You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I run an Megatron sft training(DP=8, PP=1, TP=1), but the speed is very slow. I tried profiling the training process and found that the matmul calls cutlass::Kernel<cutlass_75_tensorop_bf16_s1688gemm_bf16_256x128_tn_align1（batch_size=4， hidden_size=4096，sequence_length=4096）However if I make a copy and matmul their copies， the matmul call ampere_bf16_s16816gemm_bf16_128x256_ldg8_f2f_stages_64x3_tn？
This discussion was converted from issue #997 on August 21, 2024 18:34.
Heading
Bold
Italic
Quote
Code
Link
Numbered list
Unordered list
Task list
Attach files
Mention
Reference
Menu
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
I run an Megatron sft training(DP=8, PP=1, TP=1), but the speed is very slow. I tried profiling the training process and found that the matmul calls
cutlass::Kernel<cutlass_75_tensorop_bf16_s1688gemm_bf16_256x128_tn_align1（batch_size=4， hidden_size=4096，sequence_length=4096）However if I make a copy and matmul their copies， the matmul callampere_bf16_s16816gemm_bf16_128x256_ldg8_f2f_stages_64x3_tn？Beta Was this translation helpful? Give feedback.
All reactions