{
 "format": "neural-download-model-family-v1",
 "id": "qwen-9b",
 "primary_packet_id": "qwen35-9b-fp8-b70",
 "name": "Qwen 9B",
 "display_name": "Qwen3.5 9B",
 "publisher": "Qwen / Alibaba",
 "updated_at": "2026-09-11",
 "summary": "Alibaba's dense-hybrid Qwen3.5 9B (linear-attention layers with a full-attention layer every fourth, and a shipped MTP head), served on one Arc Pro B70 as RedHatAI's FP8-dynamic checkpoint through the lab's deterministic vLLM XPU stack with lossless speculative decoding. Two routes are measured: RedHatAI's FP8-dynamic and its W4A16 INT4, the latter faster and byte-exact at every concurrency through 64 users.",
 "architecture": {
  "class": "Qwen3_5ForConditionalGeneration",
  "model_type": "qwen3_5_text",
  "layers": 32,
  "hidden_size": 4096,
  "vocab_size": 248320,
  "full_attention_interval": 4,
  "mtp_hidden_layers": 1,
  "evidence": "repro/qwen35-9b-fp8-b70/README.md"
 },
 "dimensions": {
  "weight_revision": [
   "qwen3.5-9b-fp8-dynamic-790f0576",
   "qwen3.5-9b-w4a16-a398088c"
  ],
  "weight_quantization": [
   "FP8-dynamic",
   "W4A16"
  ],
  "runtime": [
   "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "vLLM XPU 0.27.2rc1.dev77 (R293 image 40d46730 = R276 521eb277 + class-consistent FP16 linears behind CLASSPAD)"
  ],
  "tp": [
   1,
   2
  ],
  "mtp": [
   0,
   3,
   4,
   5,
   6
  ],
  "configured_max_context_tokens": [
   1024,
   8192
  ],
  "kv": [
   "f16"
  ]
 },
 "weight_revisions": [
  {
   "id": "qwen3.5-9b-fp8-dynamic-790f0576",
   "label": "Qwen3.5 9B FP8-dynamic (RedHatAI)",
   "role": "measured FP8 weights with the publisher MTP head",
   "repository": "RedHatAI/Qwen3.5-9B-FP8-dynamic",
   "revision": "790f0576d2d77dd5322aa0603a470bd9e3a3d1f6",
   "model_manifest": "repro/qwen35-9b-fp8-b70/manifests/model-direct-redhatai-qwen35-9b-fp8-dynamic-790f0576.json"
  },
  {
   "id": "qwen3.5-9b-w4a16-a398088c",
   "label": "Qwen3.5 9B W4A16 (RedHatAI)",
   "role": "measured INT4 weights with the publisher MTP head",
   "repository": "RedHatAI/Qwen3.5-9B-quantized.w4a16",
   "revision": "a398088c4228b0ae0c8c78df88fd1e4bf445f068",
   "model_manifest": "repro/qwen35-9b-w4a16-b70/manifests/model-direct-redhatai-qwen35-9b-w4a16-a398088c.json"
  }
 ],
 "model_variants": [],
 "transfer_scope": {
  "status": "one measured FP8-dynamic revision on one and two cards, depths 0-6 strict, 2K-32K ladders on both card counts",
  "transfers": [
   "the Qwen3.5-family vLLM XPU registration, the lab GDN kernels and strict launchers, the draft-only INT4 lm_head"
  ],
  "does_not_transfer": [
   "performance to other Qwen3.5 sizes or quantizations without their own gates"
  ],
  "evidence": "repro/qwen35-9b-fp8-b70/README.md"
 },
 "model_signals": {
  "b70_fit": {
   "band": "one-card measured",
   "scope": "FP8-dynamic, 14.0 GB weights, one B70",
   "basis": "Strict pairs and identity ladders completed on one B70 on 2026-09-07.",
   "reviewed_at": "2026-09-07"
  },
  "quality_evidence": {
   "band": "deployment-identity evidence",
   "scope": "12-prompt strict suite with complete token-id identity gates against a same-configuration no-speculation oracle; canaries; no broad model-quality evaluation",
   "evidence": [
    "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
   ]
  },
  "popularity": {
   "state": "not-scored",
   "reason": "No dated popularity snapshot is stored."
