12月1月2月3月4月5月6月7月8月9月10月11月
周一周三周五
一年内 264198 次贡献
更少更多
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-module-mmap-handoff 分支08a671abed Map CUDA module handoffs across clone workers
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-module-image-store 分支019df4a3ef Preserve two vCPUs for CUDA fork pools (#820)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-module-handoff-reuse 分支8435da37b7 Map CUDA module handoffs across clone workers (#809)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-managed-rollout 分支比较 2 提交 »ffa380f5c8 Expose rollout executor timeout configurationf4a02f60b1 Keep fused rollout cleanup retry-safe
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-lineage-module-preload 分支ece2429189 Preload prepared CUDA clone modules
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-mainline-promotion 分支比较 2 提交 »ffc811de3c Refresh bundled libkrun runtimesb0254dc95b Preserve CUDA fork-pool state and VMM isolation (#779)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-image-machine-polish 分支比较 2 提交 »d6da4ecc6c Bound the CUDA shim's bring-up read so cuInit fails in seconds when no host service answers, instead of hanging the first call forever61b4213ebd Run the workload command passed to machine create --from <artifact> instead of silently ignoring it
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-independent-serving 分支比较 2 提交 »6695001898 CUDA fork isolation: capture VMM (expandable_segments) allocations0b1333700f CUDA fork independent serving: copy-on-fork isolation, graph mode, network transport
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-host-snapshot-mmap 分支82c26d6ebe Wake automatic fork pools on state changes (#815)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-guest-shim 分支比较 10 提交 »e2c2eb4626 cuda: report unregistered host memory correctly so torch pin_memory reaches the zero-copy path2b9853e7e3 cuda: forward cuBLASLt, cuDNN v8 backend + batch-norm, and strided-batched GEMM for PyTorch8c499089bd fix(cuda): gate the shm data channel to Linux for cross-platform builds396d873d3b feat(cuda): auto-enable guest-RAM zero-copy for --cuda workloadsf026c9b3fd feat(cuda): forward-to-host-libs remoting with guest-RAM zero-copy