Repository navigation

Feat/pytorch vswa kvcachemanager - #5151

Merged

byshiue merged 15 commits into

NVIDIA:mainfrom

qixiang-99:feat/pytorch-vswa-kvcache-0611

Jul 2, 2025

Merged

Feat/pytorch vswa kvcachemanager#5151
byshiue merged 15 commits into
NVIDIA:mainfrom
qixiang-99:feat/pytorch-vswa-kvcache-0611

Commits on Jun 30, 2025
- Refactor and pipeclean, can pass gemma3 test. Need to run CI to find out whether there are more issues.
  qixiang-99
  committed
  68da08e
- fix:
  qixiang-99
  committed
  31847a8
- fix: correct the test that passes the kvcache config with the wrong type (the mirror class of executor.kvcache_config)
  qixiang-99
  committed
  3401c61
- fix:
  qixiang-99
  committed
  1ced7c1
- fix:
  qixiang-99
  committed
  2010813
- test: Add GSM8K evaluation to Gemma3_1BInstruct tests
  qixiang-99
  committed
  105acac
- fix:
  qixiang-99
  committed
  72612d0
- fix: address comments
  qixiang-99
  committed
  a94798b
- chores: small fix for rebasing, will remove this after #5384 merged
  qixiang-99
  committed
  83266bc
- when it is swa, we need to make sure chunk size is 256
  qixiang-99
  committed
  d150b29
- fix: Update the test and GSM8K reference score
  qixiang-99
  committed
  c153f69
Commits on Jul 1, 2025
- Remove `KvCacheConfig` conversion as #5384 merged.
  qixiang-99
  committed
  c5919f4
- WAR for the kernel diff between Hopper and Blackwell
  qixiang-99
  committed
  a0d6a7c
- fix: check whether max_attention_window is None
  qixiang-99
  committed
  2665c44
- Revert "WAR for the kernel diff between Hopper and Blackwell"
  qixiang-99
  committed
  c2e7b91