  }
 },
 "run_measurements": [
  {
   "id": "q35-9b-fp8-tp1-mtp0-graph-c1",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 0,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "n/a"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted MTP0 oracle (c1)",
   "quality_scope": "G1/G2/G3 12/12 complete token arrays; canaries on every server; cached_tokens=0",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     50.183,
     50.15
    ],
    "ttft_ms": [
     54.33,
     53.71
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp1-mtp3-graph-fp8head-c1",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 3,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "FP8"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "superseded by the draft INT4 head (c2)",
   "quality_scope": "G1/G2/G3 12/12 complete token arrays; canaries on every server; cached_tokens=0",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     76.917,
     76.879
    ],
    "ttft_ms": [
     88.38,
     88.05
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp1-mtp3-graph-int4head-c2",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 3,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted headline (c2)",
   "quality_scope": "G1/G2/G3 12/12 complete token arrays; canaries on every server; cached_tokens=0",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     98.251,
     98.027
    ],
    "ttft_ms": [
     73.89,
     76.46
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp1-mtp4-graph-int4head-c3",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 4,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "withheld: repeat-exact (G2 12/12) but 8/12 vs the MTP0 oracle; no faster than depth 3",
   "quality_scope": "G1 12/12, G2 12/12, G3 8/12 on both servers",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     98.418,
     98.55
    ],
    "ttft_ms": [
     80.74,
     81.01
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp1-mtp3-graphoff-int4head-c4",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 3,
    "graph": "off",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted graph-off row (c4)",
   "quality_scope": "G1/G2/G3 12/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     96.779,
     96.748
    ],
    "ttft_ms": [
     76.12,
     77.32
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp2-mtp0-graph-c5",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 2,
    "mtp": 0,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "n/a"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted two-card MTP0 oracle (c5)",
   "quality_scope": "G1 12/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     79.456,
     79.4
    ],
    "ttft_ms": [
     41.73,
     41.29
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp2-mtp3-graph-int4head-c5",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 2,
    "mtp": 3,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted two-card headline (c5); LocalMaxxing cmtqzhlvn00c0pa0109xxfb2f",
   "quality_scope": "G1/G2/G3 12/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     147.712,
     147.89
    ],
    "ttft_ms": [
     49.65,
     50.05
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp1-mtp5-graph-int4head-c9",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 5,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "withheld: repeat-exact but 8/12 vs the MTP0 oracle; slower than depth 3",
   "quality_scope": "G1 12/12, G2 12/12, G3 8/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     91.301,
     91.294
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp1-mtp6-graph-int4head-c10",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 6,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "withheld: repeat-exact but 8/12 vs the MTP0 oracle; slower than depth 3",
   "quality_scope": "G1 12/12, G2 12/12, G3 8/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     88.602,
