11月12月1月2月3月4月5月6月7月8月9月10月
周一周三周五
一年内 257748 次贡献
更少更多
- frostbyte_neo 推送了仓库 github-featured/onnx--onnx 的 copilot/add-spacetodepth-mode-attribute 分支比较 350 提交 »f50b9b8c15 Merge remote-tracking branch 'origin/main' into copilot/add-spacetodepth-mode-attributefabf4590f4 Centralize SpaceToDepth documentation9b53930add Add SpaceToDepth mode in opset 2812673ebcb3 fix(inliner): handle Constant nodes without outputs (#8325)6c0785fc03 fix(shape_inference): prevent use-after-scope during node inference (#8330)
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 unified-decode-path 分支比较 5 提交 »63bdd67ac5 fix(runtime): V4 draft extend-round refresh must stay plain 1-token rows10577b62e8 fix(runtime): route the V4 draft's extend-round decode init through refresh194fbae30e fix(sampling): pin the greedy seed so tie-breaking is submission-stable59046f8aa0 test: migrate Blackwell-gated decode call sites to refresh_decode_metadata506073f6fe refactor(runtime): unify eager/graph decode and spec/non-spec sampling paths
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 ts2/tail-fusion-token-cap 分支846699bc85 perf(k3): end the fused MoE tail at its profit edge, not its capacity
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 fix/dsv4-flashinfer-precomputed-topk 分支比较 3 提交 »e43154902c refactor(moe): simplify FlashInfer routed dispatche7c440a109 chore: limit DSV4 routing change scope4cdcd79aee fix(moe): support precomputed FlashInfer MXFP4 routes
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 shared/glm-5.3-flash 分支比较 31 提交 »76b6295962 fix(ci): extend GLM-5.3-Flash AIME reasoning budgetbca62c814f fix(ci): fetch GLM-5.3-Flash tokenizer from Hugging Facea415cab73e fix(glm-5.3-flash): address review feedback8f2b7f1f7b fix(ci): cache FlashInfer JIT wheel downloads9bb6068290 ci(glm-5.3-flash): use public model repositories
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 main 分支比较 4 提交 »c006baf2dc feat(npu): support qwen3-0.6B on ascend (#1281)be0c422e5f Upgrade TokenSpeed Scheduler to 0.1.10 (#1298)4c07673100 Update TokenSpeed Scheduler to 0.1.10 (#1297)765013fa6a ci(amd): bump rocJITsu to ROCm/rocm-systems@79f87b82 (#1295)
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 feat/kda-tp-phase0 分支比较 2 提交 »b0516a2501 refactor(config): split AttnConfig into a model-wide container and component specsbb6230f308 refactor(parallel): give linear-attention layers their own parallel mapping
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 bot/slurm-flashinfer-0.6.18 分支比较 8 提交 »0b98000f49 Apply contract test formattingde672834f2 Adapt kernels to CUTLASS DSL 4.7.1723f4854b5 Adapt relative attention to latest FA4 schedulerd838c7d355 Adapt NVFP4 autotuning to FlashInfer 0.6.184d39c0ed6c Merge main into bot/slurm-flashinfer-0.6.18
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 sdk-portable-checkpoint-api 分支比较 2 提交 »be936ddfb8 Expose checkpoint restore and batch fork primitivesd0091f29ef Harden recursive portable checkpoints
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 vnc-zrle 分支比较 5 提交 »6e46c0781d Decode ZRLE in the browser client so browser sessions stop paying Raw bandwidthbab460fe5a Compress framebuffer updates with ZRLE for clients that advertise itcba995ea5c Read VNC client input on its own thread so a still screen cannot stall the keyboard and mouse8c1e7a111b Keep the VNC session alive when a client announces its clipboard or asks to resize the desktop (#1085)71eca574e2 Report why a container failed to create and stop the Vulkan injection shadowing image libraries (#1084)
- frostbyte_neo 推送了仓库 github-featured/smol-machines--smolvm 的 main 分支642b5dd41b Make checkpoint rollback roots safely forkable (#1087)
- frostbyte_neo 推送了仓库 github-featured/78--xiaozhi-esp32 的 main 分支527604dbef fix(otto): detect camera board from GPIO straps (#2219)
- frostbyte_neo 推送了仓库 github-featured/triggerdotdev--trigger.dev 的 queues-table-filter-bar 分支d4b7ad6919 fix(webapp): move Queues search and pagination above the table
