130 lines
7.8 KiB
JSON
130 lines
7.8 KiB
JSON
{
|
|
"attention_long_context_coverage": {
|
|
"by_tp": {
|
|
"1": {
|
|
"decode_kv_lengths": [
|
|
128,
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
16384,
|
|
32768,
|
|
40960
|
|
],
|
|
"true_mixed_decode_avg_kv_lengths": [
|
|
512,
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
16384,
|
|
32768
|
|
]
|
|
},
|
|
"2": {
|
|
"decode_kv_lengths": [
|
|
128,
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
16384,
|
|
32768,
|
|
40960
|
|
],
|
|
"true_mixed_decode_avg_kv_lengths": [
|
|
512,
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
16384,
|
|
32768
|
|
]
|
|
},
|
|
"4": {
|
|
"decode_kv_lengths": [
|
|
128,
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
16384,
|
|
32768,
|
|
40960
|
|
],
|
|
"true_mixed_decode_avg_kv_lengths": [
|
|
512,
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
16384,
|
|
32768
|
|
]
|
|
}
|
|
},
|
|
"included": true
|
|
},
|
|
"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-fidelity-envelope-v1/fleet-artifacts/qwen30-attention-composition-tp1-20260717-v1-20260717T014550833983Z/artifacts/artifacts/attention-composition-tp1-v1/raw/flashattn-composition-tp1.json": "ce461be92a5059e2d168b2cbf21b55d9a55ee71c3b0586a3e63cd8288938168d",
|
|
"/home/gahow/phd/aituner/runs/frontier-fidelity-envelope-v1/fleet-artifacts/qwen30-attention-composition-tp2-20260717-v1-20260717T014551831545Z/artifacts/artifacts/attention-composition-tp2-v1/raw/flashattn-composition-tp2.json": "997d67fb3a28b488c3aa4945ae7c6d4ecf8a08726ee3cb4652d16c221ff395ad",
|
|
"/home/gahow/phd/aituner/runs/frontier-fidelity-envelope-v1/fleet-artifacts/qwen30-attention-composition-tp4-20260717-v1-20260717T014552897206Z/artifacts/artifacts/attention-composition-tp4-v1/raw/flashattn-composition-tp4.json": "046cffd4c11d8a36513167ec25d3e1ca016324715679ea688a58d4a71dfe62b2",
|
|
"/home/gahow/phd/aituner/runs/frontier-fidelity-envelope-v1/fleet-artifacts/qwen30-attention-long-context-tp1-20260717-v1-20260717T022700381447Z/artifacts/artifacts/attention-long-context-tp1-v1/raw/flashattn-long-context-tp1.json": "da3b5adaa0d06f6a0ab7394d3c85d8d0524f0bd571613785f1a6183f70eb457f",
|
|
"/home/gahow/phd/aituner/runs/frontier-fidelity-envelope-v1/fleet-artifacts/qwen30-attention-long-context-tp2-20260717-v1-20260717T022701357659Z/artifacts/artifacts/attention-long-context-tp2-v1/raw/flashattn-long-context-tp2.json": "d503acb959f91b1713ff02d3251c781c240c4b057208033f4fd4a637af76e2cb",
|
|
"/home/gahow/phd/aituner/runs/frontier-fidelity-envelope-v1/fleet-artifacts/qwen30-attention-long-context-tp4-20260717-v1-20260717T022702416725Z/artifacts/artifacts/attention-long-context-tp4-v1/raw/flashattn-long-context-tp4.json": "a24acd44ab4075a3ac0aad00b91e676b325b2427dbe14282260cae6769481cf7",
|
|
"/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": "4f0c42d2c47f52880d81b8d2b2d7d52bd62d2c34e0f3a7f94c4c6cbd64d751c8",
|
|
"attention_true_mixed_fused.csv": "c84c129d466aec97b42f1255f3b7363fe7f73948dc27ed65b6ce599d6fb9af6d",
|
|
"linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
|
|
"moe.csv": "dd698113975e5ed77f51dfd8a69aebf764ed651c3f6f34b71914d25d6e001415"
|
|
},
|
|
"profile_id": "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4-fused-mixed-total-conserving-pure-prefill-batch-composition-long-context-decode-mixed",
|
|
"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_long_context": "Direct FA3 measurements at decode KV lengths 16K/32K/40K, plus true-mixed prefill/decode measurements at 16K/32K, for TP1/2/4",
|
|
"attention_pure_prefill_batch_composition": "Direct FA3 measurements for 2/4 requests at query length 2048 and 2/4/8/16 requests at query length 512 for each TP; included only when batch-composition attention inputs are supplied",
|
|
"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": 213,
|
|
"attention_true_mixed_fused_diagnostic": 48,
|
|
"linear": 36,
|
|
"moe": 72
|
|
},
|
|
"schema_version": "frontier_qwen30_vllm020_frozen_profile.v4"
|
|
}
|