 24,
   "ttft_s": 0.104,
   "decode_tok_s": 35.8
  },
  {
   "ts": "2026-08-17T16:52:16",
   "model": "qwen3.8-27b",
   "topology": "TP1",
   "tp": 1,
   "pp": 1,
   "gpus": "1",
   "ctx": "short",
   "repeat": 3,
   "load_s": 876,
   "prompt_tokens": 24,
   "ttft_s": 0.095,
   "decode_tok_s": 35.8
  },
  {
   "ts": "2026-08-17T16:52:41",
   "model": "qwen3.8-27b",
   "topology": "TP1",
   "tp": 1,
   "pp": 1,
   "gpus": "1",
   "ctx": "long",
   "repeat": 1,
   "load_s": 876,
   "prompt_tokens": 28622,
   "ttft_s": 13.312,
   "decode_tok_s": 34.3
  },
  {
   "ts": "2026-08-17T16:52:53",
   "model": "qwen3.8-27b",
   "topology": "TP1",
   "tp": 1,
   "pp": 1,
   "gpus": "1",
   "ctx": "long",
   "repeat": 2,
   "load_s": 876,
   "prompt_tokens": 28622,
   "ttft_s": 0.364,
   "decode_tok_s": 34.4
  },
  {
   "ts": "2026-08-17T16:53:05",
   "model": "qwen3.8-27b",
   "topology": "TP1",
   "tp": 1,
   "pp": 1,
   "gpus": "1",
   "ctx": "long",
   "repeat": 3,
   "load_s": 876,
   "prompt_tokens": 28622,
   "ttft_s": 0.367,
   "decode_tok_s": 34.4
  },
  {
   "ts": "2026-08-17T17:02:38",
   "model": "qwen3.8-27b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 1,
   "load_s": 552,
   "prompt_tokens": 24,
   "ttft_s": 0.076,
   "decode_tok_s": 50.7
  },
  {
   "ts": "2026-08-17T17:02:46",
   "model": "qwen3.8-27b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 2,
   "load_s": 552,
   "prompt_tokens": 24,
   "ttft_s": 0.084,
   "decode_tok_s": 50.7
  },
  {
   "ts": "2026-08-17T17:02:54",
   "model": "qwen3.8-27b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 3,
   "load_s": 552,
   "prompt_tokens": 24,
   "ttft_s": 0.079,
   "decode_tok_s": 50.7
  },
  {
   "ts": "2026-08-17T17:03:19",
   "model": "qwen3.8-27b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "long",
   "repeat": 1,
   "load_s": 552,
   "prompt_tokens": 28622,
   "ttft_s": 16.771,
   "decode_tok_s": 48.9
  },
  {
   "ts": "2026-08-17T17:03:27",
   "model": "qwen3.8-27b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "long",
   "repeat": 2,
   "load_s": 552,
   "prompt_tokens": 28622,
   "ttft_s": 0.339,
   "decode_tok_s": 48.8
  },
  {
   "ts": "2026-08-17T17:03:36",
   "model": "qwen3.8-27b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "long",
   "repeat": 3,
   "load_s": 552,
   "prompt_tokens": 28622,
   "ttft_s": 0.344,
   "decode_tok_s": 49.1
  },
  {
   "ts": "2026-08-17T17:13:52",
   "model": "qwen3.8-27b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 1,
   "load_s": 589,
   "prompt_tokens": 24,
   "ttft_s": 0.079,
   "decode_tok_s": 34.2
  },
  {
   "ts": "2026-08-17T17:14:04",
   "model": "qwen3.8-27b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 2,
   "load_s": 589,
   "prompt_tokens": 24,
   "ttft_s": 0.075,
   "decode_tok_s": 34.2
  },
  {
   "ts": "2026-08-17T17:14:15",
   "model": "qwen3.8-27b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 3,
   "load_s": 589,
   "prompt_tokens": 24,
   "ttft_s": 0.07,
   "decode_tok_s": 34.2
  },
  {
   "ts": "2026-08-17T17:14:37",
   "model": "qwen3.8-27b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "long",
   "repeat": 1,
   "load_s": 589,
   "prompt_tokens": 28622,
   "ttft_s": 8.933,
   "decode_tok_s": 32.9
  },
  {
   "ts": "2026-08-17T17:14:49",
