{
 "title": "Qwen3.8-27B on one Spark: TensorFold vs vLLM vs SGLang",
 "recorded": "2026-09-30 to 2026-10-02",
 "hardware": "1x DGX Spark (spark-78f1)",
 "model": "nvidia/Qwen3.8-27B-NVFP4",
 "drafter": "z-lab/Qwen3.8-27B-DFlash2, 7 speculative tokens",
 "engines": {
  "tensorfold-0.3.6.3": "default (one request at a time on CUDA)",
  "tensorfold-0.6.1": "17c73e1, --parallel 16",
  "vllm": "vllm/vllm-openai:v0.30.0",
  "sglang": "lmsysorg/sglang:v0.5.21-cu130"
 },
 "speed": {
  "TensorFold 0.3.6.3 (default)": {
   "1": [
    83.2,
    0.495
   ],
   "4": [
    61.2,
    18.062
   ],
   "8": [
    61.5,
    51.123
   ],
   "16": [
    61.5,
    116.816
   ],
   "16384": [
    1754.6,
    32.72
   ],
   "65536": [
    1247.7,
    18.66
   ],
   "131072": [
    881.9,
    11.07
   ],
   "rep": 0.4496
  },
  "TensorFold 0.6.1 --parallel 16": {
   "1": [
    98.5,
    0.099
   ],
   "4": [
    142.9,
    0.177
   ],
   "8": [
    244.3,
    0.206
   ],
   "16": [
    372.5,
    0.304
   ],
   "16384": [
    2709.4,
    34.09
   ],
   "65536": [
    2044.0,
    29.99
   ],
   "131072": [
    1514.6,
    21.19
   ],
   "rep": 0.4038
  },
  "vLLM 0.30.0": {
   "1": [
    57.4,
    0.256
   ],
   "4": [
    103.2,
    0.4
   ],
   "8": [
    162.4,
    0.426
   ],
   "16": [
    252.7,
    0.554
   ],
   "16384": [
    2494.1,
    20.62
   ],
   "65536": [
    1918.7,
    17.58
   ],
   "131072": [
    1456.2,
    13.22
   ],
   "rep": 1.289
  },
  "SGLang 0.5.21": {
   "1": [
    60.4,
    0.241
   ],
   "4": [
    94.6,
    0.496
   ],
   "8": [
    149.6,
    0.375
   ],
   "16": [
    223.9,
    0.634
   ],
   "16384": [
    1643.9,
    24.43
   ],
   "65536": [
    1256.2,
    16.8
   ],
   "131072": [
    1197.1,
    16.59
   ],
   "rep": 0.5869
  }
 },
 "exactness": {
  "tf061": {
   "names": [
    "numbers",
    "code",
    "json",
    "prose",
    "explain",
    "sql",
    "email",
    "math"
   ],
   "rr_on": 8,
   "cc_on": 8,
   "cc_div_on": [],
   "rr_off": 8,
   "cc_off": 8,
   "cc_div_off": [],
   "draft_vs_plain": 8,
   "draft_div": {
    "numbers": null,
    "code": null,
    "json": null,
    "prose": null,
    "explain": null,
    "sql": null,
    "email": null,
    "math": null
   }
  },
  "vllm": {
   "names": [
    "numbers",
    "code",
    "json",
    "prose",
    "explain",
    "sql",
    "email",
    "math"
   ],
   "rr_on": 8,
   "cc_on": 1,
   "cc_div_on": [
    0,
    0,
    4,
    16,
    34,
    34,
    74,
    98,
    117,
    191,
    234,
    235,
    413,
    1053
   ],
   "rr_off": 8,
   "cc_off": 1,
   "cc_div_off": [
    10,
    10,
    16,
    17,
    46,
    94,
    99,
    126,
    126,
    169,
    171,
    344,
    377,
    793
   ],
   "draft_vs_plain": 1,
   "draft_div": {
    "numbers": null,
    "code": 10,
    "json": 344,
    "prose": 17,
    "explain": 16,
    "sql": 696,
    "email": 336,
    "math": 0
   }
  },
  "sgl": {
   "names": [
    "numbers",
    "code",
    "json",
    "prose",
    "explain",
    "sql",
    "email",
    "math"
   ],
   "rr_on": 7,
   "cc_on": 1,
   "cc_div_on": [
    0,
    0,
    0,
    0,
    3,
    3,
    61,
    61,
    117,
    117,
    134,
    169,
    291,
    536
   ],
   "rr_off": 7,
   "cc_off": 1,
   "cc_div_off": [
    0,
    0,
    0,
    0,
    72,
    72,
    74,
    74,
    114,
    114,
    155,
    155,
    191,
    191
   ],
   "draft_vs_plain": 1,
   "draft_div": {
    "numbers": null,
    "code": 72,
    "json": 536,
    "prose": 17,
    "explain": 150,
    "sql": 0,
    "email": 61,
    "math": 3
   }
  }
 },
 "correction": "The first post compared vLLM/SGLang against TensorFold 0.3.6.3; 0.6.1 with --parallel was rerun on the same box and the correction posted."
}