frostbyte_neo

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 test/kda-mtp-fp32-oracle 分支

84478e96d2 test(kda): vectorize MTP reference batch

ef9df14e69 test(kda): validate MTP schedules against FP32 recurrence

比较 2 提交 »

2026-09-10 00:35:15 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 test/kda-mtp-fp32-oracle

2026-09-10 00:35:15 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 perf/gfx1250-sigmoid-topk 分支

8e8fffde5b fix(moe): keep mixed-dtype gfx950 sigmoid top-k eligible

e7a83cb047 style: format gfx1250 sigmoid top-k registration

f00ea3ca0c perf(moe): add gfx1250 fused sigmoid top-k

比较 3 提交 »

2026-09-10 00:35:13 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 perf/gfx1250-sigmoid-topk

2026-09-10 00:35:13 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 perf/gfx1250-projected-mla

2026-09-10 00:35:12 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 perf/gfx1250-projected-mla 分支

9e827e828a fix(ci): update MI450 sim node id for projected MLA decode

be9dd1def5 style: format projected MLA files

ab4221df4d perf(amd): extend gfx1250 projected MLA decode

比较 3 提交 »

2026-09-10 00:35:12 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 perf/gfx1250-latent-input-m1 分支

ff22b1ade8 style: format gfx1250 latent-input decode

14d0f35685 perf(moe): port gfx1250 M=1 latent-input decode

比较 2 提交 »

2026-09-10 00:35:11 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 perf/gfx1250-latent-input-m1

2026-09-10 00:35:11 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 perf/gfx1250-kimi3-linear-attnres

2026-09-10 00:35:10 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 perf/gfx1250-kimi3-linear-attnres 分支

f8f6f2b54a fix(kimi3): tolerate missing gfx1250 AttnRes and snapshot graph replay

9439b5482d style: format gfx1250 linear AttnRes files

47cc483db8 perf(kimi3): port linear AttnRes partials to gfx1250

比较 3 提交 »

2026-09-10 00:35:10 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 efric/refactor/iris-buffer-sizing

2026-09-10 00:35:08 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 perf/gfx1250-kimi3-largem 分支

2026-09-10 00:35:08 +00:00

frostbyte_neo 于 github-featured/lightseekorg--tokenspeed 创建了分支 perf/gfx1250-kimi3-largem

2026-09-10 00:35:08 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 efric/refactor/iris-buffer-sizing 分支

2026-09-10 00:35:08 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 main 分支

0422a3932e perf(cache): reduce cache block zeroing overhead (#1453)

dbd89e78fc perf(amd): Optimize GLM-5.3-Flash DSA decode on MI350X (#1433)

f5e0b0e571 fix(kda): preserve NVIDIA MTP launch schedule (#1467)

e67a4033c4 perf(moe): speed up gfx1250 MXFP4 prefill routing (#1463)

4d601173de refactor(kernel): move moe tactics script under benchmark (#1464)

比较 7 提交 »

2026-09-10 00:35:07 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 efric/feat/k3-iris-preallocate 分支

b8be9fa658 docs(cache): remove Iris allocation note

edb9cc8f62 fix(kimi-k3): prepare reachable Iris states

c6834394a3 nit

b900924169 feat(kimi-k3): reserve Iris buffers before KV cache

f454577af1 fix(comm): preserve zero producer-direct capacity

比较 16 提交 »

2026-09-10 00:35:06 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 Max191/glm53-flash-stack-07-tp4-0902 分支

6e4dd1f6d2 perf(comm): tune TP4 Iris decode launches

b640457632 perf(mhc): tune GLM-5.3-Flash reduction grid

19bceca275 perf(mhc): tune GLM-5.3-Flash prenorm projection

c1835d3cff perf(amd): reuse sigmoid routing kernels across row counts

d566362d7c perf(amd): fill small-route expert blocks directly

比较 12 提交 »

2026-09-10 00:35:05 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 Max191/glm53-flash-stack-06-mhc-0902 分支

b640457632 perf(mhc): tune GLM-5.3-Flash reduction grid

19bceca275 perf(mhc): tune GLM-5.3-Flash prenorm projection

c1835d3cff perf(amd): reuse sigmoid routing kernels across row counts

d566362d7c perf(amd): fill small-route expert blocks directly

98fc0a58f6 perf(amd): tune GFX950 BF16 MoE reduction tiles

比较 11 提交 »

2026-09-10 00:35:05 +00:00

frostbyte_neo 推送了仓库 github-featured/lightseekorg--tokenspeed 的 Max191/glm53-flash-stack-05-moe-0902 分支

a4d2d9fc2f Merge branch 'main' into Max191/glm53-flash-stack-05-moe-0902

dbd89e78fc perf(amd): Optimize GLM-5.3-Flash DSA decode on MI350X (#1433)

f5e0b0e571 fix(kda): preserve NVIDIA MTP launch schedule (#1467)

e67a4033c4 perf(moe): speed up gfx1250 MXFP4 prefill routing (#1463)

c1835d3cff perf(amd): reuse sigmoid routing kernels across row counts

比较 13 提交 »

2026-09-10 00:35:04 +00:00

frostbyte_neo 推送了仓库 github-featured/mpmath--mpmath 的 master 分支

c74b37d377 Ensure acos(h)/asin(h) respect special cases from applicable standards (#1173)

2026-09-10 00:34:49 +00:00