diff --git a/script/jax/a100/condor.sub b/script/jax/a100/condor.sub
new file mode 100644
index 000000000..1745a814b
--- /dev/null
+++ b/script/jax/a100/condor.sub
@@ -0,0 +1,42 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/jax/a100/auto_run_array.sh
+
+num_jobs=8
+
+# Job specific vars
+workload_or_id=$(Process)
+# workload_or_id=imagenet_vit
+framework=jax
+submission=prize_qualification_baselines/external_tuning/jax_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=a100_yesTF32_04
+study=1
+num_tuning_trials=1
+rng_seed=96
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 1000000
+request_cpus = 24
+request_gpus = 8
+requirements = (TARGET.CUDADeviceName == "NVIDIA A100-SXM4-80GB")
+# requirements = (TARGET.CUDADeviceName == "NVIDIA A100-SXM4-40GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/script/jax/v100/condor.sub b/script/jax/v100/condor.sub
new file mode 100644
index 000000000..a172967ce
--- /dev/null
+++ b/script/jax/v100/condor.sub
@@ -0,0 +1,41 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/jax/v100/auto_run_array.sh
+
+num_jobs=8
+
+# Job specific vars
+workload_or_id=$(Process)
+# workload_or_id=imagenet_vit
+framework=jax
+submission=prize_qualification_baselines/external_tuning/jax_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=v100_04
+study=1
+num_tuning_trials=1
+rng_seed=96
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 700000
+request_cpus = 24
+request_gpus = 8
+requirements = (TARGET.CUDADeviceName == "Tesla V100-SXM2-32GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/script/pytorch/2xa100/condor.sub b/script/pytorch/2xa100/condor.sub
new file mode 100644
index 000000000..e6b1c8455
--- /dev/null
+++ b/script/pytorch/2xa100/condor.sub
@@ -0,0 +1,45 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/pytorch/2xa100/auto_run_array.sh
+
+num_jobs=8
+
+# Job specific vars
+workload_or_id=$(Process)
+# workload_or_id=criteo1tb
+framework=pytorch
+submission=prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=a10040GB_x2_noTF32_01
+study=1
+num_tuning_trials=1
+rng_seed=96
+allow_tf_32=0
+halve_cuda_mem=0
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed) \
+  $(allow_tf_32) \
+  $(halve_cuda_mem)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 250000
+request_cpus = 12
+request_gpus = 2
+requirements = (TARGET.CUDADeviceName == "NVIDIA A100-SXM4-40GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/script/pytorch/4xa100/condor.sub b/script/pytorch/4xa100/condor.sub
new file mode 100644
index 000000000..ff2e3c39b
--- /dev/null
+++ b/script/pytorch/4xa100/condor.sub
@@ -0,0 +1,45 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/pytorch/4xa100/auto_run_array.sh
+
+num_jobs=8
+
+# Job specific vars
+workload_or_id=$(Process)
+# workload_or_id=imagenet_resnet
+framework=pytorch
+submission=prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=a100_x4_yesTF32_10
+study=1
+num_tuning_trials=1
+rng_seed=96
+allow_tf_32=1
+eval_num_workers=4
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed) \
+  $(allow_tf_32) \
+  $(eval_num_workers)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 700000
+request_cpus = 36
+request_gpus = 4
+requirements = (TARGET.CUDADeviceName == "NVIDIA A100-SXM4-40GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/script/pytorch/8xa100/condor.sub b/script/pytorch/8xa100/condor.sub
new file mode 100644
index 000000000..4eb11b415
--- /dev/null
+++ b/script/pytorch/8xa100/condor.sub
@@ -0,0 +1,43 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/pytorch/8xa100/auto_run_array.sh
+
+num_jobs=8
+
+# Job specific vars
+workload_or_id=$(Process)
+# workload_or_id=imagenet_vit
+framework=pytorch
+submission=prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=a100_x8_noTF32_10
+study=1
+num_tuning_trials=1
+rng_seed=96
+allow_tf_32=0
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed) \
+  $(allow_tf_32)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 700000
+request_cpus = 36
+request_gpus = 8
+requirements = (TARGET.CUDADeviceName == "NVIDIA A100-SXM4-40GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/script/pytorch/v100/condor.sub b/script/pytorch/v100/condor.sub
new file mode 100644
index 000000000..2d14bcd2f
--- /dev/null
+++ b/script/pytorch/v100/condor.sub
@@ -0,0 +1,41 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/pytorch/v100/auto_run_array.sh
+
+num_jobs=8
+
+# Job specific vars
+workload_or_id=$(Process)
+# workload_or_id=imagenet_resnet
+framework=pytorch
+submission=prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=v100_10
+study=1
+num_tuning_trials=1
+rng_seed=96
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 700000
+request_cpus = 36
+request_gpus = 8
+requirements = (TARGET.CUDADeviceName == "Tesla V100-SXM2-32GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/script/pytorch/v100_test/condor.sub b/script/pytorch/v100_test/condor.sub
new file mode 100644
index 000000000..7803590d1
--- /dev/null
+++ b/script/pytorch/v100_test/condor.sub
@@ -0,0 +1,45 @@
+# Define variables
+LOGS_DIR=/fast/najroldi/logs/algoperf
+EXE=/home/najroldi/algorithmic-efficiency/script/pytorch/v100_test/auto_run_array.sh
+
+num_jobs=1
+
+# Job specific vars
+# workload_or_id=$(Process)
+workload_or_id=imagenet_resnet
+framework=pytorch
+submission=prize_qualification_baselines/external_tuning/pytorch_nadamw_full_budget.py
+search_space=prize_qualification_baselines/external_tuning/tuning_search_space.json
+name=v100_test_05
+study=1
+num_tuning_trials=1
+rng_seed=96
+eval_workers=2
+omp_threads=2
+
+# Args
+executable = $(EXE)
+arguments = \
+  $(workload_or_id) \
+  $(framework) \
+  $(submission) \
+  $(search_space) \
+  $(name) \
+  $(study) \
+  $(num_tuning_trials) \
+  $(rng_seed) \
+  $(eval_workers) \
+  $(omp_threads)
+
+# Logs
+error = $(LOGS_DIR)/err/job.$(Cluster).$(Process).err
+output = $(LOGS_DIR)/out/job.$(Cluster).$(Process).out
+log = $(LOGS_DIR)/log/job.$(Cluster).$(Process).log
+
+# Specs
+request_memory = 700000
+request_cpus = 36
+request_gpus = 8
+requirements = (TARGET.CUDADeviceName == "Tesla V100-SXM2-32GB")
+
+queue $(num_jobs)
\ No newline at end of file
diff --git a/submission_runner.py b/submission_runner.py
index 551173bf5..60e736e9d 100644
--- a/submission_runner.py
+++ b/submission_runner.py
@@ -133,6 +133,9 @@
 flags.DEFINE_integer('max_global_steps',
                      None,
                      'Maximum number of update steps.')
+flags.DEFINE_float('max_pct_of_global_steps',
+                   0.0,
+                   'Maximum number of update steps.')
 flags.DEFINE_boolean(
     'overwrite',
     False,
@@ -160,6 +163,14 @@
     'Number of workers for ImageNet PyTorch evaluation data loaders.'
     'WARNING: Setting pytorch_eval_num_workers != 0, will result '
     'in incorrect evals currently, see issues/732.')
+flags.DEFINE_boolean(
+    'halve_CUDA_mem',
+    False,
+    'Halve the available VRAM.')
+flags.DEFINE_boolean(
+    'allow_tf32',
+    False,
+    'Allow TF32 on Ampere.')
 FLAGS = flags.FLAGS
 USE_PYTORCH_DDP, RANK, DEVICE, N_GPUS = pytorch_setup()
 
