- 27 Jan, 2026 1 commit
-
-
omkhalil authored
Fix UnembedMetrics to correctly count FLOPs for the unembedding (LM head) layer. The bug: UnembedMetrics used total_num_tokens() which counts all tokens in the batch for projection flops, vocab projections are run on just the last token for the autoregressive use case. Co-authored-by:Omar Mohamed Khalil <omarkhalil@meta.com>
-
- 18 Dec, 2025 1 commit
-
-
SungMinCho authored
Signed-off-by:
SungMinCho <tjdals4565@gmail.com> Signed-off-by:
Mark McLoughlin <markmc@redhat.com> Co-authored-by:
Mark McLoughlin <markmc@redhat.com>
-