github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-08 22:48:25 -05:00
Code Issues Packages Projects Releases Wiki Activity

bert lower learning rate and total steps (#9466)

Browse Source 
closer to the other submission with BS=240. converged with 10% less epochs
This commit is contained in:
chenyu
2025-03-16 17:21:20 -04:00
committed by GitHub
parent 09e7708b49
commit d2cfbd8a4d
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
examples/mlperf/model_train.py
View File

@@ -658,9 +658,9 @@ 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", 0.0002 * math.sqrt(BS/96))
max_lr = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.00018 * math.sqrt(BS/96))
train_steps = config["TRAIN_STEPS"] = getenv("TRAIN_STEPS", 3630000 // BS)
train_steps = config["TRAIN_STEPS"] = getenv("TRAIN_STEPS", 3300000 // BS)
warmup_steps = config["NUM_WARMUP_STEPS"] = getenv("NUM_WARMUP_STEPS", 1)
max_eval_steps = config["MAX_EVAL_STEPS"] = getenv("MAX_EVAL_STEPS", (10000 + EVAL_BS - 1) // EVAL_BS) # EVAL_BS * MAX_EVAL_STEPS >= 10000
eval_step_freq = config["EVAL_STEP_FREQ"] = getenv("EVAL_STEP_FREQ", int((math.floor(0.05 * (230.23 * BS + 3000000) / 25000) * 25000) / BS)) # Round down