{ "absolute_campaign_deadline": "2026-10-17T16:55:51.522595+00:00", "absolute_request_cutoff": "2026-10-17T15:25:51.522595+00:00", "authorization": "[REDACTED]", "campaign_id": "glm53-serving-performance-20261017", "controls": { "capacity_edge": "Nominal310272=max327680-output16384-margin1024; n4 at C1/2/4 per arm, preceded by excluded C1 warmup. Report actual prompt lengths; no simultaneous fit assumption. Edge runs last within each arm.", "comparison_scope": "whole recipes; no causal MTP or quantization attribution", "context_labels": "Targets are not actual token counts; publish actual min/p50/max. Expected approximately 8K/32K/128K; verify actual lengths and matched arms.", "controlled_requests_per_cell": 20, "hardware": "two RTX PRO 6000 Blackwell Max-Q, same order", "max_tokens": 16384, "model_max_len": 327680, "natural_requests_per_cell": 8, "prefix_cache_flush": false, "prompt_cache_mode": "unique", "reasoning_effort": "max", "request_canaries": true, "same_seed_per_matched_cell": true, "statistics": "p50/p95 descriptive with n; no stable p99 claims", "temperature": 0.6, "top_p": 0.95 }, "created_at": "2026-10-17T10:16:09.420817+00:00", "deadline_starts": "First benchmark request; root records absolute UTC cutoff before dispatch", "dispatch_clock_started_at": "2026-10-17T10:55:51.522595+00:00", "eligibility": [ "all planned requests completed and terminal", "exact usage counts", "allowed finish reason and visible output", "own canary first; no foreign marker; complete capture", "controlled cells additionally exact repeated output", "same measurement_protocol in compared arms" ], "frozen_at": "2026-10-17T10:55:51.522595+00:00", "load_bound_seconds": 3600, "matrix": [ { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "1", "--concurrency", "1", "--ctx-tokens", "10240", "--seed", "481001", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-warmup-ctx10240-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": true, "id": "baseline-warmup-ctx10240-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 1, "seed": 481001, "workload": "warmup" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "1", "--concurrency", "1", "--ctx-tokens", "163840", "--seed", "481011", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-warmup-ctx163840-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": true, "id": "baseline-warmup-ctx163840-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 163840, "requests": 1, "seed": 481011, "workload": "warmup" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "1", "--ctx-tokens", "10240", "--seed", "482001", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx10240-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-natural-ctx10240-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 8, "seed": 482001, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "2", "--ctx-tokens", "10240", "--seed", "482002", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx10240-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-natural-ctx10240-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 8, "seed": 482002, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "4", "--ctx-tokens", "10240", "--seed", "482004", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx10240-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-natural-ctx10240-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 8, "seed": 482004, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "1", "--ctx-tokens", "40960", "--seed", "482011", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx40960-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-natural-ctx40960-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 8, "seed": 482011, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "2", "--ctx-tokens", "40960", "--seed", "482012", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx40960-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-natural-ctx40960-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 8, "seed": 482012, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "4", "--ctx-tokens", "40960", "--seed", "482014", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx40960-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-natural-ctx40960-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 8, "seed": 482014, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "1", "--ctx-tokens", "163840", "--seed", "482021", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx163840-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-natural-ctx163840-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 163840, "requests": 8, "seed": 482021, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "2", "--ctx-tokens", "163840", "--seed", "482022", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx163840-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-natural-ctx163840-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 8, "seed": 482022, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "4", "--ctx-tokens", "163840", "--seed", "482024", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-natural-ctx163840-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-natural-ctx163840-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 8, "seed": 482024, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "1", "--ctx-tokens", "10240", "--seed", "483001", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx10240-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-controlled-ctx10240-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 20, "seed": 483001, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "2", "--ctx-tokens", "10240", "--seed", "483002", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx10240-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-controlled-ctx10240-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 20, "seed": 