






















A complete rewrite. Zero third-party dependencies. No PyTorch, no Triton, no transformers, no bitsandbytes. Pure-Python kernel compiler emits CUDA C, HIP C, and Metal Shading Language directly.
Install size: ~3 GB → ~5 MB.
| Metric | ZSE | vLLM | Δ |
|---|---|---|---|
| Cold start | 6.29s | 127.02s | 20.2× |
| VRAM used | 12.28 GB | 71.45 GB | 5.82× less |
| Single-seq tok/s | 37.0 | 26.5 | 1.40× |
| GPU | Cold start | vs vLLM AWQ INT4 cold |
|---|---|---|
| NVIDIA T4 (sm_75) | 7.25s | 30.2× faster |
| NVIDIA L4 (sm_89) | 5.58s | 26.0× faster |
| NVIDIA A10G (sm_86) | 6.01s | 32.1× faster |
| NVIDIA A100-80GB | 6.29s | 20.2× faster |
| AMD MI300X | 3.14s | 13.6× faster (vs vLLM-ROCm FP16) |
| Apple M1 | E2E vector_add validated, full inference pending |
pip install zse-engine zse serve model.zse --port 8000
Or run the kernel compiler standalone:
@zse.kernel Python DSL → CUDA / HIP / Metal. Warp primitives, vectorized memory, block reductions, tiling, fusion, WMMA, CDNA3 MFMA matrix cores, auto-tuning./v1/rag/*), web dashboard.zllm-zse → zse-engine on PyPIzse → zse_enginezse convertbnb / bitsandbytes backend removedFull migration guide and detailed change log: CHANGELOG.md
AMD MI300X validation, 32B-parameter benchmarks, and our ROCm wave-64 kernel development were made possible by DigitalOcean's Open Source Sponsorship Program.
447 tests passing. Zero dependencies. Three GPU backends. One package.
此内容由惯性聚合(RSS阅读器)自动聚合整理,仅供阅读参考。 原文来自 — 版权归原作者所有。