EXP
001
publisheddemo-runpileschema 1.08/16/2026

Gemma 4 NVFP4 throughput optimization on RTX 5090

Checkpoint, concurrency, GPU rotation, KV-cache, and scheduler sweeps for a prefill-heavy binary-judge workload.

binary-judgegemma-4llm-inferenceprefill-heavyrtx-5090
01

Question

Maximize total throughput for a large fixed-input binary-judge batch while retaining enough evidence to reproduce and interpret every configuration.

Method. One model replica per RTX 5090. Synthetic fixed 6,000-token prompts and greedy one-token outputs. Eight warmup requests precede timed closed-loop runs. Scheduler candidates use short screenings followed by 500-request validation near the optimum.

02

Author conclusion

The Unsloth checkpoint with a 10,240 scheduler token budget produced the best measured single-replica result around 8.65 requests/s. Larger budgets did not improve throughput and 12,288 failed at startup. Rotation results do not indicate a material device outlier.

Author-supplied interpretation · not a Runpile endorsement

03

Scheduler sweep

Token budget vs. throughput

Raw measured points · no smoothing

requests/s
Scheduler sweep raw data
Max batched tokensRequests per second
20487.43
40968.38
61448.5
81928.53
102408.65
107528.58
112648.52
117768.56
04

Runs

Complete run record

Completed, failed, interrupted, repeated, and derived runs remain visible.

Open in explorer →
#RunConfigurationResultStatusEvidence
01NVIDIA checkpoint, concurrency 8c=8 · 10,2408.07 req/scompleted35
02Unsloth checkpoint, concurrency 8c=8 · 10,2408.64 req/scompleted35
03Red Hat AI checkpoint, concurrency 8c=8 · 10,2408.21 req/scompleted35
04Concurrency sweep, c=1c=1 · 10,2401.18 req/scompleted35
05Concurrency sweep, c=2c=2 · 10,2402.31 req/scompleted35
06Concurrency sweep, c=4c=4 · 10,2404.55 req/scompleted35
07Concurrency sweep, c=8c=8 · 10,2408.64 req/scompleted35
08Scheduler token budget 2048c=8 · 2,0487.43 req/scompleted35
09Scheduler token budget 4096c=8 · 4,0968.38 req/scompleted35
10Scheduler token budget 6144c=8 · 6,1448.50 req/scompleted35
11Scheduler token budget 8192c=8 · 8,1928.53 req/scompleted35
12Scheduler token budget 10240c=8 · 10,2408.65 req/scompleted35
13Scheduler token budget 10752c=8 · 10,7528.58 req/scompleted35
14Scheduler token budget 11264c=8 · 11,2648.52 req/scompleted35
15Scheduler token budget 11776c=8 · 11,7768.56 req/scompleted35
16GPU rotation validation, device 0c=8 · 10,2408.61 req/scompleted35
17GPU rotation validation, device 1c=8 · 10,2408.58 req/scompleted35
18GPU rotation validation, device 2c=8 · 10,2408.63 req/scompleted35
19GPU rotation validation, device 3c=8 · 10,2408.60 req/scompleted35
20Full 500-request validation Ac=8 · 10,2408.54 req/scompleted35
21Full 500-request validation Bc=8 · 10,2408.59 req/scompleted35
22Full 500-request validation Cc=8 · 10,2408.64 req/scompleted35
23Scheduler token budget 12288 (startup failure)c=8 · 12,288failed35
24Interrupted preliminary screeningc=8 · 8,192interrupted35
25Derived four-replica throughputc=8 · 10,24034.48 req/scompleted35
05

