another try with specific finetune gin file

finetune_large_mt5_sentencefix_v4.gin

@@ -31,7 +31,7 @@ train_script.train:
 # `num_decodes` is equivalent to a beam size in a beam search decoding.
 models.EncoderDecoderModel.predict_batch_with_aux.num_decodes = 4
 
-partitioning.ModelBasedPjitPartitioner.num_partitions = 4
+partitioning.ModelBasedPjitPartitioner.num_partitions = 2
 
 
 #from t5.models import mesh_transformer

finetune_large_mt5_sentencefix_v4_16.gin

@@ -0,0 +1,40 @@
+from __gin__ import dynamic_registration
+import tasks_v4
+
+import __main__ as train_script
+from t5.data import mixtures
+from t5x import models
+from t5x import partitioning
+from t5x import utils
+
+include "t5x/examples/t5/mt5/large.gin"
+include "t5x/configs/runs/finetune.gin"
+
+MIXTURE_OR_TASK_NAME = "sentencefix"
+TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 256}
+TRAIN_STEPS = 1_200_000  # 1000000 pre-trained steps + 20000 fine-tuning steps.
+USE_CACHED_TASKS = False
+DROPOUT_RATE = 0.0
+RANDOM_SEED = 0
+
+# `LOSS_NORMALIZING_FACTOR`: When fine-tuning a model that was pre-trained
+# using Mesh Tensorflow (e.g. the public T5 / mT5 / ByT5 models), this should be
+# set to `pretraining batch_size` * `target_token_length`. For T5 and T5.1.1:
+# `2048 * 114`. For mT5: `1024 * 229`. For ByT5: `1024 * 189`.
+#LOSS_NORMALIZING_FACTOR = 234496
+INITIAL_CHECKPOINT_PATH = "gs://t5-data/pretrained_models/t5x/mt5_large/checkpoint_1000000"
+
+train_script.train:
+  eval_period = 500
+  partitioner = @partitioning.ModelBasedPjitPartitioner()
+
+# `num_decodes` is equivalent to a beam size in a beam search decoding.
+models.EncoderDecoderModel.predict_batch_with_aux.num_decodes = 4
+
+partitioning.PjitPartitioner.num_partitions = 4
+
+#from t5.models import mesh_transformer
+#import t5.models
+#mesh_transformer.learning_rate_schedules.constant_learning_rate.learning_rate = 0.0005
+#run.learning_rate_schedule = @learning_rate_schedules.constant_learning_rate
+

small_wmt_finetune.gin

@@ -0,0 +1,21 @@
+from __gin__ import dynamic_registration
+
+import __main__ as train_script
+from t5.data import mixtures
+from t5x import models
+from t5x import partitioning
+from t5x import utils
+
+include "t5x/examples/t5/t5_1_1/small.gin"
+include "t5x/configs/runs/finetune.gin"
+
+MIXTURE_OR_TASK_NAME = "wmt_t2t_ende_v003"
+TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 256}
+TRAIN_STEPS = 1_020_000  # 1000000 pre-trained steps + 20000 fine-tuning steps.
+DROPOUT_RATE = 0.0
+INITIAL_CHECKPOINT_PATH = "gs://t5-data/pretrained_models/t5x/t5_1_1_small/checkpoint_1000000"
+# `LOSS_NORMALIZING_FACTOR`: When fine-tuning a model that was pre-trained
+# using Mesh Tensorflow (e.g. the public T5 / mT5 / ByT5 models), this should be
+# set to `pretraining batch_size` * `target_token_length`. For T5 and T5.1.1:
+# `2048 * 114`. For mT5: `1024 * 229`. For ByT5: `1024 * 189`.
+LOSS_NORMALIZING_FACTOR = 233472

train_large_v4_16.sh

@@ -5,8 +5,7 @@ export PYTHONPATH=${PROJECT_DIR}
 
 python3 ${T5X_DIR}/t5x/train.py \
   --gin_search_paths=${PROJECT_DIR} \
-  --gin.partitioning.ModelBasedPjitPartitioner.num_partitions=4 \
-  --gin_file="finetune_large_mt5_sentencefix_v4.gin" \
+  --gin_file="finetune_large_mt5_sentencefix_v4_16.gin" \
   --gin.MODEL_DIR="'${MODEL_DIR}'" \
   --tfds_data_dir=${TFDS_DATA_DIR}