tod / latency.json
Suriyaganesh's picture
utod-picker-12b-v1 release bundle
07d34d5 verified
Raw History Blame Contribute Delete
3.75 kB
{
"device": "NVIDIA A100-SXM4-40GB",
"dtype": "bfloat16",
"attn": "chunked_eager",
"max_tokens": 49152,
"note": "harness --stage1 union refuses in bf16 (pointwise selftest |route-solo|>tol); comparison uses --stage1 retriever. No CLI flag skips/relaxes the pointwise selftest (--selftest sets record COUNT, min 1). picker argmax == harness(retriever) on all 5 rows for BOTH bundles.",
"bundles": {
"utod-picker-12b-v1": {
"temps": {
"T1": 1.7237,
"T2": 1.4778,
"T1_ret": 0.05
},
"load_test": [
{
"id": "gpqa-94",
"N": 4,
"picker_union_argmax": "canthaxanthin_biosynthesis",
"harness_retriever_pred": "canthaxanthin_biosynthesis",
"match": true,
"sum_p": 1.0
},
{
"id": "glaive-98691",
"N": 2,
"picker_union_argmax": "yes",
"harness_retriever_pred": "yes",
"match": true,
"sum_p": 1.0
},
{
"id": "tabfact-10626",
"N": 2,
"picker_union_argmax": "yes",
"harness_retriever_pred": "yes",
"match": true,
"sum_p": 1.0
},
{
"id": "banking77-train-17",
"N": 50,
"picker_union_argmax": "card_delivery_estimate",
"harness_retriever_pred": "card_delivery_estimate",
"match": true,
"sum_p": 1.0
},
{
"id": "banking77-train-23",
"N": 50,
"picker_union_argmax": "card_arrival",
"harness_retriever_pred": "card_arrival",
"match": true,
"sum_p": 1.0
}
],
"latency_small_bf16": {
"n": 50,
"stage1": "union",
"p50_ms": 541.8913,
"p95_ms": 3846.8851,
"mean_ms": 968.4763,
"mean_prompt_tokens": 942.16
},
"latency_highN_bf16": {
"n": 20,
"stage1": "retriever",
"p50_ms": 182.12849099654704,
"p95_ms": 199.31893999455497,
"mean_ms": 183.2959999483137,
"mean_prompt_tokens": 315.95
}
},
"utod-picker-12b-v1-armC": {
"temps": {
"T1": 5.4679,
"T2": 1.267,
"T1_ret": 0.05
},
"load_test": [
{
"id": "gpqa-94",
"N": 4,
"picker_union_argmax": "canthaxanthin_biosynthesis",
"harness_retriever_pred": "canthaxanthin_biosynthesis",
"match": true,
"sum_p": 1.0
},
{
"id": "glaive-98691",
"N": 2,
"picker_union_argmax": "yes",
"harness_retriever_pred": "yes",
"match": true,
"sum_p": 1.0
},
{
"id": "tabfact-10626",
"N": 2,
"picker_union_argmax": "yes",
"harness_retriever_pred": "yes",
"match": true,
"sum_p": 1.0
},
{
"id": "banking77-train-17",
"N": 50,
"picker_union_argmax": "card_delivery_estimate",
"harness_retriever_pred": "card_delivery_estimate",
"match": true,
"sum_p": 1.0
},
{
"id": "banking77-train-23",
"N": 50,
"picker_union_argmax": "card_arrival",
"harness_retriever_pred": "card_arrival",
"match": true,
"sum_p": 1.0
}
],
"latency_small_bf16": {
"n": 50,
"stage1": "union",
"p50_ms": 543.0618170066737,
"p95_ms": 3847.9274609999266,
"mean_ms": 970.9316438387032,
"mean_prompt_tokens": 942.16
},
"latency_highN_bf16": {
"n": 20,
"stage1": "retriever",
"p50_ms": 181.16832099622115,
"p95_ms": 198.3372819959186,
"mean_ms": 182.499660000758,
"mean_prompt_tokens": 315.95
},
"latency_small_8bit": {
"n": 50,
"p50_ms": 537.7577670005849,
"p95_ms": 2232.4729199899593,
"mean_ms": 758.5707962393644,
"mean_prompt_tokens": 942.16
},
"eightbit_vs_bf16": {
"argmax_agreement": 0.92,
"mean_abs_dtopp": 0.028711071880883027,
"n": 50
},
"eightbit_note": "8-bit attn=eager (chunked_eager compute-identical for short prompts); stage1=retriever letter-read path"
}
}
}