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-guest-rollout-ingress 分支比较 2 提交 »f22eab01fc Add lease-scoped guest rollout access812ac831d7 Gate leases on worker readiness
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-graph-rebuild 分支比较 3 提交 »5e940ac0a0 Format1813f4d296 Rebuild captured CUDA graphs in fork clone workers so graph replay survives the fork48c9c8411b Replay unacknowledged async CUDA ops after a fork-clone reconnect, and surface async failures at synchronize
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-graph-capture-investigation 分支比较 10 提交 »d9ee298d6f Record transactional batch fork validationece64db0ef Cover allocator policy across persistent exec256bf0100f Qualify automatic allocator across training workloads11bdfc6b7a Validate automatic fork-pool allocator policy33fffe73b9 Qualify pooled GRPO comparisons
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-golden-eviction 分支a6ec029016 Add best-effort CUDA auto-graph execution (#776)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-forkpoint-restored-wait 分支decb41564f Delay forkpoint sleeps until clone restore
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-fork-raw-handles 分支比较 10 提交 »9e4a732696 Add optional CUDA-over-network TCP transport and RPC/RTT instrumentation2040fe7656 Fix graph-mode fork isolation, make it correct-or-loud, enable by defaulta34c5713d9 Path 2 graph-isolation WIP: v2 kernel-node params, self-contained kernelParams rebuild, in-place-patch + buffer-breakdown diagnostics (still gated off; bug is invariant to patch approach and full-copy — root cause not yet found)7441d863d1 Experimental Path 2 graph-mode fork isolation (gated, WIP): re-instantiate an isolated clone's inherited cudagraph and translate its kernel/memset/memcpy node pointers to private copies; off by default (SMOLVM_CUDA_FORK_GRAPH_PATCH), default still fails loud1b818b08d2 Fail loud when an isolated fork clone launches an inherited CUDA graph (baked-in pointers would corrupt across clones) instead of silently producing garbage
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-fork-fd-reclaim 分支2c65b0f45f CUDA fork: close the daemon's copies of exported-physical fds after clone-worker spawn so a torn-down golden's VRAM is reclaimed
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-fork-fail-fast 分支比较 2 提交 »8537976246 CUDA fork: reject a clone connection whose lineage is unknown instead of serving it on the shared path it can't survive6d458a8cac CUDA fork: shut down the peer socket when one side of the daemon proxy ends so a dead clone worker fails the guest loudly instead of hanging it
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-fork-env-names 分支cdb18f8e36 Rename the CUDA fork env vars to describe behavior: SMOLVM_CUDA_FORK_WORKERS and SMOLVM_CUDA_FORK_SHARE_WEIGHTS
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-fork-enomem-retry 分支331ba0ac05 Launch CUDA clone workers without copying daemon memory (#818)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 cuda-followups 分支比较 9 提交 »479c333aa6 Reload clone-worker modules byte-identical to the golden's images45f59c5860 Add --share-weights to machine fork so sibling clones keep one copy of the base model in VRAMc8bf2cd896 Hand off a golden's cudaMalloc'd memory to fork-clone workerse93245e19d Add the new cuda/cmd/entrypoint fields to API test-struct literals0eec648c90 Accept a workload command on the machine-create API