Skip to content

Chore: advance the pypto-lib submodule to b48da63 - #190

Closed
ndleslx wants to merge 1 commit into
hw-native-sys:mainfrom
ndleslx:chore/bump-pypto-lib
Closed

ndleslx wants to merge 1 commit into
hw-native-sys:mainfrom
ndleslx:chore/bump-pypto-lib

Conversation

@ndleslx

@ndleslx ndleslx commented Aug 25, 2026

Copy link
Copy Markdown
Collaborator

Advances the pypto-lib submodule from 2cd5f828 to
b48da63
(21 commits):

Perf

  • b48da63 align DeepSeek V4 CSA indexer scoring with baseline (#1030)
  • 9ba79cf retune dspark SWA decode task partitioning (#1028)
  • df3e8a6 fuse DeepSeek V4 decode O projection reduce-scatter (#998)
  • baec009 accelerate DeepSeek-V4 Flash token generation (#1014)
  • b3b9c44 replace DSpark greedy_sample sort with a vector max scan (#1005)

Add

  • 1d7c701 1M context support for DeepSeek V4 CSA decode (#1013)
  • efd5848 dspark prompt prefill and variable decode context (#999)
  • c7f700c dspark context-parallel prefill SWA, HCA and CSA (#1010)
  • b9b60da 1M context in DeepSeek V4 HCA decode (#1008)
  • 665665a dspark context-parallel prefill KV all-gather (#1006)

Refactor

  • 4be3f5e drop the MTP hc_pre syncall implementation and restyle (#1029)
  • cb43e09 defer DeepSeek V4 MTP communication waits (#1024)
  • 634e77b rename enable_l2_swimlane to enable_chip_swimlane (#1012)
  • b7accac resolve pto-isa from the pin, not PTO_ISA_ROOT (#1011)
  • 193819b inline the dspark decode output entries and route by --tp (#1003)

Fix / CI / Docs

  • 017cb44 cap nanobind below 3.0 in the CI setup (#1018)
  • ab4ecb1 unpin pypto and resolve the toolchain from HEAD again (#1002) —
    already mirrored on this side by Revert "CI: pin the pypto revision instead of tracking its HEAD" #178
  • 5d32927 / dd69c60 / bbc4093 Qwen3-14B optimization, DeepSeek V4 decode
    optimization, and distributed programming guides

The range moves the DeepSeek V4 / MTP / dspark paths this repo's guards
exercise — notably the MTP hc_pre syncall removal and comm-wait deferral, close
to the area #144 touched before its revert (#186). The Qwen3 and DeepSeek V4
accuracy guards run on this PR unchanged and are the gate for the bump.

@coderabbitai

coderabbitai Bot commented Aug 25, 2026

Copy link
Copy Markdown

Important

  • 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 20d6c276-8ca6-4597-8993-140f0921ba7c


Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

21 commits ahead of 2cd5f828: DeepSeek V4 CSA indexer scoring
alignment, dspark context-parallel prefill (KV all-gather, SWA/HCA/CSA,
prompt prefill), 1M context support for HCA and CSA decode, fused decode
O-projection reduce-scatter, MTP comm-wait deferral and the hc_pre
syncall removal, the nanobind <3.0 cap, and pypto-lib#1002's CI unpin
(already mirrored here by hw-native-sys#178).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant