{ "metadata": { "version": "0.4.29", "engine": "vllm", "model": "glm-5.3-flash-dflash2", "server": "192.168.1.201:8000", "timestamp": "2026-09-02T16:58:14.207747", "decode_mode": "duration", "primary_decode_layer": "sustained_decode", "duration_per_test": 10.0, "request_count": 0, "warmup_request_count": 0, "run_burst": false, "prefill_mode": "skipped", "standalone_prefill": false, "prefill_only": false, "skip_prefill": true, "burst_e2e_status": "not_run_use_--run-burst", "burst_request_count": 0, "burst_warmup_request_count": 0, "burst_requests_per_concurrency": 5, "decode_warmup_seconds": 3.0, "decode_warmup_context": 0, "decode_warmup_concurrency": 1, "cell_warmup_timeout_seconds": 0.0, "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0", "show_capacity_limited_values": false, "max_tokens": 512, "temperature": 0.0, "ignore_eos": true, "max_total_tokens": 6624000, "dcp_size": 0, "metrics_available": true, "metrics_warning": "", "concurrency_levels": [ 1 ], "context_lengths": [ 0 ], "startup_diagnostics_available": true, "nvidia_p2p_override_effective": false, "p2pmark_status": "not_run", "amd_fabric_status": "not_run" }, "startup_diagnostics": { "version": "0.4.29", "server_url": "http://192.168.1.201:8000", "hostname": "macmini", "uname": "Darwin macmini 25.5.0 Darwin Kernel Version 25.5.0: Tue Jun 9 22:28:34 PDT 2026; root:xnu-12377.121.10~1/RELEASE_ARM64_T6041 arm64", "env": {}, "args": { "concurrency": "1", "contexts": "0", "max_tokens": 512, "duration": 10.0, "request_count": 0, "run_burst": false, "standalone_prefill": false, "prefill_only": false, "skip_prefill": true, "prefill_contexts": "8k,64k,128k", "prefill_metric": "client", "dcp_size": 0, "kv_budget": 0 }, "nvidia_p2p_override": { "effective": false, "configured": false, "params_path": "/proc/driver/nvidia/params", "params_available": false, "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf", "modprobe_available": false, "runtime": { "ForceP2P": "", "RMForceP2PType": "", "RMPcieP2PType": "", "GrdmaPciTopoCheckOverride": "", "EnableResizableBar": "", "DmaRemapPeerMmio": "" }, "expected": { "ForceP2P": "0x11", "RMForceP2PType": "1", "RMPcieP2PType": "2", "GrdmaPciTopoCheckOverride": "1", "EnableResizableBar": "1" }, "missing": [ "ForceP2P", "RMForceP2PType", "RMPcieP2PType", "GrdmaPciTopoCheckOverride", "EnableResizableBar" ], "mismatched": {}, "registry_dwords": "", "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"", "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded" }, "p2pmark": { "status": "not_run" }, "amd_fabric": { "status": "not_run" }, "nvidia_smi_error": "nvidia-smi not found" }, "nvidia_p2p_override": { "effective": false, "configured": false, "params_path": "/proc/driver/nvidia/params", "params_available": false, "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf", "modprobe_available": false, "runtime": { "ForceP2P": "", "RMForceP2PType": "", "RMPcieP2PType": "", "GrdmaPciTopoCheckOverride": "", "EnableResizableBar": "", "DmaRemapPeerMmio": "" }, "expected": { "ForceP2P": "0x11", "RMForceP2PType": "1", "RMPcieP2PType": "2", "GrdmaPciTopoCheckOverride": "1", "EnableResizableBar": "1" }, "missing": [ "ForceP2P", "RMForceP2PType", "RMPcieP2PType", "GrdmaPciTopoCheckOverride", "EnableResizableBar" ], "mismatched": {}, "registry_dwords": "", "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"", "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded" }, "p2pmark": { "status": "not_run" }, "amd_fabric": { "status": "not_run" }, "hardware_run_summary": {}, "event_log": [ "16:57:44 benchmark start engine=vllm", "16:57:44 startup server=http://192.168.1.201:8000 model=glm-5.3-flash-dflash2", "16:57:44 startup decode concurrency=1 contexts=0", "16:57:44 startup NVIDIA P2P override: unknown: /proc/driver/nvidia/params is not readable", "16:57:44 startup engine vLLM 0.1.dev20051+g487ecf187 models=['glm-5.3-flash-dflash2']", "16:57:44 startup KV cache budget from vLLM metrics: 6,624,000 tokens (2875 blocks x 2304)", "16:57:44 startup model context length: 262,144 tokens", "16:57:44 startup prefill tests: skipped", "16:57:44 startup startup preparation done", "16:57:44 startup hardware monitor disabled", "16:57:44 decode warmup start", "16:57:45 decode warmup start C=1 ctx=0 3s", "16:57:45 cell start C=1 ctx=0", "16:57:51 ready C=1 ctx=0 