Notes

Limitations

  • Synthetic tokens do not represent semantic prompt distributions.
  • The four-replica aggregate is derived from independent rates, not timed end-to-end.
  • Power was not measured.
Complete experiment manifest
{
  "kind": "experiment",
  "runs": [
    {
      "name": "NVIDIA checkpoint, concurrency 8",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.07,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "96d622112c81fab027dc97c56d16c7715a97e4ec3f1c94f0f3148878da03390a",
              "license": "CC-BY-4.0",
              "filename": "checkpoint-nvidia-c8.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "checkpoint-nvidia-c8",
              "client_artifact_id": "checkpoint-nvidia-c8-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "checkpoint-nvidia-c8-trial-1",
          "duration_seconds": 61.95786864931846
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "checkpoint-nvidia-c8"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 61.95786864931846,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.07,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 48420,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.07,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "03f11a2ecf5d5c682c2720f1fb9bd46001bcf651",
        "identifier": "nvidia/Gemma-4-26B-A4B-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "nvidia",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:01:01.957Z",
      "relation": {
        "type": "sweep_point",
        "group": "checkpoint"
      },
      "warnings": [
        "Trust-remote-code was required by the submitted checkpoint."
      ],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "03f11a2ecf5d5c682c2720f1fb9bd46001bcf651",
          "identifier": "nvidia/Gemma-4-26B-A4B-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "96d622112c81fab027dc97c56d16c7715a97e4ec3f1c94f0f3148878da03390a",
          "license": "CC-BY-4.0",
          "filename": "checkpoint-nvidia-c8.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "checkpoint-nvidia-c8",
          "client_artifact_id": "checkpoint-nvidia-c8-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "checkpoint-nvidia-c8",
      "duration_seconds": 61.95786864931846
    },
    {
      "name": "Unsloth checkpoint, concurrency 8",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.6446,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "959ec01dd0c84f493e5877882900ea0aba84a942ec1e655539c19a6fc51b08d8",
              "license": "CC-BY-4.0",
              "filename": "checkpoint-unsloth-c8.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "checkpoint-unsloth-c8",
              "client_artifact_id": "checkpoint-unsloth-c8-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "checkpoint-unsloth-c8-trial-1",
          "duration_seconds": 57.83957615158596
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "checkpoint-unsloth-c8"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 57.83957615158596,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.6446,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51867.600000000006,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.6446,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:57.839Z",
      "relation": {
        "type": "sweep_point",
        "group": "checkpoint"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "959ec01dd0c84f493e5877882900ea0aba84a942ec1e655539c19a6fc51b08d8",
          "license": "CC-BY-4.0",
          "filename": "checkpoint-unsloth-c8.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "checkpoint-unsloth-c8",
          "client_artifact_id": "checkpoint-unsloth-c8-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "checkpoint-unsloth-c8",
      "duration_seconds": 57.83957615158596
    },
    {
      "name": "Red Hat AI checkpoint, concurrency 8",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.21,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "192782b5bbc317bf7245b1de8311bf7a565069932e86b6a147e746d29bc9e1cf",
              "license": "CC-BY-4.0",
              "filename": "checkpoint-redhat-c8.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "checkpoint-redhat-c8",
              "client_artifact_id": "checkpoint-redhat-c8-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "checkpoint-redhat-c8-trial-1",
          "duration_seconds": 60.90133982947624
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "checkpoint-redhat-c8"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 60.90133982947624,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.21,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 49260.00000000001,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.21,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "fdd8f6ca3d63c0a1db4ae62d105fabcc788726b9",
        "identifier": "RedHatAI/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "redhat",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:01:00.901Z",
      "relation": {
        "type": "sweep_point",
        "group": "checkpoint"
      },
      "warnings": [
        "Checkpoint emitted a compressed-tensors compatibility warning at startup."
      ],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "fdd8f6ca3d63c0a1db4ae62d105fabcc788726b9",
          "identifier": "RedHatAI/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "192782b5bbc317bf7245b1de8311bf7a565069932e86b6a147e746d29bc9e1cf",
          "license": "CC-BY-4.0",
          "filename": "checkpoint-redhat-c8.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "checkpoint-redhat-c8",
          "client_artifact_id": "checkpoint-redhat-c8-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "checkpoint-redhat-c8",
      "duration_seconds": 60.90133982947624
    },
    {
      "name": "Concurrency sweep, c=1",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 1.18,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "48e9dc2d6d7adc253e7171b1c4e459ecbdd35ec6f02acaf853208b28b2ac135e",
              "license": "CC-BY-4.0",
              "filename": "concurrency-1.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "concurrency-1",
              "client_artifact_id": "concurrency-1-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "concurrency-1-trial-1",
          "duration_seconds": 423.7288135593221
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "concurrency-1"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 423.7288135593221,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 1.18,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 7080,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 1.18,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:07:03.728Z",
      "relation": {
        "type": "sweep_point",
        "group": "concurrency"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 1,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "48e9dc2d6d7adc253e7171b1c4e459ecbdd35ec6f02acaf853208b28b2ac135e",
