frostbyte_neo
0 关注者 · 0 关注中
  • 加入于 2026-07-22
仓库列表
1
 项目
1
 软件包 代码 公开活动 已点赞
66b1d307a4 fix(gfx950): preserve zero MX scale payloads
2b849a059a perf(gfx950): quantize MXFP4/MXFP8 tiles with scaled_downcast
比较 2 提交 »
2026-09-08 17:23:40 +00:00
2026-09-08 17:23:40 +00:00
fe4a2eeabc fix(cache): namespace L3 objects by runtime KV epoch
fba21f80b9 test(cache): require L3 test helper geometry arguments
fa113f491e Merge main and preserve L3 admission retries with declared cache geometry
dd8237039a refactor(cache): bind attention layers to cache groups from the plan (#1415)
651a87a7a2 fix(cache): fingerprint sharded-state patterns and npcache files
比较 7 提交 »
2026-09-08 17:23:40 +00:00
2026-09-08 17:23:40 +00:00
b82dafac61 Update tokenspeed-kernel MLA to 0.2.7 (#1438)
2cb248a090 Update tokenspeed-mla to 0.2.7 (#1437)
ff88072dba perf(dflash2): Optimize dflash2 implementation (#1399)
dd8237039a refactor(cache): bind attention layers to cache groups from the plan (#1415)
比较 4 提交 »
2026-09-08 17:23:39 +00:00
e1639af396 perf(comm): tune TP4 Iris decode launches
f2b2b996d2 perf(mhc): tune GLM-5.3-Flash reduction grid
e45eea09dc perf(mhc): tune GLM-5.3-Flash prenorm projection
60ed287334 perf(amd): reuse sigmoid routing kernels across row counts
d966a5a373 perf(amd): fill small-route expert blocks directly
比较 50 提交 »
2026-09-08 17:23:39 +00:00
f2b2b996d2 perf(mhc): tune GLM-5.3-Flash reduction grid
e45eea09dc perf(mhc): tune GLM-5.3-Flash prenorm projection
60ed287334 perf(amd): reuse sigmoid routing kernels across row counts
d966a5a373 perf(amd): fill small-route expert blocks directly
8451363234 perf(amd): tune GFX950 BF16 MoE reduction tiles
比较 49 提交 »
2026-09-08 17:23:38 +00:00
60ed287334 perf(amd): reuse sigmoid routing kernels across row counts
d966a5a373 perf(amd): fill small-route expert blocks directly
8451363234 perf(amd): tune GFX950 BF16 MoE reduction tiles
88091075b7 perf(dsa): reuse slot-conversion kernels across workspace sizes
ae7791fbe3 perf(amd): distribute DSA matrix work across columns
比较 47 提交 »
2026-09-08 17:23:38 +00:00
88091075b7 perf(dsa): reuse slot-conversion kernels across workspace sizes
ae7791fbe3 perf(amd): distribute DSA matrix work across columns
a2bb8e759d perf(amd): split BF16 DSA pooled selection widths
308c02da2c perf(cache): reuse MLA prefill writes across row alignment
9f000f9dab perf(kda): reuse prefill convolution across batch sizes
比较 44 提交 »
2026-09-08 17:23:38 +00:00
308c02da2c perf(cache): reuse MLA prefill writes across row alignment
9f000f9dab perf(kda): reuse prefill convolution across batch sizes
f93ae08427 perf(kda): consume verifier qkv as packed views
79578d1c67 perf(kda): reuse verifier scratch base rows
896e99d387 perf(kda): read verifier state from committed pages
比较 41 提交 »
2026-09-08 17:23:37 +00:00
8d4381aa1b perf(kpool): reuse decode metadata kernels across tail batches
20d7621e7d perf(kpool): keep slot-expansion table bounds dynamic
abe10f33d3 perf(kpool): keep prefill-tail source rows dynamic
9fefa8c3d7 perf(kpool): fuse decode selection bookkeeping
e40dc329d8 test(glm53-flash): cover KPool decode context bound
比较 35 提交 »
2026-09-08 17:23:37 +00:00
9f000f9dab perf(kda): reuse prefill convolution across batch sizes
f93ae08427 perf(kda): consume verifier qkv as packed views
79578d1c67 perf(kda): reuse verifier scratch base rows
896e99d387 perf(kda): read verifier state from committed pages
2849be78c6 perf(kda): tune GLM-5.3-Flash MTP verify launch
比较 40 提交 »
2026-09-08 17:23:37 +00:00
3fa8741691 chore(aihot): sync skill to v1.6.0 (auto from live aihot)
2026-09-08 17:23:21 +00:00
645d62417c refactor(templates): remove legacy disk option aliases
ad6957e220 docs(api): mark retired free disk field as ignored
9c8d372d3f feat(templates): expose configurable minimum free disk
578aaf90dc fix(cli): validate free disk option input
85d8e95213 feat(templates): expose configurable free disk space
比较 5 提交 »
2026-09-08 17:21:22 +00:00
2026-09-08 17:21:22 +00:00
2026-09-08 17:21:21 +00:00
af18843d88 fix(sdks): bound retries without request timeout
bc53da69fc fix(python): preserve retry phase timeouts
943e038ced fix(sdks): address retry review feedback
211b42a03a refactor(python): align retry module naming
d6d6fe9a30 refactor(sdks): centralize default retry count
比较 7 提交 »
2026-09-08 17:21:21 +00:00
2026-09-08 17:21:20 +00:00
395219a809 docs(desktop-python): fix docstring placement and stale parameter docs (#1806)
2026-09-08 17:21:20 +00:00
frostbyte_neo 推送了仓库 github-featured/e2b-dev--e2b 的 main 分支
b8b43232c8 feat(cli): honor NO_UPDATE_NOTIFIER to skip the update check (#1837)
6b759bf762 feat(sdk): add ServiceBusyError for 503 and expose the HTTP status on API errors (#1832)
0d7d628972 chore(desktop-python): refresh uv.lock for e2b 2.47.0 (#1829)
2a0b1d1bec ci: publish Python packages to PyPI via trusted publishing (#1828)
58c81f1019 fix: reject unrecognized onTimeout and onResume values (#1822)
比较 6 提交 »
2026-09-08 17:21:19 +00:00