Skip to content

Pull requests: hw-native-sys/pypto-lib

New pull request New
39 Open 963 Closed
39 Open 963 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Add: 512-row segment attention and 8192-token CP8 for DSv4 Flash MTP
#1121 opened Sep 3, 2026 by MaxwellF1 Contributor Loading…
5
fix: Lower the A8W8 decode layer against the requested platform
#1120 opened Sep 3, 2026 by Crystal-wzy Contributor Loading…
4 tasks done
1
Perf: parallelize DSpark HCA ratio-128 pooling
#1119 opened Sep 3, 2026 by wangqin1723-max Collaborator Loading…
1
feat(deepseek_v4_pro): expert W1/W3 matmul_mx for shared and routed experts
#1109 opened Sep 1, 2026 by yanghaoran29 Contributor Loading…
2 of 3 tasks
15
Refactor: replace the DSpark defer_wait sites with a blocking wait
#1105 opened Sep 1, 2026 by zhangqi-chen Collaborator Loading…
3
Validate the DSpark Prefill-to-Decode token loop
#1092 opened Aug 31, 2026 by hashiqiqixian Contributor Loading…
3
Refactor: fold DeepSeek-V4 Flash MTP split-K peels into matmul_acc init_cond
#1087 opened Aug 31, 2026 by Hzfengsy Member Loading…
8
Add DSpark target hidden outputs to target forwards
#1084 opened Aug 30, 2026 by hashiqiqixian Contributor Loading…
1
1
Add: fused int8 conv2d with L0C-to-L1 VREQ8
#1083 opened Aug 29, 2026 by tangalala Loading…
1
Perf: pipeline HCA AllToAll receive into O-A
#1059 opened Aug 27, 2026 by wangqin1723-max Collaborator • Draft
2
Perf: re-land the fused DeepSeek-V4 MoE combine scatter with disjoint-write opt-outs
#1053 opened Aug 26, 2026 by Hzfengsy Member • Draft
1
Refactor: fold DeepSeek-V4 Flash MTP split-K peels into matmul_acc init_cond
#1048 opened Aug 26, 2026 by Hzfengsy Member Loading…
12
Perf: coarsen HCA stream merge workers
#1038 opened Aug 25, 2026 by wangqin1723-max Collaborator Loading…
1
Perf: fuse DeepSeek V4 SWA and HCA O projection reduce-scatter
#1025 opened Aug 24, 2026 by wangqin1723-max Collaborator Loading…
3
Add: dump DeepSeek V4 MoE expert load statistics
#1016 opened Aug 23, 2026 by high-cloud Collaborator • Draft
3
Fix: fence DeepSeek V4 MoE window reuse
#948 opened Aug 12, 2026 by high-cloud Collaborator • Draft
1
1
Add: replace Qwen3 decode paged attention with PyPTO
#933 opened Aug 11, 2026 by Little-oil Contributor Loading…
18
Fix deterministic DeepSeek split-K reductions
#927 opened Aug 10, 2026 by high-cloud Collaborator • Draft
1
Add: experimental shared-expert weight TP on fixed SP
#913 opened Aug 7, 2026 by wangqin1723-max Collaborator Loading…
4
Add: use MXFP8 and MXFP4 in DeepSeek V4 Pro
#911 opened Aug 6, 2026 by yanghaoran29 Contributor • Draft
1
EXPERIMENT (do not merge): A/B the A5 daily sweep against ptoas op fusion wontfix This will not be worked on
#898 opened Aug 5, 2026 by lwDavid Collaborator • Draft
@lwDavid
1
Perf: assemble SWA decode window in L1, drop the GM gather
#882 opened Jul 31, 2026 by Hzfengsy Member Loading…
5
Fix(deepseek-v4 pro): let MoE post_ffn use automatic dependency tracking
#879 opened Jul 31, 2026 by yanghaoran29 Contributor Loading…
2 of 3 tasks
1
Add: L3 communication-computation example
#869 opened Jul 30, 2026 by Shenggan Loading…
4
Test: baseline prefill MTP CI control
#861 opened Jul 30, 2026 by Little-oil Contributor • Draft
1
ProTip! Filter pull requests by the default branch with base:main.