   "model": "qwen3.8-27b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "long",
   "repeat": 2,
   "load_s": 589,
   "prompt_tokens": 28622,
   "ttft_s": 0.356,
   "decode_tok_s": 32.9
  },
  {
   "ts": "2026-08-17T17:15:01",
   "model": "qwen3.8-27b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "long",
   "repeat": 3,
   "load_s": 589,
   "prompt_tokens": 28622,
   "ttft_s": 0.337,
   "decode_tok_s": 32.9
  },
  {
   "ts": "2026-08-17T17:25:50",
   "model": "llama-3.3-70b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 1,
   "load_s": 622,
   "prompt_tokens": 47,
   "ttft_s": 0.077,
   "decode_tok_s": 35.3
  },
  {
   "ts": "2026-08-17T17:26:02",
   "model": "llama-3.3-70b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 2,
   "load_s": 622,
   "prompt_tokens": 47,
   "ttft_s": 0.072,
   "decode_tok_s": 35.3
  },
  {
   "ts": "2026-08-17T17:26:13",
   "model": "llama-3.3-70b",
   "topology": "TP2",
   "tp": 2,
   "pp": 1,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 3,
   "load_s": 622,
   "prompt_tokens": 47,
   "ttft_s": 0.051,
   "decode_tok_s": 35.3
  },
  {
   "ts": "2026-08-17T17:36:59",
   "model": "llama-3.3-70b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 1,
   "load_s": 605,
   "prompt_tokens": 47,
   "ttft_s": 0.058,
   "decode_tok_s": 21.9
  },
  {
   "ts": "2026-08-17T17:37:17",
   "model": "llama-3.3-70b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 2,
   "load_s": 605,
   "prompt_tokens": 47,
   "ttft_s": 0.056,
   "decode_tok_s": 21.8
  },
  {
   "ts": "2026-08-17T17:37:35",
   "model": "llama-3.3-70b",
   "topology": "PP2",
   "tp": 1,
   "pp": 2,
   "gpus": "0,1",
   "ctx": "short",
   "repeat": 3,
   "load_s": 605,
   "prompt_tokens": 47,
   "ttft_s": 0.072,
   "decode_tok_s": 21.8
  }
 ],
 "experiment_9p": [
  {
   "repeat": 1,
   "source": "bind9p",
   "load_s": 763,
   "model": "Qwen3.8-27B-AWQ",
   "engine": "vllm/vllm-openai@sha256:ffb2d59b1c059a5bd8d781320c9f5189de8293693b7d95da54befddaa54abf52",
   "copy_s": 175
  },
  {
   "repeat": 1,
   "source": "volume",
   "load_s": 203,
   "model": "Qwen3.8-27B-AWQ",
   "engine": "vllm/vllm-openai@sha256:ffb2d59b1c059a5bd8d781320c9f5189de8293693b7d95da54befddaa54abf52",
   "copy_s": 175
  },
  {
   "repeat": 2,
   "source": "bind9p",
   "load_s": 762,
   "model": "Qwen3.8-27B-AWQ",
   "engine": "vllm/vllm-openai@sha256:ffb2d59b1c059a5bd8d781320c9f5189de8293693b7d95da54befddaa54abf52",
   "copy_s": 175
  },
  {
   "repeat": 2,
   "source": "volume",
   "load_s": 204,
   "model": "Qwen3.8-27B-AWQ",
   "engine": "vllm/vllm-openai@sha256:ffb2d59b1c059a5bd8d781320c9f5189de8293693b7d95da54befddaa54abf52",
   "copy_s": 175
  }
 ],
 "sweep_label_note": "pre-2026-08-18 rounds requested ~1.9x their label (chars/token guess); measured_prompt_tokens is truth",
 "topology_headline": {
  "model": "qwen3.8-27b (fits one card)",
  "long_ctx_single_stream_tok_s": {
   "TP1": 34.4,
   "TP2": 48.9,
   "PP2": 32.9
  },
  "tp2_vs_tp1_pct": 42.2,
  "pp2_vs_tp2_pct": -32.7,
  "load_min": {
   "TP1": 14.6,
   "TP2": 9.2
  },
  "scope": "one model, one context, single stream - verify under concurrency before production moves"
 },
 "experiment_9p_headline": {
  "bind_9p_s": [
   763,
   762
  ],
  "volume_ext4_s": [
   203,
   204
  ],
  "speedup": 3.75,
  "one_off_copy_s": 175,
