[gspark-etri]

Background

onggi's claim (AOT-compiled variants explode in size and share most bytes) was measured on two NPU families only. #56 measures it on a GPU with TensorRT-LLM.

Changes

• docs/evaluation.md: a GPU section with the A100 measurements: variants, storage per method, the token check, onggi diff between variants and delta apply.

Verification

Measured by a lab run (#56) on the A100 node, which was uncordoned only for the run and cordoned again afterwards. The namespace and the 116 GB scratch data were removed.

• TensorRT-LLM 1.0.0, Llama-3.1-8B-Instruct bf16, TP 1. Four engines (max batch 2/8 × max seq 2k/4k), each 16.14 GB with the weights embedded.
• Four engines: 64.5 GB raw → 16.15 GB with CDC dedup → 10.85 GB in onggi (5.95×). Fixed 1 MiB blocks reach 48.4 GB.
• A restored engine is sha256-identical and gives the same 32 greedy tokens.
• Same max batch: 99.96 % same offset, and delta apply reads 1.2 MB. A different max batch shifts the weights by 17,952 B: 0 % same offset, so the pull reads 10.8 GB. Delta apply: copy bytes the base holds at other offsets from the base file (TensorRT-LLM batch change reads 10.8 GB) #58 tracks copying those bytes from the base.

Compatibility and limits

• Docs only.

Related Issue

Closes #56

🤖 Generated with Claude Code