lower bert learning rate (tinygrad#9481)

slightly better. first sub 3hr run https://wandb.ai/chenyuxyz/MLPerf-BERT/runs/0or96ink/overview

Author: chenyuxyz

examples/mlperf/model_train.py

@@ -658,7 +658,7 @@ def train_bert():
   # ** hyperparameters **
   BS                 = config["GLOBAL_BATCH_SIZE"]      = getenv("BS", 11 * len(GPUS) if dtypes.default_float in (dtypes.float16, dtypes.bfloat16) else 8 * len(GPUS))
   EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 1 * len(GPUS))
-  max_lr             = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.00018 * math.sqrt(BS/96))
+  max_lr             = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.000175 * math.sqrt(BS/96))
 
   train_steps        = config["TRAIN_STEPS"]            = getenv("TRAIN_STEPS", 3300000 // BS)
   warmup_steps       = config["NUM_WARMUP_STEPS"]       = getenv("NUM_WARMUP_STEPS", 1)