483002, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "4", "--ctx-tokens", "10240", "--seed", "483004", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx10240-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-controlled-ctx10240-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 20, "seed": 483004, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "1", "--ctx-tokens", "40960", "--seed", "483011", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx40960-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-controlled-ctx40960-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 20, "seed": 483011, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "2", "--ctx-tokens", "40960", "--seed", "483012", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx40960-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-controlled-ctx40960-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 20, "seed": 483012, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "4", "--ctx-tokens", "40960", "--seed", "483014", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx40960-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-controlled-ctx40960-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 20, "seed": 483014, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "1", "--ctx-tokens", "163840", "--seed", "483021", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx163840-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-controlled-ctx163840-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 163840, "requests": 20, "seed": 483021, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "2", "--ctx-tokens", "163840", "--seed", "483022", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx163840-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-controlled-ctx163840-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 20, "seed": 483022, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "4", "--ctx-tokens", "163840", "--seed", "483024", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-controlled-ctx163840-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-controlled-ctx163840-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 20, "seed": 483024, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "1", "--concurrency", "1", "--ctx-tokens", "310272", "--seed", "481031", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-warmup-edge-ctx310272-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": true, "id": "baseline-warmup-edge-ctx310272-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 1, "seed": 481031, "workload": "warmup" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "4", "--concurrency", "1", "--ctx-tokens", "310272", "--seed", "484031", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-capacity-edge-ctx310272-c1.json", "--confirm" ], "arm": "baseline", "concurrency": 1, "excluded_warmup": false, "id": "baseline-capacity-edge-ctx310272-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 4, "seed": 484031, "workload": "capacity-edge" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "4", "--concurrency", "2", "--ctx-tokens", "310272", "--seed", "484032", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-capacity-edge-ctx310272-c2.json", "--confirm" ], "arm": "baseline", "concurrency": 2, "excluded_warmup": false, "id": "baseline-capacity-edge-ctx310272-c2", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 4, "seed": 484032, "workload": "capacity-edge" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31841/v1", "--model", "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--engine", "vllm-exl3-v84", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "4", "--concurrency", "4", "--ctx-tokens", "310272", "--seed", "484034", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/baseline/baseline-capacity-edge-ctx310272-c4.json", "--confirm" ], "arm": "baseline", "concurrency": 4, "excluded_warmup": false, "id": "baseline-capacity-edge-ctx310272-c4", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 4, "seed": 484034, "workload": "capacity-edge" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "1", "--concurrency", "1", "--ctx-tokens", "10240", "--seed", "481001", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-warmup-ctx10240-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": true, "id": "candidate-warmup-ctx10240-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 1, "seed": 481001, "workload": "warmup" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "1", "--concurrency", "1", "--ctx-tokens", "163840", "--seed", "481011", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-warmup-ctx163840-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": true, "id": "candidate-warmup-ctx163840-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 163840, "requests": 1, "seed": 481011, "workload": "warmup" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "1", "--ctx-tokens", "10240", "--seed", "482001", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx10240-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-natural-ctx10240-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 8, "seed": 482001, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "2", "--ctx-tokens", "10240", "--seed", "482002", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx10240-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-natural-ctx10240-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 8, "seed": 482002, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "4", "--ctx-tokens", "10240", "--seed", "482004", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx10240-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-natural-ctx10240-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 8, "seed": 482004, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "1", "--ctx-tokens", "40960", "--seed", "482011", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx40960-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-natural-ctx40960-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 