running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s", "16:57:54 cell done C=1 ctx=0 41.7 tok/s | norm 13.5 step/s len=3.10", "16:57:54 decode warmup done C=1 ctx=0", "16:57:56 cell start C=1 ctx=0", "16:58:02 ready C=1 ctx=0 running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s", "16:58:12 cell done C=1 ctx=0 38.3 tok/s | norm 13.1 step/s len=2.92" ], "prefill": {}, "results": [ { "concurrency": 1, "context_tokens": 0, "benchmark_mode": "duration", "request_count_target": 0, "warmup_request_count": 0, "measurement_seconds": 9.979069, "measurement_wall_seconds": 10.00222, "client_output_tokens": 382, "server_output_tokens": 382, "aggregate_source": "openai_continuous_usage", "aggregate_tps": 38.2801233299203, "per_request_avg_tps": 38.2801233299203, "ttft_avg": 0.24859522949554957, "ttft_p50": 0.24859522949554957, "ttft_p90": 0.26818874588352626, "ttft_p99": 0.27259728707082104, "time_to_second_token_avg": 0.07653589600522537, "time_to_second_token_p50": 0.07653589600522537, "time_to_second_token_p90": 0.07764024640491698, "time_to_second_token_p99": 0.07788872524484759, "request_latency_avg": 13.67058420900139, "request_latency_p50": 13.67058420900139, "request_latency_p90": 13.67058420900139, "request_latency_p99": 13.67058420900139, "inter_token_latency_avg": 0.022360881372554525, "inter_token_latency_p50": 0.022360881372554525, "inter_token_latency_p90": 0.025523418348860513, "inter_token_latency_p99": 0.026234989168529357, "output_tps_per_user_avg": 46.16378523847843, "output_tps_per_user_p50": 46.16378523847843, "output_tps_per_user_p90": 52.69280685218498, "output_tps_per_user_p99": 54.16183671526895, "e2e_output_tps_per_user_avg": 37.45267884476172, "e2e_output_tps_per_user_p50": 37.45267884476172, "e2e_output_tps_per_user_p90": 37.45267884476172, "e2e_output_tps_per_user_p99": 37.45267884476172, "chunk_inter_token_latency_avg": 0.07604885094410449, "chunk_inter_token_latency_p50": 0.07604885094410449, "chunk_inter_token_latency_p90": 0.07679192778881462, "chunk_inter_token_latency_p99": 0.0769591200788744, "input_seq_len_avg": 78.0, "output_seq_len_avg": 512.0, "output_seq_len_p50": 512.0, "output_seq_len_p90": 512.0, "output_seq_len_p99": 512.0, "request_count": 2, "completed_request_count": 1, "request_samples": [ { "ttft": 0.2241033340105787, "time_to_second_token": 0.07515545800561085, "latency": 13.67058420900139, "inter_token_latency_avg": 0.02631405259293701, "chunk_inter_token_latency_avg": 0.07512000488821682, "input_tokens": 78, "output_tokens": 512, "output_tps_per_user": 38.00250822134525, "e2e_output_tps_per_user": 37.45267884476172, "completed": true }, { "ttft": 0.27308712498052046, "time_to_second_token": 0.07791633400483988, "latency": 0.0, "inter_token_latency_avg": 0.01840771015217204, "chunk_inter_token_latency_avg": 0.07697769699999216, "input_tokens": 78, "output_tokens": 93, "output_tps_per_user": 54.32506225561161, "e2e_output_tps_per_user": 0.0, "completed": false } ], "total_tokens": 382, "wall_time": 15.714896708988817, "num_completed": 1, "num_errors": 0, "server_gen_throughput": 38.087877857376206, "server_utilization": 0.01217814892136393, "server_spec_accept_rate": 0.27582417582417584, "server_spec_accept_length": 2.930769230769231, "server_spec_drafts": 130, "server_spec_draft_tokens": 910, "server_spec_accepted_tokens": 251, "server_spec_pos_accept": [ 0.7154, 0.4692, 0.2846, 0.1846, 0.1154, 0.0923, 0.0692 ], "server_engine_steps": 131.0, "server_steps_per_s": 13.127476848742301, "server_accept_len_effective": 2.9160305343511452, "accept_norm_tps": 0.0, "accept_norm_ref_len": 0.0, "avg_running_reqs": 1, "max_running_reqs": 1, "effective_concurrency": 1, "avg_queue_reqs": 0, "max_queue_reqs": 0, "queue_fraction": 0.0, "underfilled": false, "warmup_timed_out": false, "warmup_duration": 5.66, "ready_reason": "running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s", "timeout_reason": "", "capacity_limited": false, "hardware_summary": {} } ], "summary_table": { "0": { "1": 38.2801233299203 } }, "burst_results": [], "burst_summary_table": {}, "methodology": { "prefill": { "name": "Prefill", "present": false, "mode": "skipped", "formula": "prompt_tokens / TTFT", "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation." }, "sustained_decode": { "name": "Sustained