{ "capacity": [ { "config": { "mns": 16, "name": "tp2_mns16", "tp": 2 }, "lower_censored": false, "maximum_tested_feasible_request_rate": 14.0, "maximum_tested_feasible_request_rate_per_gpu": 7.0, "upper_censored": true }, { "config": { "mns": 32, "name": "tp2_mns32", "tp": 2 }, "lower_censored": false, "maximum_tested_feasible_request_rate": 14.0, "maximum_tested_feasible_request_rate_per_gpu": 7.0, "upper_censored": true }, { "config": { "mns": 64, "name": "tp2_mns64", "tp": 2 }, "lower_censored": false, "maximum_tested_feasible_request_rate": 14.0, "maximum_tested_feasible_request_rate_per_gpu": 7.0, "upper_censored": true }, { "config": { "mns": 8, "name": "tp2_mns8", "tp": 2 }, "lower_censored": false, "maximum_tested_feasible_request_rate": 14.0, "maximum_tested_feasible_request_rate_per_gpu": 7.0, "upper_censored": true } ], "config_results": [ { "config": { "mns": 8, "name": "tp2_mns8", "tp": 2 }, "loads": [ { "config": { "mns": 8, "name": "tp2_mns8", "tp": 2 }, "elapsed_seconds": 10.598209381103516, "offered_request_rate": 10.0, "offered_request_rate_per_gpu": 5.0, "request_count": 64, "request_metrics_sha256": "3597984a1acaffc6b3e8b1baa3ff41bb94ee1618f8780fcc6cd8d3bdb3d7201b", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 9.814433634929404, "ttft_max_ms": 300.1784228926625, "ttft_p50_ms": 211.39433537475938, "ttft_p95_ms": 297.1650738307572 }, "status": "completed", "trace_sha256": "92e3001d8467e3f97a243bea8323440ecf254a8ca211e52226fe886ebc9404be" }, { "config": { "mns": 8, "name": "tp2_mns8", "tp": 2 }, "elapsed_seconds": 10.660333633422852, "offered_request_rate": 12.0, "offered_request_rate_per_gpu": 6.0, "request_count": 64, "request_metrics_sha256": "06641f2758bd3f801fb7b9c934670bc766372ad7353f5e14aea2988dfbec462a", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 11.55742285377119, "ttft_max_ms": 389.93612032123883, "ttft_p50_ms": 296.08352841701094, "ttft_p95_ms": 384.6764910356253 }, "status": "completed", "trace_sha256": "cbc40c05fb9e400fa5edea6dd52c4122a957418fe523d102f1e72a784321e681" }, { "config": { "mns": 8, "name": "tp2_mns8", "tp": 2 }, "elapsed_seconds": 10.409451007843018, "offered_request_rate": 14.0, "offered_request_rate_per_gpu": 7.0, "request_count": 64, "request_metrics_sha256": "94cf0f3c510015ba50495eb38fd6fb452dd23d16ce0c788dc926229ecd557812", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 13.374732701347012, "ttft_max_ms": 388.3013907870674, "ttft_p50_ms": 309.4635231444016, "ttft_p95_ms": 386.4490667329015 }, "status": "completed", "trace_sha256": "3e43801bb724218ceaf0d52f17cd4d0fb3863e70f95c4d8d5a43ce7f85c75c01" } ] }, { "config": { "mns": 16, "name": "tp2_mns16", "tp": 2 }, "loads": [ { "config": { "mns": 16, "name": "tp2_mns16", "tp": 2 }, "elapsed_seconds": 10.360321521759033, "offered_request_rate": 10.0, "offered_request_rate_per_gpu": 5.0, "request_count": 64, "request_metrics_sha256": "3597984a1acaffc6b3e8b1baa3ff41bb94ee1618f8780fcc6cd8d3bdb3d7201b", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 9.814433634929404, "ttft_max_ms": 300.1784228926625, "ttft_p50_ms": 211.39433537475938, "ttft_p95_ms": 297.1650738307572 }, "status": "completed", "trace_sha256": "92e3001d8467e3f97a243bea8323440ecf254a8ca211e52226fe886ebc9404be" }, { "config": { "mns": 16, "name": "tp2_mns16", "tp": 2 }, "elapsed_seconds": 10.402549028396606, "offered_request_rate": 12.0, "offered_request_rate_per_gpu": 6.0, "request_count": 