     88.677
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-fp8-tp2-mtp4-graph-int4head-c6",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "variant": "FP8-dynamic",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 2,
    "mtp": 4,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "withheld: repeat-exact but 9/12 vs the MTP0 oracle",
   "quality_scope": "G1 12/12, G2 12/12, G3 9/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     150.694,
     151.737
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json"
  },
  {
   "id": "q35-9b-w4a16-tp1-mtp0-graph-w1",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-w4a16-a398088c",
   "variant": "W4A16",
   "quantization": "W4A16",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 0,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "n/a"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted W4A16 MTP0 oracle (w1); ladder exact at every rung through 64 users",
   "quality_scope": "G1 12/12; identity ladder 64/64 at every rung, both passes",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     64.332,
     64.338
    ],
    "ttft_ms": [
     48.19,
     47.02
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-w4a16-matrix-result.json"
  },
  {
   "id": "q35-9b-w4a16-tp1-mtp3-graph-int4head-w1",
   "state": "lab-measured",
   "revision": "qwen3.5-9b-w4a16-a398088c",
   "variant": "W4A16",
   "quantization": "W4A16",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "config": {
    "tp": 1,
    "mtp": 3,
    "graph": "on",
    "kv": "f16",
    "configured_max_context_tokens": 1024,
    "draft_head": "draft INT4"
   },
   "profile_id": "qwen35-9b-fp8-strict-completions-v1",
   "measurement_class": "strict fresh-server pair",
   "promotion_status": "promoted W4A16 headline (w1); LocalMaxxing cmtrhoyl1000cps01o43bhl72",
   "quality_scope": "G1/G2/G3 12/12",
   "workload": "strict 12-prompt six-class suite over the completions API, 512-token cap, cache zero, class-balanced median tok/s over tokens 1-100 after TTFT, two fresh servers",
   "metrics": {
    "decode_tok_s": [
     113.627,
     112.904
    ],
    "ttft_ms": [
     67.65,
     67.61
    ]
   },
   "evidence": "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-w4a16-matrix-result.json"
  }
 ],
 "series_measurements": [],
 "estimates": [],
 "packets": [
  {
   "id": "qwen35-9b-fp8-b70",
   "label": "Qwen3.5 9B FP8-dynamic \u00b7 one B70 \u00b7 MTP depth 3 with the draft INT4 head",
   "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
   "quantization": "FP8-dynamic",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
   "status": "candidate",
   "evidence_level": "B70-measured strict pairs and ladders",
   "coverage": [
    "decode",
    "TTFT",
    "TP1",
    "TP2",
    "MTP0/3",
    "graph on/off",
    "identity ladders c1-c64",
    "cache-zero strict suite"
   ],
   "grades": {
    "evidence": {
     "grade": "B",
     "scope": "one-card strict pairs with identity gates; matrix in progress",
     "basis": "two fresh-server pairs per configuration, G1-G3 exact, ladders; no clean-host replay yet",
     "reviewed_at": "2026-09-07",
     "evidence": [
      "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-fp8-matrix-result.json",
      "repro/qwen35-9b-fp8-b70/README.md"
     ]
    }
   },
   "projection": {
    "model": "qwen3.5_9b",
    "quant": "fp8",
    "runtime": "vllm",
    "spec": "mtp:3",
    "prompt_tokens": 128,
    "output_tokens": 100
   },
   "manifest": "packages/qwen35-9b-fp8-b70/package.json",
   "topologies": [
    1,
    2
   ]
  },
  {
   "id": "qwen35-9b-w4a16-b70",
   "label": "Qwen3.5 9B W4A16 \u00b7 one or two B70 \u00b7 MTP depth 3 with the draft INT4 head",
   "revision": "qwen3.5-9b-w4a16-a398088c",
   "quantization": "W4A16",
   "runtime": "vLLM XPU 0.27.2rc1.dev77 (R293 image 40d46730 = R276 521eb277 + class-consistent FP16 linears behind CLASSPAD)",
   "topologies": [
