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 263619 次贡献
更少更多
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-tp-joint-moe 分支比较 23 提交 »9ddf36898f perf(kimi3): join TP MoE reductions43cd6034a4 perf(comm): extend fused AttnRes through M1610863f1956 perf(comm): add two-stage producer-direct reduction377b28925f perf(amd): tune TP-local A16W4 decode205030c641 perf(moe): rebalance multi-token decode work
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-small-m-moe 分支比较 20 提交 »377b28925f perf(amd): tune TP-local A16W4 decode205030c641 perf(moe): rebalance multi-token decode workc6facd848f perf(kimi-k3): fuse the MoE path through four tokens2b1d76a434 perf(moe): tune small-batch router wavesbcebe08a43 feat(moe): select gfx950 TP A8W4 SiTU
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-kda-decode 分支比较 10 提交 »7759d332ae perf(kda): consume strided decode projections978ed2cfdc perf(cache): sparsify and budget Kimi-K3 state cache (#1147)53acfa2608 ci: use explicit GB300 runner labels (#1151)94f4d07cf6 perf(kimi3): avoid packed KDA QKV copies (#1129)cb1b5ed9fd perf(kda): forward host cu_seqlens to the CuteDSL prefill wrapper (#1120)
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-situ-foundation 分支比较 14 提交 »2c32487d76 feat(amd): prepare gfx950 SiTU prefill3a351c31d7 perf(gemm): tune Kimi latent projection by batchf2a912b54e perf(kimi3): use Gluon projection for small batchesf36484c664 perf(gemm): extend Kimi AttnRes projection to M47759d332ae perf(kda): consume strided decode projections
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-sharded-finalize 分支比较 24 提交 »ab93343c54 perf(kimi3): shard the TP MoE final projection9ddf36898f perf(kimi3): join TP MoE reductions43cd6034a4 perf(comm): extend fused AttnRes through M1610863f1956 perf(comm): add two-stage producer-direct reduction377b28925f perf(amd): tune TP-local A16W4 decode
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-latent-projection 分支比较 13 提交 »3a351c31d7 perf(gemm): tune Kimi latent projection by batchf2a912b54e perf(kimi3): use Gluon projection for small batchesf36484c664 perf(gemm): extend Kimi AttnRes projection to M47759d332ae perf(kda): consume strided decode projections978ed2cfdc perf(cache): sparsify and budget Kimi-K3 state cache (#1147)
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-attnres-runtime 分支比较 25 提交 »d5fee37923 perf(kimi3): enable batched AttnRes runtime fusionab93343c54 perf(kimi3): shard the TP MoE final projection9ddf36898f perf(kimi3): join TP MoE reductions43cd6034a4 perf(comm): extend fused AttnRes through M1610863f1956 perf(comm): add two-stage producer-direct reduction
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-a8w4-core 分支比较 15 提交 »07fded4624 feat(amd): add gfx950 A8W4 SiTU core2c32487d76 feat(amd): prepare gfx950 SiTU prefill3a351c31d7 perf(gemm): tune Kimi latent projection by batchf2a912b54e perf(kimi3): use Gluon projection for small batchesf36484c664 perf(gemm): extend Kimi AttnRes projection to M4
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-a8w4-select 分支比较 16 提交 »bcebe08a43 feat(moe): select gfx950 TP A8W4 SiTU07fded4624 feat(amd): add gfx950 A8W4 SiTU core2c32487d76 feat(amd): prepare gfx950 SiTU prefill3a351c31d7 perf(gemm): tune Kimi latent projection by batchf2a912b54e perf(kimi3): use Gluon projection for small batches
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-attnres-m4 分支比较 11 提交 »f36484c664 perf(gemm): extend Kimi AttnRes projection to M47759d332ae perf(kda): consume strided decode projections978ed2cfdc perf(cache): sparsify and budget Kimi-K3 state cache (#1147)53acfa2608 ci: use explicit GB300 runner labels (#1151)94f4d07cf6 perf(kimi3): avoid packed KDA QKV copies (#1129)
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-iris-attnres 分支比较 22 提交 »43cd6034a4 perf(comm): extend fused AttnRes through M1610863f1956 perf(comm): add two-stage producer-direct reduction377b28925f perf(amd): tune TP-local A16W4 decode205030c641 perf(moe): rebalance multi-token decode workc6facd848f perf(kimi-k3): fuse the MoE path through four tokens
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 sanket/wip-k3-iris-producer-direct 分支比较 21 提交 »10863f1956 perf(comm): add two-stage producer-direct reduction377b28925f perf(amd): tune TP-local A16W4 decode205030c641 perf(moe): rebalance multi-token decode workc6facd848f perf(kimi-k3): fuse the MoE path through four tokens2b1d76a434 perf(moe): tune small-batch router waves
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 borontion/amd-dsv4-1 分支比较 10 提交 »5f4aa800f8 perf(moe): use direct precomputed MXFP4 decodecba4153da6 feat(moe): support caller-provided MXFP4 outputs3b99ae680e feat(fp8): support caller-provided linear outputs5a4c9fde59 perf(v4): fuse paged attention metadata12846bc10b perf(v4): reuse precomputed routing in MoE
- frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 main 分支比较 11 提交 »1832f62a06 Bump TokenSpeed Scheduler to 0.1.8 (#1163)fdf3aae748 feat(kda): eager replay commit for speculative target verify (#1156)dc1d49b6da refactor(kda): disallow None host extend lengths — raise instead of silent fallback (#1161)652151d1f6 ci: add GB300 two-node AIME26 gates for K3-NVFP4 and K3-NVFP4+DSpark (#1159)9196fa2762 ci: cancel stale GB300 main runs (#1160)
