phibench-mi210-rows / task_catalog.json
Erland's picture
Calibrate the whole-layer MoE row and keep executable bits
323e515
Raw History Blame Contribute Delete
3.7 kB
{
"schema": "fai_bench_task_catalog_v2",
"task_count": 3,
"by_bench": {
"lh": 3
},
"tasks": [
{
"bench": "lh",
"task_dir": "mi210-jais3-moe-m256",
"infra_name": "fused_moe_w8a16_m256",
"reward_class": "performance",
"gpus": 1,
"max_submissions": 16,
"ref_speedup": 1.201437,
"summary": "This is the routed-expert computation of one Mixture-of-Experts layer on one tensor-parallel rank of a vLLM server on AMD Instinct MI210, for a block-quantized FP8 model with BF16 activations. The implementation in moe_w8a16.py and its two tile tables is what the server runs today, and it is correct. You must make the whole layer call faster (token sort, the two expert matmuls, SiLU-and-mul and the top-k sum) for one batch size, 256 tokens, with routing drawn from expert counts recorded on a real server. Inputs are the hidden states, the FP8 weights with their block scales, and the top-k expert ids and weights; the output is this rank's partial sum. Metric: continuous performance score on that one batch size after a hard correctness gate over all batch sizes (error within 1.02x of the baseline's, deterministic, no compilation during inference), on a logarithmic curve relative to the oracle anchor, calibrated on MI210, not comparable with H20 anchors."
},
{
"bench": "lh",
"task_dir": "mi210-jais3-moe-m1024",
"infra_name": "fused_moe_w8a16_m1024",
"reward_class": "performance",
"gpus": 1,
"max_submissions": 16,
"ref_speedup": 1.249117,
"summary": "This is the routed-expert computation of one Mixture-of-Experts layer on one tensor-parallel rank of a vLLM server on AMD Instinct MI210, for a block-quantized FP8 model with BF16 activations. The implementation in moe_w8a16.py and its two tile tables is what the server runs today, and it is correct. You must make the whole layer call faster (token sort, the two expert matmuls, SiLU-and-mul and the top-k sum) for one batch size, 1024 tokens, with routing drawn from expert counts recorded on a real server. Inputs are the hidden states, the FP8 weights with their block scales, and the top-k expert ids and weights; the output is this rank's partial sum. Metric: continuous performance score on that one batch size after a hard correctness gate over all batch sizes (error within 1.02x of the baseline's, deterministic, no compilation during inference), on a logarithmic curve relative to the oracle anchor, calibrated on MI210, not comparable with H20 anchors."
},
{
"bench": "lh",
"task_dir": "mi210-jais3-routed-moe-w8a16",
"infra_name": "fused_moe_w8a16_routed",
"reward_class": "performance",
"gpus": 1,
"max_submissions": 16,
"ref_speedup": 1.132514,
"summary": "This is the routed-expert computation of one Mixture-of-Experts layer on one tensor-parallel rank of a vLLM server on AMD Instinct MI210, for a block-quantized FP8 model with BF16 activations. The implementation in moe_w8a16.py and its two tile tables is what the server runs today: already tuned, correct, and hard to improve. You must make the whole layer call faster (token sort, the two expert matmuls, SiLU-and-mul and the top-k sum) over token counts from 1 to 8192, with decode counts carrying most of the weight, replaying routing recorded on a real server. Inputs are the hidden states, the FP8 weights with their block scales, and the top-k expert ids and weights; the output is this rank's partial sum. Metric: continuous performance score after a hard correctness gate (error within 1.02x of the baseline's, deterministic, no compilation during inference), on a logarithmic curve relative to the oracle anchor, calibrated on MI210, not comparable with H20 anchors."
}
]
}