8, "seed": 482011, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "2", "--ctx-tokens", "40960", "--seed", "482012", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx40960-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-natural-ctx40960-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 8, "seed": 482012, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "4", "--ctx-tokens", "40960", "--seed", "482014", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx40960-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-natural-ctx40960-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 8, "seed": 482014, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "1", "--ctx-tokens", "163840", "--seed", "482021", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx163840-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-natural-ctx163840-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 163840, "requests": 8, "seed": 482021, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "2", "--ctx-tokens", "163840", "--seed", "482022", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx163840-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-natural-ctx163840-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 8, "seed": 482022, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "8", "--concurrency", "4", "--ctx-tokens", "163840", "--seed", "482024", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-natural-ctx163840-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-natural-ctx163840-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 8, "seed": 482024, "workload": "natural" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "1", "--ctx-tokens", "10240", "--seed", "483001", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx10240-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-controlled-ctx10240-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 20, "seed": 483001, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "2", "--ctx-tokens", "10240", "--seed", "483002", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx10240-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-controlled-ctx10240-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 20, "seed": 483002, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "4", "--ctx-tokens", "10240", "--seed", "483004", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx10240-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-controlled-ctx10240-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 10240, "requests": 20, "seed": 483004, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "1", "--ctx-tokens", "40960", "--seed", "483011", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx40960-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-controlled-ctx40960-c1", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 20, "seed": 483011, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "2", "--ctx-tokens", "40960", "--seed", "483012", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx40960-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-controlled-ctx40960-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 20, "seed": 483012, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "4", "--ctx-tokens", "40960", "--seed", "483014", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx40960-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-controlled-ctx40960-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 40960, "requests": 20, "seed": 483014, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "1", "--ctx-tokens", "163840", "--seed", "483021", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx163840-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-controlled-ctx163840-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 163840, "requests": 20, "seed": 483021, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "2", "--ctx-tokens", "163840", "--seed", "483022", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx163840-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-controlled-ctx163840-c2", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 20, "seed": 483022, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "20", "--concurrency", "4", "--ctx-tokens", "163840", "--seed", "483024", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "strict", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "128", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-controlled-ctx163840-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-controlled-ctx163840-c4", "outer_timeout_seconds": 1200, "planned_context_tokens": 163840, "requests": 20, "seed": 483024, "workload": "controlled" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "1", "--concurrency", "1", "--ctx-tokens", "310272", "--seed", "481031", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-warmup-edge-ctx310272-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": true, "id": "candidate-warmup-edge-ctx310272-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 1, "seed": 481031, "workload": "warmup" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "4", "--concurrency", "1", "--ctx-tokens", "310272", "--seed", "484031", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-capacity-edge-ctx310272-c1.json", "--confirm" ], "arm": "candidate", "concurrency": 1, "excluded_warmup": false, "id": "candidate-capacity-edge-ctx310272-c1", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 4, "seed": 484031, "workload": "capacity-edge" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "4", "--concurrency", "2", "--ctx-tokens", "310272", "--seed", "484032", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-capacity-edge-ctx310272-c2.json", "--confirm" ], "arm": "candidate", "concurrency": 