          "license": "CC-BY-4.0",
          "filename": "concurrency-1.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "concurrency-1",
          "client_artifact_id": "concurrency-1-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "concurrency-1",
      "duration_seconds": 423.7288135593221
    },
    {
      "name": "Concurrency sweep, c=2",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 2.31,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "2a08f16881b942d7242c7415e5d851ce8171434b0c9131f261430a61c3a1f3b8",
              "license": "CC-BY-4.0",
              "filename": "concurrency-2.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "concurrency-2",
              "client_artifact_id": "concurrency-2-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "concurrency-2-trial-1",
          "duration_seconds": 216.45021645021646
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "concurrency-2"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 216.45021645021646,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 2.31,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 13860,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 2.31,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:03:36.450Z",
      "relation": {
        "type": "sweep_point",
        "group": "concurrency"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 2,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "2a08f16881b942d7242c7415e5d851ce8171434b0c9131f261430a61c3a1f3b8",
          "license": "CC-BY-4.0",
          "filename": "concurrency-2.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "concurrency-2",
          "client_artifact_id": "concurrency-2-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "concurrency-2",
      "duration_seconds": 216.45021645021646
    },
    {
      "name": "Concurrency sweep, c=4",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 4.55,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "9a2e687a197735ab05807e7747e83ba83f579630839b358e3719990dbc3e4c52",
              "license": "CC-BY-4.0",
              "filename": "concurrency-4.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "concurrency-4",
              "client_artifact_id": "concurrency-4-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "concurrency-4-trial-1",
          "duration_seconds": 109.8901098901099
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "concurrency-4"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 109.8901098901099,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 4.55,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 27300,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 4.55,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:01:49.890Z",
      "relation": {
        "type": "sweep_point",
        "group": "concurrency"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 4,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "9a2e687a197735ab05807e7747e83ba83f579630839b358e3719990dbc3e4c52",
          "license": "CC-BY-4.0",
          "filename": "concurrency-4.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "concurrency-4",
          "client_artifact_id": "concurrency-4-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "concurrency-4",
      "duration_seconds": 109.8901098901099
    },
    {
      "name": "Concurrency sweep, c=8",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.6446,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "ce9a510753fdc8b695fe36d15b2698dfc1043c3fc2b90de750c5dd2af166bb84",
              "license": "CC-BY-4.0",
              "filename": "concurrency-8.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "concurrency-8",
              "client_artifact_id": "concurrency-8-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "concurrency-8-trial-1",
          "duration_seconds": 57.83957615158596
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "concurrency-8"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 57.83957615158596,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.6446,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51867.600000000006,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.6446,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:57.839Z",
      "relation": {
        "type": "sweep_point",
        "group": "concurrency"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "ce9a510753fdc8b695fe36d15b2698dfc1043c3fc2b90de750c5dd2af166bb84",
          "license": "CC-BY-4.0",
          "filename": "concurrency-8.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "concurrency-8",
          "client_artifact_id": "concurrency-8-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "concurrency-8",
      "duration_seconds": 57.83957615158596
    },
    {
      "name": "Scheduler token budget 2048",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 7.43,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "eda6e7673a06238b74d12e16d5a01c4f9b8acafb5f0317c508a21c2836040777",
              "license": "CC-BY-4.0",
              "filename": "scheduler-2048.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-2048",
              "client_artifact_id": "scheduler-2048-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-2048-trial-1",
          "duration_seconds": 67.29475100942126
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-2048"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 67.29475100942126,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 7.43,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 44580,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 7.43,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 2048
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 2048 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:01:07.294Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "eda6e7673a06238b74d12e16d5a01c4f9b8acafb5f0317c508a21c2836040777",
          "license": "CC-BY-4.0",
          "filename": "scheduler-2048.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-2048",
          "client_artifact_id": "scheduler-2048-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-2048",
      "duration_seconds": 67.29475100942126
    },
    {
      "name": "Scheduler token budget 4096",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.38,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "e1365c3b671792993b0211e49c30a8163596b698263cec24c1a5980d4e823c94",
              "license": "CC-BY-4.0",
              "filename": "scheduler-4096.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-4096",
              "client_artifact_id": "scheduler-4096-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-4096-trial-1",
          "duration_seconds": 59.665871121718375
