AI Workloads

Pantheon includes controlled GPU stress tests for modern AI execution paths. They measure synthetic diagnostic work, not model quality or end-to-end serving performance.

Inference

pantheon --test inference --duration 60 --gpu 0 --mem 50 --verify --profile

Workload	Focus
`llm_decode`	Synthetic dependent KV-cache gather and projection pressure.
`llm_prefill`	Synthetic long-context causal-scan and projection pressure.
`kv_cache_churn`	Synthetic paged/ragged cache-update pressure.
`fused_attention`	Attention-style compute and memory proxy. It is not a fused-attention kernel.
`rope_stress`	RoPE-style rotary-position math proxy.
`quantized_gemm`	Packed low-precision projection proxy. It is not a full GEMM benchmark.
`serving_mix`	Mixed-request pressure proxy. It does not model a production scheduler or queue.
`speculative_decode`	Draft/verify execution proxy. It does not run draft and target models.
`moe_router`	Sparse Top-K routing proxy. It does not perform expert all-to-all communication.

Training, runtime, and auxiliary AI

Workload	Command	Focus
`transformer_train_step`	`pantheon --test training --duration 60 --gpu 0 --mem 50 --verify --profile`	Transformer forward/backward and optimizer-style proxy.
`allocation_fragmentation`	`pantheon --test allocation_fragmentation --duration 60 --gpu 0 --mem 25 --verify --profile`	Runtime allocation-pressure proxy. It does not yet execute real allocation/free fragmentation cycles.
`graph_replay`	`pantheon --test graph_replay --duration 60 --gpu 0 --mem 25 --verify --profile`	Actual CUDA Graph or HIP Graph capture, instantiation, and replay test.
`rag_embedding`	`pantheon --test ai_auxiliary --duration 60 --gpu 0 --mem 50 --verify --profile`	RAG embedding projection proxy.
`vision_encoder`	`pantheon --test ai_auxiliary --duration 60 --gpu 0 --mem 50 --verify --profile`	Vision-encoder projection proxy.

--mem selects the percentage of free VRAM used by the synthetic working set. Start at 25% if the GPU has other active workloads.

Note

Use Pantheon to isolate hardware and runtime behavior. Measure production throughput, queueing, time to first token, inter-token latency, and model accuracy with the actual serving framework and model separately.