  "refutes": "2026-08-12 dd-based conclusion (530MB/s sequential) that migration buys nothing - dd measured a workload vLLM never runs"
 },
 "downloads": {
  "authenticated_vs_not_mb_min": {
   "unauthenticated_8_streams": 223,
   "authenticated": 1090
  },
  "concurrency_mb_min": {
   "8_streams": 230,
   "6_streams": 692
  },
  "conclusion": "HF throttles per account; token=5x, fewer streams=3x. Queue downloads, never fan out.",
  "silent_death": "snapshot_download dies without error mid-transfer (observed 7+ times); restart loops required; a repo can ship the weights TWICE (gpt-oss root + original/ = 121.6GB on disk for a 65.8GB servable model)"
 },
 "disk_contention": {
  "shard_load_s_quiet": 104,
  "shard_load_s_during_download": 669,
  "factor": 6.7
 },
 "tokenizer_divergence": {
  "same_text_tokens": {
   "qwen": 29715,
   "lfm2.5": 33139
  },
  "pct": 11.5,
  "filler_tokens_per_model": {
   "qwen3.8": 28,
   "glm-4.7": 27
  },
  "note": "from /tokenize calibration; the 14-chars/token guess was ~1.9x off"
 },
 "quant_capability_probe": {
  "date": "2026-08-17",
  "method": "vllm/quant-support.py asks the pinned engine",
  "gpu_capability": 89,
  "min_capability": {
   "mxfp4": 80,
   "gpt_oss_mxfp4": 80,
   "modelopt_fp4": 75,
   "nvfp4_per_token": 75,
   "fp8": 75,
   "fp8_per_block": 75,
   "awq": 75,
   "gptq_marlin": 60
  },
  "caveat": "min_capability proves a PERMITTED code path, not a fast one"
 },
 "mtp_finding": {
  "verdict": "NEUTRAL on vLLM 0.26.0 (this rig): 47.4/46.9/45.9 tok/s off vs 51.1/44.6/47.8 on, at ctx ~0/~4k/~15k - within the noise floor",
  "history": "the same preset measured ~+1.9x on the 0.25.x engine (repo bench history) - an engine-version regression to neutral, NOT a reversal",
  "acceptance": {
   "drafts": 504,
   "draft_tokens": 1008,
   "accepted": 696,
   "rate": 0.69,
   "per_pos": [
    0.788,
    0.593
   ],
   "note": "healthy drafter; the cost of the drafter forwards eats the gain"
  },
  "spec_config_verified_engaged": "Qwen3_5MTP resolved, drafter loaded, config echoed",
  "deadlock_hazard_still_stands": "long cold prefill deadlock on 0.26.0, /health stays 200",
  "measurement_forensics": [
   "first sweep reported 26.1 tok/s (-44%): the counter counted SSE CHUNKS, and spec decode bundles ~2.4 tokens/chunk (measured: 400 tokens in 165-171 chunks; non-spec runs are exactly 1/chunk, which is why every other row was unaffected)",
   "a quick wall-clock curl A/B then suggested +26%: overcorrection - total time with n=2 including prefill and warmup variance. The token-true streaming measure across three contexts is the instrument of record"
  ]
 },
 "doctrines_overturned": [
  {
   "claim": "Never tensor-parallelise a model that fits one card (x4 link taxes all-reduce)",
   "status": "REFUTED 2026-08-17",
   "evidence": "topology-study: TP2 +42.2% over TP1 on qwen3.8; PP2 -32.7% vs TP2"
  },
  {
   "claim": "Ada has no kernels for block-FP8 / NVFP4 / MXFP4",
   "status": "REFUTED 2026-08-17",
   "evidence": "engine capability probe: every format permitted at capability 89 (quant-support.py)"
  },
  {
   "claim": "MXFP4 upconverts to bf16 on Ada (a 21B wants 42GB)",
   "status": "REFUTED 2026-08-18",
   "evidence": "gpt-oss-120b (117B, native MXFP4) serves at ctx=131072 in 92.7GB total; bf16 weights would be ~234GB - existence proof. Then measured at 153 tok/s"