64, "request_metrics_sha256": "06641f2758bd3f801fb7b9c934670bc766372ad7353f5e14aea2988dfbec462a", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 11.55742285377119, "ttft_max_ms": 389.93612032123883, "ttft_p50_ms": 296.08352841701094, "ttft_p95_ms": 384.6764910356253 }, "status": "completed", "trace_sha256": "cbc40c05fb9e400fa5edea6dd52c4122a957418fe523d102f1e72a784321e681" }, { "config": { "mns": 16, "name": "tp2_mns16", "tp": 2 }, "elapsed_seconds": 10.40966272354126, "offered_request_rate": 14.0, "offered_request_rate_per_gpu": 7.0, "request_count": 64, "request_metrics_sha256": "94cf0f3c510015ba50495eb38fd6fb452dd23d16ce0c788dc926229ecd557812", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 13.374732701347012, "ttft_max_ms": 388.3013907870674, "ttft_p50_ms": 309.4635231444016, "ttft_p95_ms": 386.4490667329015 }, "status": "completed", "trace_sha256": "3e43801bb724218ceaf0d52f17cd4d0fb3863e70f95c4d8d5a43ce7f85c75c01" } ] }, { "config": { "mns": 32, "name": "tp2_mns32", "tp": 2 }, "loads": [ { "config": { "mns": 32, "name": "tp2_mns32", "tp": 2 }, "elapsed_seconds": 10.155991315841675, "offered_request_rate": 10.0, "offered_request_rate_per_gpu": 5.0, "request_count": 64, "request_metrics_sha256": "3597984a1acaffc6b3e8b1baa3ff41bb94ee1618f8780fcc6cd8d3bdb3d7201b", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 9.814433634929404, "ttft_max_ms": 300.1784228926625, "ttft_p50_ms": 211.39433537475938, "ttft_p95_ms": 297.1650738307572 }, "status": "completed", "trace_sha256": "92e3001d8467e3f97a243bea8323440ecf254a8ca211e52226fe886ebc9404be" }, { "config": { "mns": 32, "name": "tp2_mns32", "tp": 2 }, "elapsed_seconds": 10.103453874588013, "offered_request_rate": 12.0, "offered_request_rate_per_gpu": 6.0, "request_count": 64, "request_metrics_sha256": "06641f2758bd3f801fb7b9c934670bc766372ad7353f5e14aea2988dfbec462a", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 11.55742285377119, "ttft_max_ms": 389.93612032123883, "ttft_p50_ms": 296.08352841701094, "ttft_p95_ms": 384.6764910356253 }, "status": "completed", "trace_sha256": "cbc40c05fb9e400fa5edea6dd52c4122a957418fe523d102f1e72a784321e681" }, { "config": { "mns": 32, "name": "tp2_mns32", "tp": 2 }, "elapsed_seconds": 10.255317687988281, "offered_request_rate": 14.0, "offered_request_rate_per_gpu": 7.0, "request_count": 64, "request_metrics_sha256": "94cf0f3c510015ba50495eb38fd6fb452dd23d16ce0c788dc926229ecd557812", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 13.374732701347012, "ttft_max_ms": 388.3013907870674, "ttft_p50_ms": 309.4635231444016, "ttft_p95_ms": 386.4490667329015 }, "status": "completed", "trace_sha256": "3e43801bb724218ceaf0d52f17cd4d0fb3863e70f95c4d8d5a43ce7f85c75c01" } ] }, { "config": { "mns": 64, "name": "tp2_mns64", "tp": 2 }, "loads": [ { "config": { "mns": 64, "name": "tp2_mns64", "tp": 2 }, "elapsed_seconds": 10.106109380722046, "offered_request_rate": 10.0, "offered_request_rate_per_gpu": 5.0, "request_count": 64, "request_metrics_sha256": "3597984a1acaffc6b3e8b1baa3ff41bb94ee1618f8780fcc6cd8d3bdb3d7201b", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 9.814433634929404, "ttft_max_ms": 300.1784228926625, "ttft_p50_ms": 211.39433537475938, "ttft_p95_ms": 297.1650738307572 }, "status": "completed", "trace_sha256": "92e3001d8467e3f97a243bea8323440ecf254a8ca211e52226fe886ebc9404be" }, { "config": { "mns": 64, "name": "tp2_mns64", "tp": 2 }, "elapsed_seconds": 10.112579107284546, "offered_request_rate": 12.0, "offered_request_rate_per_gpu": 6.0, "request_count": 