    1,
    2
   ],
   "status": "candidate",
   "evidence_level": "B70-measured strict pairs and ladders",
   "coverage": [
    "decode",
    "TTFT",
    "TP1",
    "TP2",
    "MTP0/3",
    "identity ladders c1-c64 exact without speculation",
    "cache-zero strict suite",
    "stagger-exact concurrency",
    "CLASSPAD many-user mode"
   ],
   "grades": {
    "evidence": {
     "grade": "B",
     "scope": "one- and two-card strict pairs with identity gates and full-range ladders",
     "basis": "two fresh-server pairs per card count, G1-G3 exact on both; every no-speculation ladder rung exact through 64 users in both passes on one card, while two cards hold 32 users exactly and drop to 63/64 at 64 users; no clean-host replay yet. 2026-09-11 (R293, CLASSPAD=1): strict gates 12/12 on both card counts at 112/62 (one card) and 165/94 (two cards) tok/s; no-speculation exact to c24 and at c64 on one card (1644 tok/s, 1955 at c128 with the larger admission budget, 512/512), to c32 on two (3227 at c128, 512/512); with a 5 ms admission stagger 64 users on two cards are byte-identical over twenty passes at 2557 tok/s.",
     "reviewed_at": "2026-09-11",
     "evidence": [
      "experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-w4a16-matrix-result.json",
      "repro/qwen35-9b-w4a16-b70/README.md",
      "experiments/qwen35-9b-b70/data/qwen35-9b-w4a16-tp2-mtp3-graph1-dhint4-20260907-w3-strict-result.json",
      "experiments/qwen35-9b-b70/data/2026-09-11-qwen35-9b-r293-classpad.json",
      "experiments/qwen35-9b-b70/notes/2026-09-11-r293-on-the-9b.md"
     ]
    }
   },
   "projection": {
    "model": "qwen3.5_9b",
    "quant": "int4",
    "runtime": "vllm",
    "spec": "mtp:3",
    "prompt_tokens": 128,
    "output_tokens": 100
   },
   "manifest": "packages/qwen35-9b-w4a16-b70/package.json"
  }
 ],
 "views": [
  {
   "id": "qwen-9b-mtp-depth",
   "title": "Speculative depth on one B70",
   "subtitle": "FP8-dynamic \u00b7 strict completions suite \u00b7 graph capture on \u00b7 MTP0 oracle vs depth 3 with the FP8 draft head vs the draft INT4 head \u00b7 depth 4 with the INT4 head is withheld (8/12 vs the oracle)",
   "x_label": "MTP depth",
   "discrete": true,
   "metrics": [
    "decode_tok_s",
    "ttft_ms"
   ],
   "series": [
    {
     "label": "FP8 draft head",
     "measurement_ids": [
      "q35-9b-fp8-tp1-mtp0-graph-c1",
      "q35-9b-fp8-tp1-mtp3-graph-fp8head-c1"
     ],
     "x_from": "config.mtp"
    },
    {
     "label": "draft INT4 head",
     "measurement_ids": [
      "q35-9b-fp8-tp1-mtp0-graph-c1",
      "q35-9b-fp8-tp1-mtp3-graph-int4head-c2",
      "q35-9b-fp8-tp1-mtp4-graph-int4head-c3",
      "q35-9b-fp8-tp1-mtp5-graph-int4head-c9",
      "q35-9b-fp8-tp1-mtp6-graph-int4head-c10"
     ],
     "x_from": "config.mtp"
    }
   ]
  },
  {
   "id": "qwen-9b-tp",
   "title": "One card vs two cards",
   "subtitle": "FP8-dynamic \u00b7 strict completions suite \u00b7 graph capture on \u00b7 MTP0 and depth 3 with the draft INT4 head",
   "x_label": "tensor parallel cards",
   "discrete": true,
   "metrics": [
    "decode_tok_s",
    "ttft_ms"
   ],
   "series": [
    {
     "label": "MTP0",
     "measurement_ids": [
      "q35-9b-fp8-tp1-mtp0-graph-c1",
      "q35-9b-fp8-tp2-mtp0-graph-c5"
     ],
     "x_from": "config.tp"
    },
    {
     "label": "depth 3 + draft INT4 head",
     "measurement_ids": [
      "q35-9b-fp8-tp1-mtp3-graph-int4head-c2",
      "q35-9b-fp8-tp2-mtp3-graph-int4head-c5"
     ],
     "x_from": "config.tp"
    }
   ]
  },
  {
   "id": "qwen-9b-quant",
   "title": "INT4 versus FP8 on one B70",
   "subtitle": "Same model, same launcher, same suite. W4A16 routes to the row-invariant INT4 kernel; FP8 does not.",
   "x_label": "MTP depth",
   "discrete": true,
   "metrics": [
    "decode_tok_s",
    "ttft_ms"
   ],
   "series": [
    {
     "label": "W4A16",
     "measurement_ids": [
      "q35-9b-w4a16-tp1-mtp0-graph-w1",