  },
  {
   "claim": "9P bind mount is irrelevant to load time (dd reads 530MB/s)",
   "status": "REFUTED 2026-08-18",
   "evidence": "controlled A/B, mount the only variable: 3.75x (experiment-9p.py)"
  }
 ],
 "exclusions": [
  {
   "model": "Kimi-Linear-48B-AWQ",
   "reason": "tokenizer requires executing shipped Python (no tokenizer.json fallback); architecture IS vLLM-native; official upstream packaged identically - policy exclusion"
  },
  {
   "model": "Muse-Glimmer-30B-FP8-block",
   "reason": "architecture absent from pinned 0.26.0 registry (block-FP8 wrongly blamed at first)"
  },
  {
   "model": "gpt-oss-120b-AWQ (twhitworth)",
   "reason": "faulty community conversion - MoE experts never quantised (HF metadata: F16 100%); OOMs via unquantized fused-MoE path. Replaced by official MXFP4"
  }
 ],
 "instrument_failures": [
  "error capture kept the HEAD of a log tail (banner), then the TAIL (vLLM's 'see root cause above' wrapper) - cause lines must be scanned, deduped, from the whole log",
  "a fix does not reach a running process - stale imports recorded useless errors for hours",
  "the launcher shell is not the job - 6+ completed/failed notifications described a shell that exited immediately while the work ran on",
  "PS 5.1 vs 7 JSON escaping is INVERTED; no literal works in both; 5.1 silently dropped a study arm (build args from PSVersionTable at runtime)",
  "production vLLM releases VRAM when idle and reloads on request (684MiB -> 45GB in 18s) - preflight is necessary; stopping production for both-card work is sufficient",
  "a GPU sampler indexed characters not fields (clock 52 / draw 94 on a 210MHz card) - parse lines into objects first",
  "a study applied conditions and sampled one GPU of a two-GPU model - scope conditions AND sampling to every card the model occupies",
  "harness default flags appended AFTER per-model flags override them (last-wins) - the measured flag lost to the plumbing",
  "the sweep counted SSE chunks as tokens - correct for normal decoding (1/chunk), silently ~2.4x wrong under speculative decoding, reporting the accelerator as a brake (26.1 vs a true ~47). count from usage.completion_tokens, never from chunks",
  "the verdict tool read the 0.92-utilisation allocation as evidence of upconversion - vLLM fills 92% of each card BY DESIGN",
  "dd was perfectly accurate about a workload vLLM never runs - validate instruments against the real access pattern",
  "clocks.max.graphics reports the hardware ceiling locked or not; only clocks under load verify a lock; -lgc SUCCEEDS on GPU 0 but enforcement under load is unproven"
 ],
 "open_items": [
  "GLM-4.7-Flash returns 'no tokens generated' at ~16k ctx (soft quirk, undiagnosed)",
  "GPU 0 clock-lock enforcement under load unproven (LOCKCHECK rerun pending: elevation + load)",
  "pre-08-18 sweep rounds label ~1.9x under; measured_prompt_tokens carried per row",
  "Nemotron-3.5-Lightning at 158 tok/s = 3.4x production daily driver - delegation gauntlet pending",
  "MTP neutral on 0.26.0 vs +1.9x on 0.25.x: engine regression worth a changelog/issue search before P1 quotes the old number",
  "TP=2 production question pending a concurrency test",
  "harness mounts not yet switched to the migrated vllm_models volume"
 ]
}