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 253673 次贡献
更少更多
- frostbyte_neo 推送了仓库 github-featured/mvanhorn--last30days-skill 的 chore/release-3.18.0 分支ca30c86cd9 chore(release): bump version to 3.18.0
- frostbyte_neo 推送了仓库 github-featured/mvanhorn--last30days-skill 的 main 分支01aef34ca4 docs(solutions): compound the checkpointed host-judgment protocol learning (#858)
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 yechan/qwen35-mrope-opt 分支比较 10 提交 »965bf28434 fix(multimodal): remove stale model imports14b9612755 fix(multimodal): preserve Qwen3 Omni M-RoPE APIa0a538be2a perf(multimodal): keep text-only M-RoPE fast path9e392a9d2d refactor(multimodal): narrow M-RoPE optimization scopeb2c64019b3 perf(multimodal): reduce M-RoPE decode host overhead
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 zt/dev/ts_qwen35_gdn 分支比较 3 提交 »e47363dfb8 perf(gdn): update to gdn decode API to flashinfer 0.6.15360295496c fix: fix dtype error.71f75d4044 perf: use flashinfer decode gdn and ssm_states with k-last shape.
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 xiangyi/v4-mhc-fused-backend-select 分支比较 3 提交 »d5bc88f3b9 perf(deepseek-v4): extend two-stage fused mHC launch to small token countsa3b42c02e0 perf(deepseek-v4): tune big-fuse block size in fused mHC launch tablefad225396d perf(deepseek-v4): select two-stage fused mHC launch by token count
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 yechan/mm-shm-transport-opt 分支比较 6 提交 »7afff1105d fix(multimodal): order TP input broadcast on model stream1ad208f6e4 perf(multimodal): copy TP source SHM directly to GPU61bbde26f5 perf(multimodal): overlap SHM unlink with H2D95a0e32f20 perf(multimodal): broadcast large SHM inputs across TPeb9e0d05f1 perf(multimodal): overlap large SHM copies with H2D
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 xiangyi/v4-sparse-compress-warps 分支6caad4b1e1 perf(deepseek-v4): widen sparse compress cache launch for large ratios
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 yechan/qwen-vision-runtime-opt 分支比较 3 提交 »68e23cb628 refactor(multimodal): narrow Qwen vision runtime scopee885283705 perf(multimodal): optimize Qwen vision embedding assemblye1c383d315 perf(multimodal): optimize Qwen vision CUDA graph replay