64, "request_metrics_sha256": "06641f2758bd3f801fb7b9c934670bc766372ad7353f5e14aea2988dfbec462a", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 11.55742285377119, "ttft_max_ms": 389.93612032123883, "ttft_p50_ms": 296.08352841701094, "ttft_p95_ms": 384.6764910356253 }, "status": "completed", "trace_sha256": "cbc40c05fb9e400fa5edea6dd52c4122a957418fe523d102f1e72a784321e681" }, { "config": { "mns": 64, "name": "tp2_mns64", "tp": 2 }, "elapsed_seconds": 9.847446918487549, "offered_request_rate": 14.0, "offered_request_rate_per_gpu": 7.0, "request_count": 64, "request_metrics_sha256": "94cf0f3c510015ba50495eb38fd6fb452dd23d16ce0c788dc926229ecd557812", "score": { "feasible": true, "pass_rate": 1.0, "passed": 64, "throughput_requests_per_second": 13.374732701347012, "ttft_max_ms": 388.3013907870674, "ttft_p50_ms": 309.4635231444016, "ttft_p95_ms": 386.4490667329015 }, "status": "completed", "trace_sha256": "3e43801bb724218ceaf0d52f17cd4d0fb3863e70f95c4d8d5a43ce7f85c75c01" } ] } ], "contract": { "arrival": "open_loop_uniform", "input_tokens": 2048, "output_tokens": 1, "prefix_caching": false, "rates": [ 10.0, 12.0, 14.0 ], "requests_per_anchor": 64, "target_pass_rate": 0.95, "ttft_slo_ms": 1256.0 }, "frontier": { "git_head": "d9cfeb6d8791fbf2f295dd9744c56a666171776e", "git_status_short": " M frontier/config/config.py\n M frontier/entities/request.py\n M frontier/events/cluster_schedule_event.py\n M frontier/execution_time_predictor/sklearn_execution_time_predictor.py\n M frontier/metrics/constants.py\n M frontier/metrics/metrics_store.py\n M frontier/profiling/common/layers/rotary_embedding.py\n M frontier/profiling/moe/moe_impl.py\n M frontier/profiling/moe/moe_vllm_kernel.py\n M frontier/scheduler/cluster_scheduler/__init__.py\n M frontier/scheduler/cluster_scheduler/base_cluster_scheduler.py\n M frontier/scheduler/cluster_scheduler/cluster_scheduler_registry.py\n M frontier/scheduler/cluster_scheduler/sticky_lor_cluster_scheduler.py\n M frontier/scheduler/replica_scheduler/base_replica_scheduler.py\n M frontier/scheduler/replica_scheduler/vllm_v1_engine_replica_scheduler.py\n M frontier/scheduler/replica_stage_scheduler/replica_stage_schduler.py\n M frontier/simulator.py\n M frontier/types/cluster_scheduler_type.py\n?? data/profiling/compute/h20/\n?? frontier/scheduler/cluster_scheduler/prefix_lor_cluster_scheduler.py\n?? runs/\n?? tests/unit/test_attn_prefill_prediction_fallback.py\n", "source": "/tmp/replayserve-frontier-rs1b" }, "profiles": { "coverage": { "attention": { "1": { "exact_prefill_2048_rows": 1, "profile_batch_size": 1 }, "2": { "exact_prefill_2048_rows": 1, "profile_batch_size": 1 }, "4": { "exact_prefill_2048_rows": 1, "profile_batch_size": 1 } }, "manifest": { "attention_tp_coverage": [ 1, 2, 4 ], "environment_contract": { "dtype": "bfloat16", "frontier_commit": "d9cfeb6d8791fbf2f295dd9744c56a666171776e", "hardware": "NVIDIA H20", "model": "Qwen3-30B-A3B", "tensor_parallel_sizes": [ 1, 2, 4 ], "vllm_source_commit": "88d34c6409e9fb3c7b8ca0c04756f061d2099eb1", "vllm_version": "0.20.0" }, "inputs": { "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp2-20260716-v1-dispatch-aware-20260716T140743025781Z/artifacts/artifacts/allreduce-full-tp2-v1/raw/allreduce-tp2.json": "97c3c76b5a04e95bd9192423c2b891667c668f39cc0dfecbd097d749939f2d0a", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp4-20260716-v1-dispatch-aware-20260716T141106009788Z/artifacts/artifacts/allreduce-full-tp4-v1/raw/allreduce-tp4.json": "809df9baa6f468cf12bf0c99827475acc67894dd9f3f948976