Decode", "present": true, "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable", "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline." }, "burst_e2e_decode": { "name": "Burst / E2E Decode", "present": false, "status": "not run; use --run-burst", "formula": "sum(completion_tokens) / profiling_wall_time", "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion." }, "acceptance_normalization": { "name": "Acceptance-normalized decode (MTP / speculative)", "present": true, "formula": "engine_steps = spec_drafts + max(0, output_tokens - (accepted_tokens + spec_drafts)); accept_len_effective = output_tokens / engine_steps; steps_per_s = aggregate_tps / accept_len_effective", "notes": "With speculative decoding tok/s = steps_per_s * accept_len, so raw tok/s mixes engine speed with data-dependent acceptance. steps_per_s (target-model forward passes per second) is the acceptance-independent speed used to compare runs; server_spec_pos_accept holds per-draft-position acceptance probabilities. Counters are vLLM window deltas; SGLang falls back to its lifetime accept-length gauge." }, "coding_peak": { "name": "Coding Peak", "present": true, "formula": "usage.completion_tokens / (last_stream_time - first_token_time)", "notes": "Sequential cc1 Sieve-of-Eratosthenes coding prompt, matching /mnt/test.py throughput semantics. Uses OpenAI stream usage with continuous_usage_stats when the server supports it." } }, "coding_peak": { "mode": "coding_peak", "prompt": "Write a Python script that implements the Sieve of Eratosthenes.", "runs_requested": 3, "runs_ok": 3, "max_tokens": 2000, "temperature": 0.0, "summary": { "mean_generation_tok_s": 73.16228317941676, "median_generation_tok_s": 73.68997497304997, "max_generation_tok_s": 74.74192110738443, "min_generation_tok_s": 71.0549534578159, "cjk_runs": 0 }, "samples": [ { "run": 1, "ok": true, "finish_reason": "stop", "completion_tokens": 1874, "ttft": 0.24347924999892712, "gen_elapsed": 26.373952958994778, "total_elapsed": 26.617432208993705, "generation_tok_s": 71.0549534578159, "total_tok_s": 70.40498817789036, "content_chars": 7012, "reasoning_chars": 0, "cjk_chars": 0, "content_preview": "The user wants a Python script implementing the Sieve of Eratosthenes. This is a classic algorithm for finding all prime numbers up to a given limit. Let me think about what makes a good implementation.\n\nThe Sieve of Eratosthenes works as follows:\n1. Create a boolean array of size n+1, initialized to True (assuming all numbers are prime initially)\n2. Mark 0 and 1 as not prime\n3. Starting from 2, for each number that is still marked as prime, mark all its multiples as not prime\n4. The key optimiz" }, { "run": 2, "ok": true, "finish_reason": "length", "completion_tokens": 2000, "ttft": 0.19760791698354296, "gen_elapsed": 27.14073387501412, "total_elapsed": 27.338341791997664, "generation_tok_s": 73.68997497304997, "total_tok_s": 73.15732663000905, "content_chars": 7295, "reasoning_chars": 0, "cjk_chars": 0, "content_preview": "The user wants a Python script implementing the Sieve of Eratosthenes. This is a classic algorithm for finding all prime numbers up to a given limit. Let me think about what makes a good implementation:\n\n1. **The algorithm basics:**\n - Create a boolean array of size n+1, initialized to True\n - Mark 0 and 1 as not prime\n - For each number p starting from 2, if p is still marked prime, mark all multiples of p (starting from p²) as not prime\n - Continue until p² > n\n - Collect all indices" }, { "run": 3, "ok": true, "finish_reason": "stop", "completion_tokens": 1829, "ttft": 0.19229045798419975, "gen_elapsed": 24.470872208010405, "total_elapsed": 24.663162665994605, "generation_tok_s": 74.74192110738443, "total_tok_s": 74.15918326329707, "content_chars": 6561, "reasoning_chars": 0, "cjk_chars": 0, "content_preview": "The user wants a Python script implementing the Sieve of Eratosthenes. This is a classic algorithm for finding all prime numbers up to a given limit. Let me think about what makes a good response here.\n\nThe Sieve of Eratosthenes algorithm:\n1. Create a boolean list of size n+1, initialized to True (assuming all numbers are prime)\n2. Mark 0 and 1 as not prime\n3. Starting from 2, for each number p that is still marked prime, mark all multiples of p (starting from p²) as not prime\n4. Continue until " } ] } }