DeepSpeed
fix(rollout): pass a pre-built 4-D mask to the static graph decode forward
#9004
Merged

Commits
  • Pass a prebuilt 4-D mask to the static rollout CUDA graph
    anyadontfly committed 12 days ago
  • fix(rollout): reject flash_attention_2 on the static graph path and test real models
    anyadontfly committed 12 days ago
  • Merge master into graph-decode-4d-mask
    anyadontfly committed 11 days ago
Loading