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-4096"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 59.665871121718375,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.38,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 50280.00000000001,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.38,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 4096
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 4096 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:59.665Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "e1365c3b671792993b0211e49c30a8163596b698263cec24c1a5980d4e823c94",
          "license": "CC-BY-4.0",
          "filename": "scheduler-4096.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-4096",
          "client_artifact_id": "scheduler-4096-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-4096",
      "duration_seconds": 59.665871121718375
    },
    {
      "name": "Scheduler token budget 6144",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.5,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "dd045ff5fbfff084caafa2eff08531aeb88dc4c3ffd170bc869da14506765b0b",
              "license": "CC-BY-4.0",
              "filename": "scheduler-6144.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-6144",
              "client_artifact_id": "scheduler-6144-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-6144-trial-1",
          "duration_seconds": 58.8235294117647
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-6144"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.8235294117647,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.5,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51000,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.5,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 6144
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 6144 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.823Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "dd045ff5fbfff084caafa2eff08531aeb88dc4c3ffd170bc869da14506765b0b",
          "license": "CC-BY-4.0",
          "filename": "scheduler-6144.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-6144",
          "client_artifact_id": "scheduler-6144-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-6144",
      "duration_seconds": 58.8235294117647
    },
    {
      "name": "Scheduler token budget 8192",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.53,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "5f2578bef81269c3cd1f75cf57662f234da71a440c0fc60384f9befb8a03b185",
              "license": "CC-BY-4.0",
              "filename": "scheduler-8192.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-8192",
              "client_artifact_id": "scheduler-8192-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-8192-trial-1",
          "duration_seconds": 58.6166471277843
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-8192"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.6166471277843,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.53,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51179.99999999999,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.53,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 8192
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 8192 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.616Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "5f2578bef81269c3cd1f75cf57662f234da71a440c0fc60384f9befb8a03b185",
          "license": "CC-BY-4.0",
          "filename": "scheduler-8192.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-8192",
          "client_artifact_id": "scheduler-8192-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-8192",
      "duration_seconds": 58.6166471277843
    },
    {
      "name": "Scheduler token budget 10240",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.65,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "3202c76774de95899b36ddbb2f920b9f826cc022e25647844428c7865e799c03",
              "license": "CC-BY-4.0",
              "filename": "scheduler-10240.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-10240",
              "client_artifact_id": "scheduler-10240-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-10240-trial-1",
          "duration_seconds": 57.80346820809248
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-10240"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 57.80346820809248,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.65,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51900,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.65,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:57.803Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "3202c76774de95899b36ddbb2f920b9f826cc022e25647844428c7865e799c03",
          "license": "CC-BY-4.0",
          "filename": "scheduler-10240.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-10240",
          "client_artifact_id": "scheduler-10240-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-10240",
      "duration_seconds": 57.80346820809248
    },
    {
      "name": "Scheduler token budget 10752",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.58,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "782b5e9940b819009c30237612eb05b522e98291a1825e59f44f9eb71322372f",
              "license": "CC-BY-4.0",
              "filename": "scheduler-10752.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-10752",
              "client_artifact_id": "scheduler-10752-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-10752-trial-1",
          "duration_seconds": 58.27505827505828
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-10752"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.27505827505828,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.58,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51480,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.58,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10752
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10752 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.275Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "782b5e9940b819009c30237612eb05b522e98291a1825e59f44f9eb71322372f",
          "license": "CC-BY-4.0",
          "filename": "scheduler-10752.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-10752",
          "client_artifact_id": "scheduler-10752-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-10752",
      "duration_seconds": 58.27505827505828
    },
    {
      "name": "Scheduler token budget 11264",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.52,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "9ca84451584837616509201c58a0b1b9ee9ed44fec11a84c4ec6629f3cb568c4",
              "license": "CC-BY-4.0",
              "filename": "scheduler-11264.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-11264",