      "q35-9b-w4a16-tp1-mtp3-graph-int4head-w1"
     ],
     "x_from": "config.mtp"
    },
    {
     "label": "FP8-dynamic",
     "measurement_ids": [
      "q35-9b-fp8-tp1-mtp0-graph-c1",
      "q35-9b-fp8-tp1-mtp3-graph-int4head-c2"
     ],
     "x_from": "config.mtp"
    }
   ]
  }
 ],
 "coverage_views": [
  {
   "id": "qwen-9b-mtp-by-tp",
   "label": "MTP \u00d7 TP (FP8 route)",
   "fixed": "FP8-dynamic revision, R276 vLLM XPU, graph capture on, draft INT4 head where applicable; depth 4 rows are withheld (not lossless).",
   "row_axis": {
    "key": "mtp",
    "label": "MTP",
    "prefix": "MTP"
   },
   "column_axis": {
    "key": "tp",
    "label": "TP",
    "prefix": "TP"
   },
   "fixed_selectors": {
    "revision": "qwen3.5-9b-fp8-dynamic-790f0576",
    "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)"
   },
   "rows": [
    0,
    3,
    4,
    5,
    6
   ],
   "columns": [
    1,
    2
   ],
   "cells": {
    "0:1": {
     "state": "lab-measured",
     "label": "D50.18",
     "evidence_id": "q35-9b-fp8-tp1-mtp0-graph-c1",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "3:1": {
     "state": "lab-measured",
     "label": "D98.25",
     "evidence_id": "q35-9b-fp8-tp1-mtp3-graph-int4head-c2",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "0:2": {
     "state": "lab-measured",
     "label": "D79.46",
     "evidence_id": "q35-9b-fp8-tp2-mtp0-graph-c5",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "3:2": {
     "state": "lab-measured",
     "label": "D147.71",
     "evidence_id": "q35-9b-fp8-tp2-mtp3-graph-int4head-c5",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "4:1": {
     "state": "lab-measured",
     "label": "D98.42 withheld (8/12)",
     "evidence_id": "q35-9b-fp8-tp1-mtp4-graph-int4head-c3",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "4:2": {
     "state": "lab-measured",
     "label": "D150.69 withheld (9/12)",
     "evidence_id": "q35-9b-fp8-tp2-mtp4-graph-int4head-c6",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "5:1": {
     "state": "lab-measured",
     "label": "D91.30 withheld (8/12)",
     "evidence_id": "q35-9b-fp8-tp1-mtp5-graph-int4head-c9",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "5:2": {
     "state": "missing",
     "label": "not run (depth 4 already withheld on two cards)"
    },
    "6:1": {
     "state": "lab-measured",
     "label": "D88.60 withheld (8/12)",
     "evidence_id": "q35-9b-fp8-tp1-mtp6-graph-int4head-c10",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "6:2": {
     "state": "missing",
     "label": "not run (depth 4 already withheld on two cards)"
    }
   }
  },
  {
   "id": "qwen-9b-quant-by-mtp",
   "label": "quantization \u00d7 MTP",
   "fixed": "One B70, R276 vLLM XPU, graph capture on, draft INT4 head at depth 3.",
   "row_axis": {
    "key": "variant",
    "label": "Quantization",
    "prefix": ""
   },
   "column_axis": {
    "key": "mtp",
    "label": "MTP",
    "prefix": "MTP"
   },
   "fixed_selectors": {
    "runtime": "vLLM XPU 0.27.2rc1.dev77 (R276 image 521eb277)",
    "tp": 1
   },
   "rows": [
    "FP8-dynamic",
    "W4A16"
   ],
   "columns": [
    0,
    3
   ],
   "cells": {
    "FP8-dynamic:0": {
     "state": "lab-measured",
     "label": "D50.18",
     "evidence_id": "q35-9b-fp8-tp1-mtp0-graph-c1",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "FP8-dynamic:3": {
     "state": "lab-measured",
     "label": "D98.25",
     "evidence_id": "q35-9b-fp8-tp1-mtp3-graph-int4head-c2",
     "packet_id": "qwen35-9b-fp8-b70"
    },
    "W4A16:0": {
     "state": "lab-measured",
     "label": "D64.33",
     "evidence_id": "q35-9b-w4a16-tp1-mtp0-graph-w1",
     "packet_id": "qwen35-9b-w4a16-b70"
    },
    "W4A16:3": {
     "state": "lab-measured",
     "label": "D113.63",
     "evidence_id": "q35-9b-w4a16-tp1-mtp3-graph-int4head-w1",
     "packet_id": "qwen35-9b-w4a16-b70"
    }
   }
  }
 ],
 "family_closures": []
}