Eliminate zero-ops (VIEW/RESHAPE/TRANSPOSE/PERMUTE) #27

Closed
opened 2026-04-30 18:11:33 +02:00 by sleepy · 0 comments
Owner

No code change needed. Zero-ops (VIEW/RESHAPE/TRANSPOSE/PERMUTE) are already filtered at ggml-metal-ops.cpp:59 before encoding — they return 1 and skip to next node without GPU dispatch. The 682 zero-ops/tick incur zero dispatch overhead.

No code change needed. Zero-ops (VIEW/RESHAPE/TRANSPOSE/PERMUTE) are already filtered at ggml-metal-ops.cpp:59 before encoding — they return 1 and skip to next node without GPU dispatch. The 682 zero-ops/tick incur zero dispatch overhead.
sleepy added the perf label 2026-04-30 18:11:33 +02:00
sleepy changed title from Eliminate zero-ops VIEW RESHAPE TRANSPOSE PERMUTE to Eliminate zero-ops (VIEW/RESHAPE/TRANSPOSE/PERMUTE) 2026-04-30 18:16:37 +02:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: sleepy/llama.cpp#27