debug

Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com>

Author: kmehant

scripts/benchmarks/compare_with_reference.py

@@ -10,15 +10,16 @@
 
 # default columns to compare
 DEFAULT_PLOT_COLUMNS = [
-    "mem_torch_mem_alloc_in_bytes",
-    "mem_peak_torch_mem_alloc_in_bytes",
+    # "mem_torch_mem_alloc_in_bytes",
+    # "mem_peak_torch_mem_alloc_in_bytes",
+    'mem_nvidia_mem_reserved',
     "train_loss",
     "train_tokens_per_second",
 ]
 # Used as combined identifier of experiment
 DEFAULT_INDICES = [
     "framework_config",
-    "peft_method",
+    # "peft_method",
     "model_name_or_path",
     "num_gpus",
     "per_device_train_batch_size",
@@ -29,7 +30,7 @@
     "train_runtime",
     "train_steps_per_second",
     "train_samples_per_second",
-    "mem_nvidia_mem_reserved",
+    # "mem_nvidia_mem_reserved",
 ]
 
 DEFAULT_REFERENCE_FILEPATH = "scripts/benchmarks/refs/a100_80gb.csv"

scripts/benchmarks/scenarios-moe.yaml

@@ -41,13 +41,13 @@ scenarios:
             - # without acceleration
             - moe-scattermoe-granite-ep1
             - moe-scattermoe-granite-ep2
-            - moe-scattermoe-granite-ep4
-            - moe-scattermoe-granite-ep1-padding-free
-            - moe-scattermoe-granite-ep1-padding-free-foak
-            - moe-scattermoe-granite-ep2-padding-free
-            - moe-scattermoe-granite-ep2-padding-free-foak
-            - moe-scattermoe-granite-ep4-padding-free
-            - moe-scattermoe-granite-ep4-padding-free-foak
+            # - moe-scattermoe-granite-ep4
+            # - moe-scattermoe-granite-ep1-padding-free
+            # - moe-scattermoe-granite-ep1-padding-free-foak
+            # - moe-scattermoe-granite-ep2-padding-free
+            # - moe-scattermoe-granite-ep2-padding-free-foak
+            # - moe-scattermoe-granite-ep4-padding-free
+            # - moe-scattermoe-granite-ep4-padding-free-foak
         arguments:
             learning_rate: 5e-5
             torch_dtype: bfloat16
@@ -58,8 +58,8 @@ scenarios:
             adam_epsilon: 1e-8
             model_name_or_path: 
                 - 'ibm-granite/granite-4.0-tiny-preview'
-                - 'ibm-granite/granite-3.0-3b-a800m-instruct'
-                - 'ibm-research/moe-7b-1b-active-shared-experts'
+                # - 'ibm-granite/granite-3.0-3b-a800m-instruct'
+                # - 'ibm-research/moe-7b-1b-active-shared-experts'
 
 
     -   name: accelerated-moe-full-mixtral