Add six measured single-card B70 results (vLLM, OpenVINO GenAI, Cascadia, llama.cpp)

results/SergiioB/2026-08-30-qwen3-6-27b-int4-vllm.json

@@ -0,0 +1,16 @@
+{
+  "rig": "1",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-6-27b",
+  "quant": "int4",
+  "runtime": "vllm",
+  "runtimeVersion": "0.26.1rc1.dev457+gc810e5ee9.xpu (vllm-xpu-kernels 0.1.12)",
+  "runtimeFlags": "native MTP4 spec decode; fp8 KV (required at 128K); prefix cache on; scheduler 8192; gpu-memory-utilization 0.88",
+  "decodeTps": 69.3,
+  "promptTps": 1755,
+  "contextLength": 131072,
+  "batchSize": 1,
+  "runDate": "2026-08-30",
+  "notes": "Dense 27B. C1 decode at p512/g128, client post-first-token rate, median of n=5 after one same-output same-shape warmup; prefix cache enabled with unique entropy-first cold prefixes and zero cache-hit delta. promptTps is the cold-input rate at p2048 for the same MTP4 mode (actual input tokens / client TTFT), not isolated engine prefill. Single card of a 2x B70 rig. Configured cap 230 W (not measured draw). Checkpoint llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-GPTQ-Int4 (GPTQ-INT4 G128, native MTP head preserved), served from image vllm/vllm-openai-xpu@sha256:2c427ef477da092eb6f2cdbbbd24950b5fa171565b916db69d4c7bb10e68ca97 with patch_mtp_nightly.py + patch_mtp_boundary.py. Correctness: coherent-output smoke. Evidence: cookbook docs/qwen36-27/QWEN36-DENSE-VLLM-XPU.md, evidence commit a4ba8754; runDate is that evidence-record date. Self-reported, E2."
+}

results/SergiioB/2026-08-30-qwen3-6-35b-a3b-int4-vllm.json

@@ -0,0 +1,16 @@
+{
+  "rig": "1",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-6-35b-a3b",
+  "quant": "int4",
+  "runtime": "vllm",
+  "runtimeVersion": "0.26.1rc1.dev457+gc810e5ee9.xpu (vllm-xpu-kernels 0.1.12)",
+  "runtimeFlags": "native MTP4 spec decode; prefix cache on; scheduler 8192",
+  "decodeTps": 170.91,
+  "promptTps": 4532,
+  "contextLength": 131072,
+  "batchSize": 1,
+  "runDate": "2026-08-30",
+  "notes": "MoE (35B total, 3B active). C1 decode, client post-first-token rate, median of n=5 after one same-output same-shape warmup; prefix cache enabled with unique entropy-first cold prefixes and zero cache-hit delta; scheduler 8192; context 131072; C1. promptTps is the cold-input rate at p512 for the same MTP4 mode (actual input tokens / client TTFT), not isolated engine prefill. Single card of a 2x B70 rig. Configured cap 165 W (not measured draw). Checkpoint llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GPTQ-Int4 (GPTQ-INT4 G128, native MTP head preserved), served from image vllm/vllm-openai-xpu@sha256:2c427ef477da092eb6f2cdbbbd24950b5fa171565b916db69d4c7bb10e68ca97. Correctness: coherent-output smoke. Evidence: cookbook docs/qwen36-35a3/QWEN36-MOE-VLLM-XPU.md section 2, evidence commit a4ba8754; runDate is that evidence-record date. Self-reported, E2."
+}

results/SergiioB/2026-09-11-qwen3-8-27b-int4-cascadia.json

@@ -0,0 +1,15 @@
+{
+  "rig": "1",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "int4",
+  "runtime": "cascadia",
+  "runtimeVersion": "Cascadia 0.2.3 qwen35 backend (OpenVINO 2026.3.1)",
+  "runtimeFlags": "no speculative decode; OV LATENCY hint; INFERENCE_PRECISION=f16; prefix cache 1 GB",
+  "decodeTps": 4.97,
+  "contextLength": 131072,
+  "batchSize": 1,
+  "runDate": "2026-09-11",
+  "notes": "Dense 27B hybrid-attention (GDN). Wall tok/s over 128 generated tokens including prefill, p128/g128, C1, median of n=5 - wall-based, so it slightly understates steady decode. No speculative decode. Research engine: one-stage qwen35 backend over OpenVINO 2026.3.1; OpenVINO GenAI on the same IR measured 15.79 tok/s pre-MTP and 79.5 tok/s with a grafted MTP5 head (n=3 screen, not submitted here), so this number reflects the Cascadia backend, not the weights. Single card of a 2x B70 rig. Configured cap 230 W; measured 101.9 W mean draw, 65 C max temp (card is not power-bound here). Coherent-output quality PASS on all five samples. Checkpoint IR: SergiioB/Qwen3.8-27B-int4-gdn8-ov (INT4 with GDN kept in float). Evidence: raw n5-cascadia-int4-gdn8-230w.json (2026-09-11); published in cookbook docs/qwen38-27b/OPENVINO-CASCADIA-REPORT.md. Self-reported, E2."
+}