              "client_artifact_id": "scheduler-11264-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-11264-trial-1",
          "duration_seconds": 58.68544600938967
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-11264"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.68544600938967,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.52,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51120,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.52,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 11264
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 11264 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.685Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "9ca84451584837616509201c58a0b1b9ee9ed44fec11a84c4ec6629f3cb568c4",
          "license": "CC-BY-4.0",
          "filename": "scheduler-11264.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-11264",
          "client_artifact_id": "scheduler-11264-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-11264",
      "duration_seconds": 58.68544600938967
    },
    {
      "name": "Scheduler token budget 11776",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.56,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "1263bf649b4c58f5dc79bb52c5f6d776e9a4196ed2fc9fe07d005569d5da3e30",
              "license": "CC-BY-4.0",
              "filename": "scheduler-11776.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "scheduler-11776",
              "client_artifact_id": "scheduler-11776-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "scheduler-11776-trial-1",
          "duration_seconds": 58.41121495327103
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "scheduler-11776"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.41121495327103,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.56,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51360,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.56,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 11776
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 11776 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.411Z",
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "1263bf649b4c58f5dc79bb52c5f6d776e9a4196ed2fc9fe07d005569d5da3e30",
          "license": "CC-BY-4.0",
          "filename": "scheduler-11776.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-11776",
          "client_artifact_id": "scheduler-11776-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "scheduler-11776",
      "duration_seconds": 58.41121495327103
    },
    {
      "name": "GPU rotation validation, device 0",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.61,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "973ae9f36fea7eb0e527a1430dc72b0c063b80a4c0850ae30d990822847689d4",
              "license": "CC-BY-4.0",
              "filename": "benchmark-20260815T160000.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "gpu-rotation-0",
              "client_artifact_id": "gpu-rotation-0-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "gpu-rotation-0-trial-1",
          "duration_seconds": 23.228803716608596
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 1200200,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "gpu-rotation-0"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 23.228803716608596,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.61,
          "measurement_status": "measured"
        },
        "successful_requests": 200,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51660,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.61,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:23.228Z",
      "relation": {
        "type": "sweep_point",
        "group": "gpu_device"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 200,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "973ae9f36fea7eb0e527a1430dc72b0c063b80a4c0850ae30d990822847689d4",
          "license": "CC-BY-4.0",
          "filename": "benchmark-20260815T160000.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "gpu-rotation-0",
          "client_artifact_id": "gpu-rotation-0-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "gpu-rotation-0",
      "duration_seconds": 23.228803716608596
    },
    {
      "name": "GPU rotation validation, device 1",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.58,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "db9be6d742923bbfdfe15edc396516cd1bf68b53e1791f936aef7baa4e56b5cf",
              "license": "CC-BY-4.0",
              "filename": "benchmark-20260815T160000.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "gpu-rotation-1",
              "client_artifact_id": "gpu-rotation-1-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "gpu-rotation-1-trial-1",
          "duration_seconds": 23.31002331002331
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 1200200,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "gpu-rotation-1"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 23.31002331002331,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.58,
          "measurement_status": "measured"
        },
        "successful_requests": 200,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51480,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.58,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:23.310Z",
      "relation": {
        "type": "sweep_point",
        "group": "gpu_device"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 200,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "db9be6d742923bbfdfe15edc396516cd1bf68b53e1791f936aef7baa4e56b5cf",
          "license": "CC-BY-4.0",
          "filename": "benchmark-20260815T160000.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "gpu-rotation-1",
          "client_artifact_id": "gpu-rotation-1-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 1
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              1
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "gpu-rotation-1",
      "duration_seconds": 23.31002331002331
    },
    {
      "name": "GPU rotation validation, device 2",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.63,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "9d1dd0b12cc4185915ab281d047e885d691f8a3cd21d19d5a423bffdb05efc92",
              "license": "CC-BY-4.0",
              "filename": "benchmark-20260815T160000.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "gpu-rotation-2",
              "client_artifact_id": "gpu-rotation-2-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "gpu-rotation-2-trial-1",
          "duration_seconds": 23.17497103128621
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 1200200,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "gpu-rotation-2"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 23.17497103128621,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.63,
          "measurement_status": "measured"
        },
