mirror of https://github.com/harvard-edge/cs249r_book.git synced 2026-07-16 06:07:17 -05:00

Files

Rocky c46ec4734d fix(mlperf-edu): anomaly-ae-train model-size gate impossible to pass (#1937 )

max_model_size_kb: 32 could never be satisfied by the reference model it
gates: AnomalyDetectionAE (640/784 -> 128x4 -> 8 -> 128x4 -> 640) has
~266K parameters -- matching this same file's own `params: 0.3M` -- which
is ~1.03MB at FP32 and ~260KB even fully INT8-quantized. A submitter who
trains exactly this reference model and reports the true model size would
automatically fail the size gate regardless of reconstruction quality.

Corrected the budget to 300KB, consistent with the file's own stated 0.3M
parameter count at 1 byte/param (INT8), the standard deployment target for
this suite's "microcontroller" framing.

2026-07-15 11:47:33 +02:00

suites

…

README.md

…

README.md

MLPerf EDU Native Registry

This directory is the native suite/workload/variant registry layout. Edit these files first. workloads.yaml is a compatibility mirror generated by tools/export_flat_registry.py. tools/export_registry_layout.py is retained as a migration helper; normal edits should happen in this native layout and then refresh the generated flat mirror.

Suite	Workload	Internal ID	Public status
`agent`	`nano-codegen-agent`	`nano-codegen-agent`	`systems-only`
`agent`	`nano-rag-agent`	`nano-rag-agent`	`systems-only`
`agent`	`nano-react-agent`	`nano-react-agent`	`systems-only`
`agent`	`nano-toolcall-agent`	`nano-toolcall-agent`	`systems-only`
`distributed`	`micro-dlrm-distributed`	`micro-dlrm-distributed`	`systems-only`
`graph`	`micro-gnn-train`	`micro-gnn-train`	`systems-only`
`language`	`micro-bert-train`	`micro-bert-train`	`systems-only`
`language`	`nano-lora-finetune`	`nano-lora-finetune`	`systems-only`
`language`	`nano-moe-train`	`nano-moe-train`	`systems-only`
`language`	`nanogpt-inference --variant decode`	`nanogpt-decode`	`performance-bearing`
`language`	`nanogpt-inference --variant fp16-b16`	`nanogpt-decode-fp16-b16`	`systems-only`
`language`	`nanogpt-inference --variant fp32-b16`	`nanogpt-decode-fp32-b16`	`systems-only`
`language`	`nanogpt-inference --variant prefill`	`nanogpt-prefill`	`performance-bearing`
`language`	`nanogpt-inference --variant speculative`	`nanogpt-decode-spec`	`systems-only`
`language`	`nanogpt-train`	`nanogpt-train`	`score-bearing`
`recommender`	`micro-dlrm-dram-train`	`micro-dlrm-dram-train`	`systems-only`
`recommender`	`micro-dlrm-train`	`micro-dlrm-train`	`score-bearing`
`rl`	`micro-rl-train`	`micro-rl-train`	`systems-only`
`slm`	`smollm2-chat-inference --variant baseline`	`slm-decode`	`performance-bearing`
`slm`	`smollm2-chat-inference --variant batched-b4`	`slm-batched-decode`	`systems-only`
`slm`	`smollm2-chat-inference --variant long-context`	`slm-long-context-decode`	`systems-only`
`slm`	`smollm2-chat-inference --variant quantized-int8`	`slm-quantized-decode`	`performance-bearing`
`timeseries`	`micro-lstm-train`	`micro-lstm-train`	`systems-only`
`tiny`	`anomaly-ae-train`	`anomaly-ae-train`	`score-bearing`
`tiny`	`dscnn-kws-train`	`dscnn-kws-train`	`systems-only`
`tiny`	`wake-vision-vww`	`wake-vision-vww`	`systems-only`
`vision`	`micro-diffusion-train`	`micro-diffusion-train`	`systems-only`
`vision`	`mobilenet-cifar100-composed-fp16`	`mobilenet-cifar100-composed-fp16`	`systems-only`
`vision`	`mobilenetv2-train`	`mobilenetv2-train`	`score-bearing`
`vision`	`resnet18-train`	`resnet18-train`	`score-bearing`