Merge pull request #452 from denys-fridman/dfridman/moe_aux_loss_weight

Add moe_aux_loss_coeff constant and compliance checks for DeepSeek V3

Author: ShriyaRishab

mlperf_logging/compliance_checker/training_6.0.0/closed_deepseekv3_671b.yaml

@@ -69,6 +69,11 @@
     REQ:   EXACTLY_ONE
     CHECK: " v['value'] == 1.0 "
 
+- KEY:
+    NAME:  moe_aux_loss_coeff
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.01 "
+
 - KEY:
     NAME:  gradient_accumulation_steps
     REQ:   EXACTLY_ONE

mlperf_logging/compliance_checker/training_6.0.0/open_deepseekv3_671b.yaml

@@ -55,6 +55,10 @@
     NAME:  opt_gradient_clip_norm
     REQ:   EXACTLY_ONE
 
+- KEY:
+    NAME:  moe_aux_loss_coeff
+    REQ:   EXACTLY_ONE
+
 - KEY:
     NAME:  gradient_accumulation_steps
     REQ:   EXACTLY_ONE

mlperf_logging/mllog/constants.py

@@ -177,6 +177,7 @@
 START_WARMUP_STEP = "start_warmup_step"
 INIT_CHECKPOINT_STEP = "init_checkpoint_step"
 LORA_ALPHA = "lora_alpha"
+MOE_AUX_LOSS_COEFF = "moe_aux_loss_coeff"
 # Log keys - misc.
 BBOX = "bbox"
 SEGM = "segm"