        "successful_requests": 200,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51780.00000000001,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.63,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:23.174Z",
      "relation": {
        "type": "sweep_point",
        "group": "gpu_device"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 200,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "9d1dd0b12cc4185915ab281d047e885d691f8a3cd21d19d5a423bffdb05efc92",
          "license": "CC-BY-4.0",
          "filename": "benchmark-20260815T160000.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "gpu-rotation-2",
          "client_artifact_id": "gpu-rotation-2-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 2
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              2
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "gpu-rotation-2",
      "duration_seconds": 23.17497103128621
    },
    {
      "name": "GPU rotation validation, device 3",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.6,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "28d1069bd4d13b2ec72a75a0bad436de638914511cee8b3cff6d98918e54e08f",
              "license": "CC-BY-4.0",
              "filename": "benchmark-20260815T160000.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "gpu-rotation-3",
              "client_artifact_id": "gpu-rotation-3-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "gpu-rotation-3-trial-1",
          "duration_seconds": 23.255813953488374
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 1200200,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "gpu-rotation-3"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 23.255813953488374,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.6,
          "measurement_status": "measured"
        },
        "successful_requests": 200,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51600,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.6,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:23.255Z",
      "relation": {
        "type": "sweep_point",
        "group": "gpu_device"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 200,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "28d1069bd4d13b2ec72a75a0bad436de638914511cee8b3cff6d98918e54e08f",
          "license": "CC-BY-4.0",
          "filename": "benchmark-20260815T160000.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "gpu-rotation-3",
          "client_artifact_id": "gpu-rotation-3-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 3
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              3
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "gpu-rotation-3",
      "duration_seconds": 23.255813953488374
    },
    {
      "name": "Full 500-request validation A",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.54,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "1d9a21c83ecddfe9b773fca0b49b6e171d87a4a9eb77d987bb6620a972d98f16",
              "license": "CC-BY-4.0",
              "filename": "validation-500-a.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "validation-500-a",
              "client_artifact_id": "validation-500-a-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "validation-500-a-trial-1",
          "duration_seconds": 58.5480093676815
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "validation-500-a"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.5480093676815,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.54,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51239.99999999999,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.54,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.548Z",
      "relation": {
        "type": "sweep_point",
        "group": "repetition"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "1d9a21c83ecddfe9b773fca0b49b6e171d87a4a9eb77d987bb6620a972d98f16",
          "license": "CC-BY-4.0",
          "filename": "validation-500-a.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "validation-500-a",
          "client_artifact_id": "validation-500-a-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "validation-500-a",
      "duration_seconds": 58.5480093676815
    },
    {
      "name": "Full 500-request validation B",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.59,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "703b6b48dba8c0504795932829b2e275690e0c3c0ad1fb8d63548a8173119ce9",
              "license": "CC-BY-4.0",
              "filename": "validation-500-b.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "validation-500-b",
              "client_artifact_id": "validation-500-b-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "validation-500-b-trial-1",
          "duration_seconds": 58.20721769499418
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "validation-500-b"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 58.20721769499418,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.59,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51540,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.59,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.207Z",
      "relation": {
        "type": "sweep_point",
        "group": "repetition"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "703b6b48dba8c0504795932829b2e275690e0c3c0ad1fb8d63548a8173119ce9",
          "license": "CC-BY-4.0",
          "filename": "validation-500-b.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "validation-500-b",
          "client_artifact_id": "validation-500-b-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "validation-500-b",
      "duration_seconds": 58.20721769499418
    },
    {
      "name": "Full 500-request validation C",
      "status": "completed",
      "trials": [
        {
          "status": "completed",
          "results": {
            "requests_per_second": {
              "unit": "req/s",
              "value": 8.64,
              "measurement_status": "measured"
            }
          },
          "artifacts": [
            {
              "kind": "benchmark_result_json",
              "sha256": "8a51bbb7c93cd36378e17860ee8513afd424d3e4b203f91e5145e1a86dc0bbbf",
              "license": "CC-BY-4.0",
              "filename": "validation-500-c.json",
              "media_type": "application/json",
              "size_bytes": 18432,
              "visibility": "public",
              "run_client_id": "validation-500-c",
              "client_artifact_id": "validation-500-c-raw",
              "raw_requests_included": false
            }
          ],
          "client_trial_id": "validation-500-c-trial-1",
          "duration_seconds": 57.87037037037037
        }
      ],
      "results": {
        "total_tokens": {
          "unit": "tokens",
          "value": 3000500,
          "derivation": "successful_requests × (fixed_input_tokens + fixed_output_tokens)",