@@ -205,6 +216,7 @@ def train_once(
     rng: spec.RandomState,
     profiler: Profiler,
     max_global_steps: int = None,
+    max_pct_of_global_steps: float = None,
     log_dir: Optional[str] = None,
     save_checkpoints: Optional[bool] = True
 ) -> Tuple[spec.Timing, Dict[str, Any]]:
@@ -365,6 +377,10 @@ def train_once(
     global_step += 1
     if (max_global_steps is not None) and (global_step == max_global_steps):
       train_state['training_complete'] = True
+    # (nico): train for a fixed pct of step_hint
+    if (max_pct_of_global_steps is not None) and \
+        (global_step / workload.step_hint >= max_pct_of_global_steps):
+      train_state['training_complete'] = True
 
     train_step_end_time = get_time()
 
@@ -498,6 +514,7 @@ def score_submission_on_workload(workload: spec.Workload,
                                  tuning_ruleset: str,
                                  profiler: Optional[Profiler] = None,
                                  max_global_steps: Optional[int] = None,
+                                 max_pct_of_global_steps: Optional[float] = None,
                                  imagenet_v2_data_dir: Optional[str] = None,
                                  tuning_search_space: Optional[str] = None,
                                  num_tuning_trials: Optional[int] = None,
@@ -595,6 +612,7 @@ def score_submission_on_workload(workload: spec.Workload,
                                      rng,
                                      profiler,
                                      max_global_steps,
+                                     max_pct_of_global_steps,
                                      tuning_dir_name,
                                      save_checkpoints=save_checkpoints,)
       all_timings[hi] = timing
@@ -631,6 +649,19 @@ def score_submission_on_workload(workload: spec.Workload,
 
 
 def main(_):
+
+  if FLAGS.framework == 'pytorch':
+
+    if FLAGS.halve_CUDA_mem:
+      torch.cuda.set_per_process_memory_fraction(0.5, device=DEVICE)
+
+    if FLAGS.allow_tf32:
+      torch.backends.cuda.matmul.allow_tf32 = True
+      torch.backends.cudnn.allow_tf32 = True
+    else:
+      torch.backends.cuda.matmul.allow_tf32 = False
+      torch.backends.cudnn.allow_tf32 = False
+
   if FLAGS.profile:
     profiler = Profiler()
   else:
@@ -687,6 +718,7 @@ def main(_):
       tuning_ruleset=FLAGS.tuning_ruleset,
       profiler=profiler,
       max_global_steps=FLAGS.max_global_steps,
+      max_pct_of_global_steps=FLAGS.max_pct_of_global_steps,
       imagenet_v2_data_dir=FLAGS.imagenet_v2_data_dir,
       tuning_search_space=FLAGS.tuning_search_space,
       num_tuning_trials=FLAGS.num_tuning_trials,