2, "excluded_warmup": false, "id": "candidate-capacity-edge-ctx310272-c2", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 4, "seed": 484032, "workload": "capacity-edge" }, { "argv": [ "[PRIVATE_OPERATOR_ROOT]/.venv/bin/python", "-m", "anvil_serving.cli", "eval", "benchmark", "capacity", "--base-url", "http://127.0.0.1:31843/v1", "--model", "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--engine", "vllm-b12x-release-10152026", "--gpu", "dual-rtx-pro-6000-maxq", "--requests", "4", "--concurrency", "4", "--ctx-tokens", "310272", "--seed", "484034", "--max-tokens", "16384", "--prompt-cache-mode", "unique", "--request-canaries", "--controlled-output-policy", "observe", "--max-model-len", "327680", "--reasoning-effort", "max", "--temperature", "0.6", "--top-p", "0.95", "--timeout", "900", "--response-words", "0", "--output", "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-serving-performance/native/candidate/candidate-capacity-edge-ctx310272-c4.json", "--confirm" ], "arm": "candidate", "concurrency": 4, "excluded_warmup": false, "id": "candidate-capacity-edge-ctx310272-c4", "outer_timeout_seconds": 1800, "planned_context_tokens": 310272, "requests": 4, "seed": 484034, "workload": "capacity-edge" } ], "metric_profile": "serving-comparison/v1", "metrics": [ "effective_prefill_tok_s (queue-inclusive prompt_tokens/TTFO)", "TTFO (first reasoning or visible delta)", "visible TTFT", "combined reasoning+visible decode", "TPOT per-request proxy", "E2E", "aggregate completion throughput", "actual input/output distributions", "engine prefill/decode/queue histogram deltas when exposed", "GPU weight/KV allocations and allocated token pool", "configured max context/max sequences separately from tested actual context/concurrency", "observed engine active/queued/KV occupancy", "GPU used/free/utilization; host available and cgroup current/peak/limits", "OOM events and admission/preemption/errors" ], "models": { "baseline": { "container": "intelligence-glm-exl3-327k-r11-dcp1-vision8-apc-nospec", "engine": "vllm-exl3-v84", "model": "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "port": 31841, "registry": "[PRIVATE_OPERATOR_ROOT]/serve-recipes.glm53-exl3-327k-r11-dcp1-vision8-apc-nospec.toml", "repository": "brandonmusic/GLM-5.3-Flash-tr3-4bpw", "revision": "a5fee929cf4888b1824323e33e8a19b60129e025" }, "candidate": { "container": "glm53-nvidia-fp4-local-ab-r2", "engine": "vllm-b12x-release-10152026", "model": "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "port": 31843, "registry": "[PRIVATE_OPERATOR_ROOT]/serve-recipes.glm53-nvidia-fp4-ab-20261017-r2.toml", "repository": "nvidia/GLM-5.3-Flash-FP4-Local", "revision": "76348ddb2be2e35e39ecfc00e75dc07703f4f3d5" } }, "owner": "root sole lifecycle owner", "publication": { "bundle": "docs/findings/2026-10-17-glm53-serving-performance-evidence", "finding": "docs/findings/2026-10-17-glm53-serving-performance.md", "old_evidence": "preserve original immutable campaign; link follow-up" }, "quality_ledger": { "SWE": "not run; outside serving-performance follow-up; cannot claim complete model quality qualification", "capacity": "fresh complete frozen matrix above", "context": "reuse actual316K evidence; candidate17/18 original plus18/18 unchanged retry andtwo1/1 isolated remain distinct", "direct_text_tools_JSON": "reuse exact arms original applicable passes; fresh smoke before each arm", "stability_endurance": "single old baseline overlap scout only; not endurance; long cells provide bounded operational exposure only", "synthetic_agentic": "reuse18/18 each; not repository-scale coding", "vision_OCR": "reuse exact arms original18/18 andOCR1/1; text performance only" }, "request_cutoff_seconds": 16200, "restoration": { "argv_source": "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-nvidia-fp4-ab/plans/restore-retained-argv.json", "candidate_load_argv_source": "[PRIVATE_OPERATOR_ROOT]/evidence/2026-10-17-glm53-nvidia-fp4-ab/plans/candidate-load-r2-argv.json", "postconditions": [ "candidate absent", "same retained r11 ID/image/revision", "health and identity match", "direct JSON and authenticated llm.primary checks", "primary readmitted", "starting config and manifest hashes unchanged" ], "preview_precondition": "restore-retained requires stopped target and free GPU custody; expected refusal while incumbent is running. Before stop independently verify exact identity and source hashes; repeat preview after keep-container stop before candidate load.", "retained_container_id": "container-id-redacted" }, "restoration_reserve_seconds": 5400, "source_revision": "5e1f92fbe2624b2121f443c146db422dd36d62cb", "status": "frozen-before-dispatch", "stops": { "OOM_engine_death_CUDA_NCCL": "stop all dispatch and restore exact r11", "admission_or_KV_refusal": "stop only proven supersets with logs and telemetry; timeout is not proof of physical limit", "canary_or_request_loss": "stop higher concurrency on arm; retain cell; safe independent C1 may continue", "capacity_edge": "Proven C1 capacity refusal blocks C2/C4; proven C2 refusal blocks C4. Preserve OOM/restoration stop and request cutoff. Unequal actual prompt distributions make latency comparison non-comparable; do not normalize away differences.", "deadline": "stop new requests at cutoff; terminate owned CLI if outer cap; verify requests drained before unload; reserve protected", "identity_or_protocol_drift": "no further requests on invalid arm" }, "telemetry": "Before and after every cell: raw engine metrics, managed GPU inventory, managed recipe identity/cgroup status and /proc/meminfo. During every cell: timestamped raw engine metrics at 0.5-1 seconds and GPU/cgroup/host samples about every 2 seconds. Retain histogram bucket/count/sum deltas, actual sampling intervals and collector errors. Sampled engine maxima are lower bounds; native producer records exact client stream-call overlap.", "total_bound_seconds": 21600 }