          "input_run_ids": [
            "validation-500-c"
          ],
          "measurement_status": "derived"
        },
        "failed_requests": 0,
        "duration_seconds": {
          "unit": "s",
          "value": 57.87037037037037,
          "measurement_status": "measured"
        },
        "requests_per_second": {
          "unit": "req/s",
          "value": 8.64,
          "measurement_status": "measured"
        },
        "successful_requests": 500,
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 51840,
          "measurement_status": "measured"
        },
        "output_tokens_per_second": {
          "unit": "tok/s",
          "value": 8.64,
          "measurement_status": "measured"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype auto"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "auto",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:57.870Z",
      "relation": {
        "type": "sweep_point",
        "group": "repetition"
      },
      "warnings": [],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "8a51bbb7c93cd36378e17860ee8513afd424d3e4b203f91e5145e1a86dc0bbbf",
          "license": "CC-BY-4.0",
          "filename": "validation-500-c.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "validation-500-c",
          "client_artifact_id": "validation-500-c-raw",
          "raw_requests_included": false
        }
      ],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "validation-500-c",
      "duration_seconds": 57.87037037037037
    },
    {
      "name": "Scheduler token budget 12288 (startup failure)",
      "status": "failed",
      "failure": {
        "stage": "runtime_startup",
        "category": "kv_cache_allocation",
        "exit_code": 1,
        "reproducible": true,
        "message_sanitized": "KV cache allocation failed at the requested scheduler token budget.",
        "replacement_client_run_id": "scheduler-11776"
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "max_model_len": 6144,
            "max_num_batched_tokens": 12288
          }
        }
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "revision_kind": "commit"
      },
      "relation": {
        "type": "sweep_point",
        "group": "max_num_batched_tokens"
      },
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "concurrency": 8,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        }
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "d6d09aecbaecfb556265244550b804c38093d029f545384a6e60148635efc7ac",
          "license": "CC-BY-4.0",
          "filename": "startup-sanitized.log",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "scheduler-12288-failed",
          "client_artifact_id": "scheduler-12288-failed-raw",
          "raw_requests_included": false
        }
      ],
      "client_run_id": "scheduler-12288-failed"
    },
    {
      "name": "Interrupted preliminary screening",
      "status": "interrupted",
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "max_num_batched_tokens": 8192
          }
        }
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "revision_kind": "commit"
      },
      "relation": {
        "type": "other",
        "group": "screening"
      },
      "warnings": [
        "Interrupted after 73 of 500 requests; excluded from completed-run rankings."
      ],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "concurrency": 8,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 500
      },
      "artifacts": [
        {
          "kind": "benchmark_result_json",
          "sha256": "01541ded8566c149dd63e6cd577bef82790752293bc07cbc056e457c9c101686",
          "license": "CC-BY-4.0",
          "filename": "screening-interrupted.json",
          "media_type": "application/json",
          "size_bytes": 18432,
          "visibility": "public",
          "run_client_id": "screening-interrupted",
          "client_artifact_id": "screening-interrupted-raw",
          "raw_requests_included": false
        }
      ],
      "client_run_id": "screening-interrupted"
    },
    {
      "name": "Derived four-replica throughput",
      "status": "completed",
      "trials": [],
      "results": {
        "requests_per_second": {
          "unit": "req/s",
          "value": 34.48,
          "derivation": "Sum of four independently timed single-GPU replica rates; not a shared end-to-end workload.",
          "input_run_ids": [
            "gpu-rotation-0",
            "gpu-rotation-1",
            "gpu-rotation-2",
            "gpu-rotation-3"
          ],
          "measurement_status": "derived"
        },
        "input_tokens_per_second": {
          "unit": "tok/s",
          "value": 206880,
          "derivation": "Derived requests/s × fixed 6000 input tokens.",
          "input_run_ids": [
            "derived-four-replica"
          ],
          "measurement_status": "derived"
        }
      },
      "runtime": {
        "engine": "vllm",
        "version": "0.27.1",
        "parameters": {
          "vllm": {
            "moe_backend": "FLASHINFER_CUTLASS",
            "max_model_len": 6144,
            "attention_backend": "TRITON_ATTN",
            "enable_prefix_caching": false,
            "future_scheduler_knob": "preserved",
            "gpu_memory_utilization": 0.95,
            "max_num_batched_tokens": 10240
          }
        },
        "parallelism": {
          "data": 1,
          "tensor": 1,
          "pipeline": 1,
          "replicas": 1
        },
        "command_sanitized": "vllm serve MODEL --max-model-len 6144 --max-num-batched-tokens 10240 --gpu-memory-utilization 0.95 --kv-cache-dtype fp8_e4m3"
      },
      "subject": {
        "type": "model",
        "source": "huggingface",
        "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
        "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4",
        "modalities": [
          "text"
        ],
        "architecture": "Gemma4ForConditionalGeneration",
        "quantization": {
          "scope": "all_linears",
          "format": "compressed-tensors",
          "method": "nvfp4",
          "provider": "unsloth",
          "weight_dtype": "nvfp4",
          "kv_cache_dtype": "fp8_e4m3",
          "activation_dtype": "bfloat16"
        },
        "revision_kind": "commit",
        "parameter_count_total": 26000000000,
        "parameter_count_active": 4000000000
      },
      "ended_at": "2026-08-15T16:00:58.004Z",
      "relation": {
        "type": "aggregate",
        "group": "replica_aggregate"
      },
      "warnings": [
        "This aggregate is derived from independent replicas and was not measured end-to-end."
      ],
      "workload": {
        "mode": "online",
        "task": "text_generation",
        "client": {
          "placement": "same-host",
          "overhead_included": true
        },
        "dataset": {