590b665fac0e76", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp1-20260716-v2-20260716T135132587012Z/artifacts/artifacts/flashattn-kv-full-v2-tp1/raw/flashattn-tp1.json": "dcb4c1bf7e76b9c765f78ddd2b8a734f2d7ba2adac13ce017689a8a77fe69a27", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp2-20260716-v2-20260716T135134194295Z/artifacts/artifacts/flashattn-kv-full-v2-tp2/raw/flashattn-tp2.json": "43ce042556ba887c8860614b43ccf0f564e5cebc1a0cffbce299d0acb9fa8d07", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp4-20260716-v2-20260716T135135197200Z/artifacts/artifacts/flashattn-kv-full-v2-tp4/raw/flashattn-tp4.json": "84eef31bcad0f556907a093318a420959d14fdc94474823d11f659704bdfec73", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-frontier-linear-full-20260716-v2-max-tokens-20260716T144444676943Z/artifacts/artifacts/frontier-linear-full-v2/profiles/compute/h20/qwen3-a3b-30b-moe/linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-moe-full-20260716-v1-local-shard-20260716T141334565164Z/artifacts/artifacts/moe-full-v1/raw/moe-full.json": "588f6ad0d69c9636d1b852e3df0a12d13cfe731f050ea7ec7aea457cceefbde8", "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-router-full-20260716-v3-tp-context-20260716T145446098505Z/artifacts/artifacts/router-full-v3/raw/router.json": "1962972e983bff3e06a721ef4ae4ec65728ff669681497a4a7e7f769b88b4931" }, "outputs": { "allreduce.json": "b38d14f990578d668523d25b107aceed433da5020d8ada3b6e44d3562261a3b3", "attention.csv": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d", "attention_true_mixed_fused.csv": "43ef4be90bddc9aeac6dbbe339feec24162cd1f2129a08fbd959e6ee4eaf5f60", "linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c", "moe.csv": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0" }, "profile_id": "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4-fused-mixed-total-conserving", "projection_contract": { "allreduce": "Frozen exact runtime measurements; base profile-only comparison keeps the historical Frontier CC backend fixed to isolate compute profile fidelity", "attention": "Pure prefill/extend/decode FA3 core plus separately measured KV update; input/output reshape assumed zero; exported mean is used as median target; true mixed rows use a total-conserving compatibility projection", "attention_true_mixed": "The directly measured fused total is preserved in diagnostics. Frontier's two targets are projected by the same-TP pure prefill/decode reference ratio, with projected prefill + decode exactly equal to the fused total; the split is a schema compatibility attribution, not an observation", "linear": "Frontier profiler using vLLM 0.20 CUDA operators", "moe": "Replicated gate and fused top-k plus TP-local modular expert kernel; expert measurement already includes prepare/finalize so shuffling is zero" }, "row_counts": { "allreduce": 24, "attention_frontier_compatible": 132, "attention_true_mixed_fused_diagnostic": 30, "linear": 36, "moe": 72 }, "schema_version": "frontier_qwen30_vllm020_frozen_profile.v2" } }, "root": "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/frozen/profile-v2", "sha256": { "attention": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d", "linear": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c", "manifest": "af40545e75aff55c6333cd2d5379ccf042a5a0b7d7fc7df4f745ce256cb290eb", "moe": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0" } }, "schema": "frontier-qwen30-prefill-surface-v1", "status": "partial_not_decision_bearing" }