results/SergiioB/2026-09-14-qwen3-8-27b-int4-vllm.json

@@ -0,0 +1,16 @@
+{
+  "rig": "1",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "int4",
+  "runtime": "vllm",
+  "runtimeVersion": "0.27.2rc1.dev77+gac7509e2b",
+  "runtimeFlags": "MTP4 (Draft INT4 S+M1); prefix cache on; XPU graphs on; v5 scheduler patch",
+  "decodeTps": 106.7,
+  "ttftMs": 335,
+  "contextLength": 131072,
+  "batchSize": 1,
+  "runDate": "2026-09-14",
+  "notes": "Dense 27B hybrid-attention (GDN). C1 decode with calibrated realistic Pi prompts, client post-first-token rate, median of n=5 (spread 103.2-111.3), MTP acceptance 89-96%. Prefix cache enabled. Single card of a 2x B70 rig. Configured cap 230 W (not measured draw). Checkpoint SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 (GPTQ-INT4 G128) with the separate INT4 draft head (S+M1); recommended-model-card sampling regimes kept separate from greedy diagnostics. Measured through the LocalMaxxing short-prompt harness against our pinned server (record cmt03mj040eh8ms01trjvhm75). Correctness: coherent-output smoke. Evidence: cookbook docs/qwen38-27b/QWEN38-VLLM-XPU.md current-stack C1 table. Self-reported, E2."
+}

results/SergiioB/2026-09-14-qwen3-8-27b-int8-openvino-genai.json

@@ -0,0 +1,16 @@
+{
+  "rig": "1",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "int8",
+  "runtime": "openvino-genai",
+  "runtimeVersion": "OpenVINO GenAI 2026.5 nightly (VLMPipeline)",
+  "runtimeFlags": "MTP nat5; KV f16; greedy; PERFORMANCE_HINT=LATENCY; INFERENCE_PRECISION=f16",
+  "decodeTps": 62.07,
+  "promptTps": 1536.5,
+  "ttftMs": 353.4,
+  "batchSize": 1,
+  "runDate": "2026-09-14",
+  "notes": "Dense 27B hybrid-attention (GDN). Short-context cell of the 2026-09-14 cross-engine sweep: p512 prompt (543 actual input tokens), g128 requested with ignore_eos, C1, client post-first-token rate, median of n=5. Greedy synthetic counting task - MTP acceptance and rate on natural text differ. KV f16 is clean to 48K context; long-context ceilings measured separately (u8/u4 paths, 128K full-protocol best). promptTps is the cold-input rate (actual input tokens / TTFT), not isolated engine prefill. Single card of a 2x B70 rig. Configured cap 230 W; measured 202 W mean card draw (energy1_input counter over the measured interval). Weights 25.8 GB; 4.0 GiB VRAM free after load. Checkpoint OpenVINO/Qwen3.8-27B-int8-ov (int8-ov VL split IR). Correctness: coherent-output smoke. Evidence: raw ctx-sweep-openvino.json in the 2026-09-14 sweep; published in cookbook docs/qwen38-27b/ENGINE-COMPARISON-CTX-20260914.md. Self-reported, E2."
+}

results/SergiioB/2026-09-14-qwen3-8-27b-q4-k-m-llamacpp.json

@@ -0,0 +1,16 @@
+{
+  "rig": "1",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "q4_k_m",
+  "runtime": "llamacpp",
+  "runtimeVersion": "SYCL 0.3.0-dev 1692f9e50 (build 361)",
+  "runtimeFlags": "-fa 1, -ngl 99, -b 8192, -ub 4096, KV q8_0 K / q4_1 V, no speculative decode",
+  "decodeTps": 20.01,
+  "promptTps": 611.73,
+  "contextLength": 8192,
+  "batchSize": 1,
+  "runDate": "2026-09-14",
+  "notes": "Dense 27B hybrid-attention (GDN). llama-bench engine-native tg128 at p512, mean of 5 repetitions (raw emitted repetition data); engine-native timing, not client-observed streaming. promptTps is llama-bench engine-native pp512 from the same repetitions, not the HTTP cold-input rate. No speculative decode; the same GGUF family with native draft-MTP reaches ~38 tok/s at this cell (n=3 screen, not submitted as representative). Q4_K_M 16.81 GB freshly quantized from the SHA-verified BF16 release (not a third-party quant); decode is cap-insensitive (15.2 tok/s at both 150 W and 230 W on the Q8_0 arm). Single card of a 2x B70 rig. Configured cap 150 W. Correctness: coherent-output smoke plus llama-quantize perplexity arms. Evidence: cookbook docs/qwen38-27b/QUANT-FORMAT-SHOWDOWN-B70.md (commit 22bd3253). Self-reported, E2."
+}