Cleanups for intel gptj int4

mlcommons · Mar 14, 2024 · 6d2b1c3 · 6d2b1c3
1 parent 171dcbc
commit 6d2b1c3
Show file tree

Hide file tree

Showing 4 changed files with 33 additions and 15 deletions.
diff --git a/cm-mlops/script/app-mlperf-inference-intel/_cm.yaml b/cm-mlops/script/app-mlperf-inference-intel/_cm.yaml
@@ -426,13 +426,18 @@ variations:
       CM_LOCAL_MLPERF_INFERENCE_INTEL_RUN_MODE: calibration
     new_env_keys:
       - CM_ML_MODEL_*
+      - INT4_CALIBRATION_DIR
 
   calibration,gptj_:
     deps:
       - tags: get,generic-python-lib,_package.texttable
         names:
           - pip-package
           - texttable
+      - tags: get,generic-python-lib,_package.transformers
+        names:
+          - pip-package
+        version: "4.21.2"
 
   build-harness,bert_:
     deps:

diff --git a/cm-mlops/script/app-mlperf-inference-intel/build_gptj_harness.sh b/cm-mlops/script/app-mlperf-inference-intel/build_gptj_harness.sh
@@ -33,5 +33,7 @@ export CALIBRATION_DATA_JSON=${WORKLOAD_DATA}/calibration-data/cnn_dailymail_cal
 export VALIDATION_DATA_JSON=${WORKLOAD_DATA}/validation-data/cnn_dailymail_validation.json
 export INT4_CALIBRATION_DIR=${WORKLOAD_DATA}/quantized-int4-model
 #sudo -E bash run_quantization.sh
-bash run_quantization.sh
+#bash run_quantization.sh
+echo "${RUN_QUANTIZATION_CMD}"
+eval "${RUN_QUANTIZATION_CMD}"
 test $? -eq 0 || exit $?
diff --git a/cm-mlops/script/app-mlperf-inference-intel/calibrate_gptj_int4_model.sh b/cm-mlops/script/app-mlperf-inference-intel/calibrate_gptj_int4_model.sh
@@ -1,12 +1,14 @@
 #!/bin/bash
+
 export PATH=${CM_CONDA_BIN_PATH}:$PATH
+
 cd ${CM_MLPERF_INFERENCE_INTEL_CALIBRATION_PATH}
 CUR_DIR=$(pwd)
 export WORKLOAD_DATA=${CUR_DIR}/data
 mkdir -p ${WORKLOAD_DATA}
 
-echo $PATH
 python download-calibration-dataset.py --calibration-list-file calibration-list.txt --output-dir ${WORKLOAD_DATA}/calibration-data
+test $? -eq 0 || exit $?
 
 export CALIBRATION_DATA_JSON=${WORKLOAD_DATA}/calibration-data/cnn_dailymail_calibration.json
 
@@ -20,4 +22,3 @@ export QUANTIZED_MODEL_DIR=${WORKLOAD_DATA}/quantized-int4-model
 mkdir -p ${QUANTIZED_MODEL_DIR}
 bash run_calibration_int4.sh
 test $? -eq 0 || exit $?
-exit 0
diff --git a/cm-mlops/script/app-mlperf-inference-intel/customize.py b/cm-mlops/script/app-mlperf-inference-intel/customize.py
@@ -63,10 +63,7 @@ def preprocess(i):
             i['run_script_input']['script_name'] = "calibrate_gptj_int4_model"
             calibration_path = os.path.join(calibration_root, "INT4")
             env['CM_MLPERF_INFERENCE_INTEL_CALIBRATION_PATH'] = calibration_path
-            '''env['CALIBRATION_DATA_JSON'] =
-            env['CHECKPOINT_DIR'] =
-            env['QUANTIZED_MODEL_DIR'] =
-            '''
+            env['INT4_CALIBRATION_DIR'] = os.path.join(calibration_path, "data", "quantized-int4-model")
 
 
     elif env['CM_LOCAL_MLPERF_INFERENCE_INTEL_RUN_MODE'] == "build_harness":
@@ -80,13 +77,22 @@ def preprocess(i):
             env['CM_MLPERF_INFERENCE_INTEL_HARNESS_PATH'] = os.path.join(os.getcwd(), "harness", "build", "gptj_inference")
             env['DATA_PATH'] = os.path.join(os.getcwd(), "harness", "gptj")
             env['MLPERF_INFERENCE_ROOT'] = env['CM_MLPERF_INFERENCE_SOURCE']
-            final_model_path = os.path.join(harness_root, "data", "gpt-j-int8-model", "best_model.pt")
-            env['INT8_MODEL_DIR'] = os.path.dirname(final_model_path)
-            if not os.path.exists(env['INT8_MODEL_DIR']):
-                os.makedirs(env['INT8_MODEL_DIR'])
-            env['CM_ML_MODEL_PATH'] = env['INT8_MODEL_DIR']
-            if env.get('CM_MLPERF_INFERENCE_INTEL_GPTJ_INT8_MODEL_PATH', '') != '':
-                shutil.copy(env['CM_MLPERF_INFERENCE_INTEL_GPTJ_INT8_MODEL_PATH'], env['INT8_MODEL_DIR'])
+            if env.get('INTEL_GPTJ_INT4', '') == 'yes':
+                model_precision = "int4"
+                env['RUN_QUANTIZATION_CMD'] = "bash run_quantization_int4.sh"
+            else:
+                model_precision = "int8"
+                env['RUN_QUANTIZATION_CMD'] = "bash run_quantization.sh"
+            final_model_path = os.path.join(harness_root, "data", f"gpt-j-{model_precision}-model", "best_model.pt")
+            model_dir_name = f"{model_precision.upper()}_MODEL_DIR"
+            env[model_dir_name] = os.path.dirname(final_model_path)
+            if not os.path.exists(env[model_dir_name]):
+                os.makedirs(env[model_dir_name])
+            env['CM_ML_MODEL_PATH'] = env[model_dir_name]
+            if env.get('CM_MLPERF_INFERENCE_INTEL_GPTJ_INT8_MODEL_PATH', '') != '' and env.get('INT8_MODEL_DIR', '') != '':
+                shutil.copy(env['CM_MLPERF_INFERENCE_INTEL_GPTJ_INT8_MODEL_PATH'], env[model_dir_name])
+            if env.get('CM_MLPERF_INFERENCE_INTEL_GPTJ_INT4_MODEL_PATH', '') != '' and env.get('INT4_MODEL_DIR', '') != '':
+                shutil.copy(env['CM_MLPERF_INFERENCE_INTEL_GPTJ_INT4_MODEL_PATH'], env[model_dir_name])
 
     elif env['CM_LOCAL_MLPERF_INFERENCE_INTEL_RUN_MODE'] == "run_harness":
         print(f"Harness Root: {harness_root}")
@@ -104,7 +110,11 @@ def preprocess(i):
                 env['LOADGEN_MODE'] = 'Accuracy'
             else:
                 env['LOADGEN_MODE'] = 'Performance'
-            env['INT8_MODEL_DIR'] = env['CM_ML_MODEL_PATH']
+            if env.get('INTEL_GPTJ_INT4', '') == 'yes':
+                model_precision = "int4"
+                env['INT4_MODEL_DIR'] = env['CM_ML_MODEL_PATH']
+            else:
+                env['INT8_MODEL_DIR'] = env['CM_ML_MODEL_PATH']
             env['CM_RUN_DIR'] = i['run_script_input']['path']
             env['CM_RUN_CMD'] = "bash run_gptj_harness.sh "