[agent:cc] benchmarks: null borrowed conc=1 prefill/power on conc=4 rows (honest structured fields, not prose-only)

data/results.jsonl

@@ -102,5 +102,5 @@
 {"variant":"desktop","vram_gb":24,"memory_type":"gddr7","bus_width_bit":192,"bandwidth_gbps":672,"tdp_w":140,"pcie_lanes":16,"gpu_count":2,"model":"qwopus36-27b-q4km","params_b":27,"quant":"gguf-q4_k_m","backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","driver_version":"610.43.02","os":"Ubuntu 24.04.4 LTS","context_len":2048,"concurrency":1,"tok_s_prefill":64.89,"tok_s_decode":32.79,"vram_used_gb":16.855,"power_w":206.083,"thinking":"off","source":"owner-measured","submitter":"Andrew","date":"2026-06-18","notes":"Seeded from llm-lab card-bench qwopus36-27b-q4km-2gpu; one row per context; median of kept runs; production stack stopped for clean measurement."}
 {"variant":"desktop","vram_gb":24,"memory_type":"gddr7","bus_width_bit":192,"bandwidth_gbps":672,"tdp_w":140,"pcie_lanes":16,"gpu_count":2,"model":"qwopus36-27b-q4km","params_b":27,"quant":"gguf-q4_k_m","backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","driver_version":"610.43.02","os":"Ubuntu 24.04.4 LTS","context_len":8192,"concurrency":1,"tok_s_prefill":65.15,"tok_s_decode":31.86,"vram_used_gb":17.102,"power_w":205.447,"thinking":"off","source":"owner-measured","submitter":"Andrew","date":"2026-06-18","notes":"Seeded from llm-lab card-bench qwopus36-27b-q4km-2gpu; one row per context; median of kept runs; production stack stopped for clean measurement."}
 {"variant": "desktop", "vram_gb": 24, "memory_type": "gddr7", "bus_width_bit": 192, "bandwidth_gbps": 672, "tdp_w": 140, "pcie_lanes": 16, "gpu_count": 2, "model": "qwen36-35b-a3b-nvfp4-verifier", "params_b": 35, "quant": "nvfp4", "backend": "vllm", "backend_version": "0.23.0-a24a03aa", "driver_version": "610.43.02", "os": "Ubuntu 24.04.4 LTS", "context_len": 8192, "concurrency": 1, "tok_s_prefill": 14178.05, "tok_s_decode": 198.99, "vram_used_gb": 44.912, "power_w": 232.333, "thinking": "off", "source": "owner-measured", "submitter": "Andrew", "date": "2026-06-22", "notes": "vLLM concurrency bench 2026-06-22: stock NVFP4 ModelOpt verifier of Qwen3.6-35B-A3B served alone (no DFlash draft) on the 0.23.0-cu129 container, dual-GPU tensor-parallel + expert-parallel. Single-stream decode, 3-run median, client-stream timing. Same-session llama.cpp Q4_K_M baseline 130.69 tok/s decode at this context. Production stack stopped for clean measurement."}
-{"variant": "desktop", "vram_gb": 24, "memory_type": "gddr7", "bus_width_bit": 192, "bandwidth_gbps": 672, "tdp_w": 140, "pcie_lanes": 16, "gpu_count": 2, "model": "qwen36-35b-a3b-nvfp4-verifier", "params_b": 35, "quant": "nvfp4", "backend": "vllm", "backend_version": "0.23.0-a24a03aa", "driver_version": "610.43.02", "os": "Ubuntu 24.04.4 LTS", "context_len": 4096, "concurrency": 4, "tok_s_prefill": 14178.05, "tok_s_decode": 391.79, "vram_used_gb": 45.12, "power_w": 232.333, "thinking": "off", "source": "owner-measured", "submitter": "Andrew", "date": "2026-06-22", "notes": "vLLM concurrency bench 2026-06-22: 4-way concurrent decode, tok_s_decode is the aggregate over 4 streams (256 tok/stream, ignore_eos for equal length), EP = tensor-parallel-size 2 + expert-parallel. tok_s_prefill and power_w are the single-stream (concurrency-1) references for the same config; vram_used_gb is the served dual-card footprint at 0.90 util. Same-session llama.cpp Q4_K_M 4-way aggregate 253.44 tok/s. Production stopped for clean measurement. Caveat (2026-06-22 adversarial review): sustained 256-tok/stream decode (ignore_eos forces equal length), NOT natural-chat throughput; single measured volley, not a 3-run median. Re-measure script committed at bench/vllm_concurrent_remeasure.sh."}
