sasarkar/minor_fixes_sdxl

examples/stable-diffusion/requirements.txt

@@ -1 +1,2 @@
 opencv-python
+imagesize

examples/stable-diffusion/training/run_1x.sh

@@ -23,4 +23,5 @@ python train_text_to_image_sdxl.py \
   --validation_prompt="a robotic cat with wings" \
   --validation_epochs 48 \
   --checkpointing_steps 2500 \
-  --logging_step 10 --discount_chkpoint_saving_in_throughput 2>&1 | tee log_1x_r512.txt
+  --logging_step 10 \
+  --adjust_throughput 2>&1 | tee log_1x_r512.txt

examples/stable-diffusion/training/run_8x.sh

@@ -25,4 +25,4 @@ python ../../gaudi_spawn.py --world_size 8 --use_mpi train_text_to_image_sdxl.py
   --validation_epochs 48  \
   --checkpointing_steps 336 \
   --mediapipe dataset_sdxl_pokemon \
-  --discount_chkpoint_saving_in_throughput 2>&1 | tee log_8x_r512.txt
+  --adjust_throughput 2>&1 | tee log_8x_r512.txt

examples/stable-diffusion/training/train_text_to_image_sdxl.py

@@ -525,10 +525,10 @@ def parse_args(input_args=None):
         case 3: a non empty path is passed -> images from that location are used ",
     )
     parser.add_argument(
-        "--discount_chkpoint_saving_in_throughput",
+        "--adjust_throughput",
         default=False,
         action="store_true",
-        help="Checkpoitn saving takes a lot of time. Ignore time for checkpoint saving for throughput calculations"
+        help="Checkpoint saving takes a lot of time. Ignore time for checkpoint saving for throughput calculations"
     )
 
 
@@ -1340,7 +1340,7 @@ def compute_time_ids(original_size, crops_coords_top_left):
 
                 del pipeline
 
-    duration = time.perf_counter() - t0 - (checkpoint_time if args.discount_chkpoint_saving_in_throughput else 0)
+    duration = time.perf_counter() - t0 - (checkpoint_time if args.adjust_throughput else 0)
     ttt = time.perf_counter() - t_start
     throughput = (args.max_train_steps - args.throughput_warmup_steps) * total_batch_size / duration