          "license": "CC0-1.0",
          "generator": "vllm-random",
          "synthetic": true
        },
        "sampling": {
          "temperature": 0,
          "max_output_tokens": 1
        },
        "stopping": {
          "exact_output_tokens": true
        },
        "streaming": false,
        "tokenizer": {
          "revision": "20df0542b1a86ce19f495ac2eca2c7c12bce82f9",
          "identifier": "unsloth/gemma-4-26B-A4B-it-NVFP4"
        },
        "concurrency": 8,
        "random_seed": 42,
        "input_tokens": {
          "kind": "fixed",
          "value": 6000
        },
        "endpoint_type": "openai-compatible-chat",
        "output_tokens": {
          "kind": "fixed",
          "value": 1
        },
        "request_count": 2000,
        "arrival_process": "closed_loop",
        "prefix_behavior": {
          "cache_state": "cold",
          "reset_policy": "server_restart",
          "cache_enabled": false,
          "repetition_ratio": 0,
          "shared_prefix_tokens": 0
        },
        "warmup_request_count": 8,
        "request_rate_per_second": null
      },
      "artifacts": [],
      "started_at": "2026-08-15T16:00:00.000Z",
      "environment": {
        "extensions": {
          "runpile.demo.gpu_rotation": 0
        },
        "accelerators": [
          {
            "count": 1,
            "model": "GeForce RTX 5090",
            "driver": "575.64.03",
            "vendor": "NVIDIA",
            "architecture": "Blackwell",
            "device_indices": [
              0
            ],
            "power_limit_watts": 575,
            "memory_bytes_per_device": 34359738368
          }
        ]
      },
      "client_run_id": "derived-four-replica",
      "duration_seconds": 58.004640371229705
    }
  ],
  "tags": [
    "binary-judge",
    "gemma-4",
    "llm-inference",
    "prefill-heavy",
    "rtx-5090"
  ],
  "title": "Gemma 4 NVFP4 throughput optimization on RTX 5090",
  "shared": {
    "workload": {
      "mode": "online",
      "task": "text_generation",
      "dataset": {
        "license": "CC0-1.0",
        "generator": "vllm-random",
        "synthetic": true
      },
      "sampling": {
        "temperature": 0,
        "max_output_tokens": 1
      },
      "streaming": false,
      "concurrency": 8,
      "random_seed": 42,
      "input_tokens": {
        "kind": "fixed",
        "value": 6000
      },
      "output_tokens": {
        "kind": "fixed",
        "value": 1
      },
      "request_count": 500,
      "arrival_process": "closed_loop",
      "prefix_behavior": {
        "cache_state": "cold",
        "cache_enabled": false,
        "shared_prefix_tokens": 0
      },
      "warmup_request_count": 8,
      "request_rate_per_second": null
    },
    "environment": {
      "os": "Ubuntu 24.04.2 LTS",
      "cpu": {
        "model": "AMD Ryzen Threadripper PRO",
        "logical_cores": 64,
        "physical_cores": 32
      },
      "kernel": "6.11.0-26-generic",
      "topology": "Four discrete PCIe GPUs, one serving process per GPU.",
      "accelerators": [
        {
          "count": 4,
          "model": "GeForce RTX 5090",
          "driver": "575.64.03",
          "vendor": "NVIDIA",
          "architecture": "Blackwell",
          "interconnect": "PCIe 5.0; no NVLink",
          "device_indices": [
            0,
            1,
            2,
            3
          ],
          "power_limit_watts": 575,
          "memory_bytes_per_device": 34359738368
        }
      ],
      "driver_versions": {
        "nvidia": "575.64.03",
        "cuda_toolkit": "12.9"
      },
      "host_architecture": "x86_64",
      "system_memory_bytes": 274877906944,
      "virtual_environment": {
        "python": "3.12.10",
        "dependency_lock_hash": "074c2809a1a0ddf856a6b751e0b20773f8cc0828e7015e49d4daa7678dc38f54"
      }
    },
    "measurement_method": {
      "harness": "vllm bench serve",
      "version": "0.27.1",
      "clock_source": "CLOCK_MONOTONIC",
      "warmup_semantics": "Eight requests completed before the timed batch.",
      "duration_semantics": "First timed request dispatch through final response completion.",
      "tokenization_included": true,
      "model_loading_included": false,
      "retry_failure_treatment": "Failed requests remain in the failure count and are not retried.",
      "client_overhead_included": true,
      "metric_definition_version": "runpile-generation-1"
    }
  },
  "license": "CC-BY-4.0",
  "summary": "Checkpoint, concurrency, GPU rotation, KV-cache, and scheduler sweeps for a prefill-heavy binary-judge workload.",
  "artifacts": [
    {
      "kind": "methodology_markdown",
      "sha256": "a62e7baea6125e3df45961e257f36b4518e2d301df21e137b7f2b706f5016329",
      "license": "CC-BY-4.0",
      "filename": "methodology.md",
      "media_type": "text/markdown",
      "size_bytes": 4200,
      "visibility": "public",
      "client_artifact_id": "experiment-notes",
      "raw_requests_included": false
    }
  ],
  "objective": "Maximize total throughput for a large fixed-input binary-judge batch while retaining enough evidence to reproduce and interpret every configuration.",
  "extensions": {
    "runpile.demo.fixture": {
      "generated": true,
      "production_seed": true
    }
  },
  "provenance": {
    "tool": "Codex",
    "notes": "Synthetic demonstration data modeled on a motivating local benchmark. Values are clearly labeled demo data outside tests.",
    "source": "agent",
    "generated_at": "2026-08-16T14:00:00.000Z",
    "tool_version": "demo-fixture-generator-1"
  },
  "visibility": "public",
  "conclusions": "The Unsloth checkpoint with a 10,240 scheduler token budget produced the best measured single-replica result around 8.65 requests/s. Larger budgets did not improve throughput and 12,288 failed at startup. Rotation results do not indicate a material device outlier.",
  "limitations": [
    "Synthetic tokens do not represent semantic prompt distributions.",
    "The four-replica aggregate is derived from independent rates, not timed end-to-end.",
    "Power was not measured."
  ],
  "methodology": "One model replica per RTX 5090. Synthetic fixed 6,000-token prompts and greedy one-token outputs. Eight warmup requests precede timed closed-loop runs. Scheduler candidates use short screenings followed by 500-request validation near the optimum.",
  "schema_version": "1.0",
  "publisher_notes": {
    "warnings": [
      "Two GPU rotation artifacts intentionally share a display filename to demonstrate why IDs and hashes are authoritative."
    ],
    "redactions": [
      "Local username, hostname, absolute paths, and LAN addresses were removed."
    ],
    "omitted_files": [
      "unsanitized-server.log",
      "raw-prompts.jsonl"
    ],
    "public_data_license_affirmed": true
  },
  "client_submission_id": "0198b812-cde0-7a2d-8cd7-51d426080001"
}