54 lines
4.2 KiB
JSON
54 lines
4.2 KiB
JSON
{
|
|
"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": "674e010598eea4c06bdee04ccba413f4c34194fbe2d70b143c6ec2169b638cc5",
|
|
"attention_true_mixed_fused.csv": "92cd22f8b669abc7af488cf3d067ed880a6cced2dba1a3213a93a4f02a0001a8",
|
|
"linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
|
|
"moe.csv": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
|
|
},
|
|
"profile_id": "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4",
|
|
"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",
|
|
"attention_true_mixed": "Preserved as fused-total diagnostics and excluded from attention.csv because Frontier requires an unobservable prefill/decode split",
|
|
"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": 102,
|
|
"attention_true_mixed_fused_diagnostic": 30,
|
|
"linear": 36,
|
|
"moe": 72
|
|
},
|
|
"schema_version": "frontier_qwen30_vllm020_frozen_profile.v1"
|
|
}
|