-{"variant": "desktop", "vram_gb": 24, "memory_type": "gddr7", "bus_width_bit": 192, "bandwidth_gbps": 672, "tdp_w": 140, "pcie_lanes": 16, "gpu_count": 2, "model": "qwen36-35b-a3b-q4-k-m", "params_b": 35, "quant": "gguf-q4_k_m", "backend": "llama.cpp", "backend_version": "94a220cd6745e6e3f8de62870b66fd5b9bc92700", "driver_version": "610.43.02", "os": "Ubuntu 24.04.4 LTS", "context_len": 4096, "concurrency": 4, "tok_s_prefill": 183.55, "tok_s_decode": 253.44, "vram_used_gb": 22.368, "power_w": 178.307, "thinking": "off", "source": "owner-measured", "submitter": "Andrew", "date": "2026-06-22", "notes": "vLLM concurrency bench 2026-06-22: llama.cpp production baseline, 4-way concurrent decode, tok_s_decode is the aggregate over 4 streams (256 tok/stream), dual-GPU layer-split, q8_0 KV. tok_s_prefill is the single-stream (concurrency-1) reference. Same-session vLLM NVFP4 EP 4-way aggregate 391.79 tok/s. Production stopped for clean measurement. Caveat (2026-06-22 adversarial review): sustained 256-tok/stream decode (ignore_eos forces equal length), NOT natural-chat throughput; single measured volley, not a 3-run median. Re-measure script committed at bench/vllm_concurrent_remeasure.sh."}
+{"variant": "desktop", "vram_gb": 24, "memory_type": "gddr7", "bus_width_bit": 192, "bandwidth_gbps": 672, "tdp_w": 140, "pcie_lanes": 16, "gpu_count": 2, "model": "qwen36-35b-a3b-nvfp4-verifier", "params_b": 35, "quant": "nvfp4", "backend": "vllm", "backend_version": "0.23.0-a24a03aa", "driver_version": "610.43.02", "os": "Ubuntu 24.04.4 LTS", "context_len": 4096, "concurrency": 4, "tok_s_prefill": null, "tok_s_decode": 391.79, "vram_used_gb": 45.12, "power_w": null, "thinking": "off", "source": "owner-measured", "submitter": "Andrew", "date": "2026-06-22", "notes": "vLLM concurrency bench 2026-06-22 (CC-PROMPT-vllm-concurrency-bench). 4-way concurrent SUSTAINED decode: tok_s_decode is the aggregate over 4 streams at 256 tok/stream forced via ignore_eos (equal-length vs llama's natural 256); EP = tensor-parallel-size 2 + expert-parallel, 0.23.0-cu129 container. NOT natural-chat throughput; single measured volley (one-shot), not a 3-run median. tok_s_prefill and power_w are null - not separately measured at concurrency 4; vram_used_gb is the served dual-card footprint at 0.90 util. Same-session llama.cpp Q4_K_M 4-way aggregate 253.44 tok/s. Re-measure script: bench/vllm_concurrent_remeasure.sh. Production stopped for clean measurement."}
+{"variant": "desktop", "vram_gb": 24, "memory_type": "gddr7", "bus_width_bit": 192, "bandwidth_gbps": 672, "tdp_w": 140, "pcie_lanes": 16, "gpu_count": 2, "model": "qwen36-35b-a3b-q4-k-m", "params_b": 35, "quant": "gguf-q4_k_m", "backend": "llama.cpp", "backend_version": "94a220cd6745e6e3f8de62870b66fd5b9bc92700", "driver_version": "610.43.02", "os": "Ubuntu 24.04.4 LTS", "context_len": 4096, "concurrency": 4, "tok_s_prefill": null, "tok_s_decode": 253.44, "vram_used_gb": 22.368, "power_w": 178.307, "thinking": "off", "source": "owner-measured", "submitter": "Andrew", "date": "2026-06-22", "notes": "vLLM concurrency bench 2026-06-22: llama.cpp production baseline, 4-way concurrent SUSTAINED decode: tok_s_decode is the aggregate over 4 streams at 256 tok/stream (natural, hit the 256 cap), dual-GPU layer-split, q8_0 KV. Single measured volley (one-shot), not a 3-run median. tok_s_prefill is null - not separately measured at concurrency 4; power_w is the real 4-way concurrent measurement; vram_used_gb is the concurrent footprint. Same-session vLLM NVFP4 EP 4-way aggregate 391.79 tok/s. Production stopped for clean measurement."}