add block_mask & retrain_free features (#775)

Signed-off-by: Zhang, Weiwei1 <[email protected]>
Co-authored-by: wenhuach21 <[email protected]>

Author: WeiweiZhang1

examples/pytorch/nlp/huggingface_models/language-modeling/pruning/eager/run_clm_no_trainer.py

@@ -0,0 +1,17 @@
+#!/bin/bash
+set -x
+    python \
+    examples/pytorch/nlp/huggingface_models/language-modeling/pruning/eager/run_clm_no_trainer.py \
+    --model_name_or_path /path/to/your/model \
+    --dataset_name lambada \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 16 \
+    --max_train_steps 3002 \
+    --weight_decay  0 \
+    --block_size 512 \
+    --do_prune \
+    --auto_slim \
+    --output_dir sparse_clm_models/ \
+    --target_sparsity 0.2 \
+    --pruning_pattern channelx1 \
+    --pruning_frequency 500 \

examples/pytorch/nlp/huggingface_models/language-modeling/pruning/eager/run_llm_pruning.sh

@@ -0,0 +1,33 @@
+import time
+import torch
+class CPUTimer:
+    def __init__(self, timelogs):
+        self.timelogs = timelogs
+
+    def __enter__(self):
+        self.start = time.time()
+
+    def __exit__(self):
+        end = time.time()
+        self.timelogs.append((end - self.start) * 1000) # ms
+    
+    def get_avg_time(self):
+        return sum(self.timelogs) / len(self.timelogs)
+
+class GPUTimer:
+    def __init__(self, timelogs):
+        self.timelogs = timelogs
+
+    def __enter__(self):
+        self.start_event = torch.cuda.Event(enable_timing=True)
+        self.end_event = torch.cuda.Event(enable_timing=True)
+        self.start_event.record()
+
+    def __exit__(self):
+        self.end_event.record()
+        self.end_event.synchronize()
+        elapsed_time = self.start_event.elapsed_time(self.end_event)
+        self.timelogs.append(elapsed_time)
+
+    def get_avg_time(self):
+        return sum(self.timelogs) / len(self.timelogs)