use IPEX.device instead of dpcpp && change api

examples/README.md

@@ -71,7 +71,7 @@
 ```
   git clone https://github.com/facebookresearch/dlrm
   cd dlrm
-  pip install -r requirement.txt
+  pip install -r requirements.txt
   git checkout 52b77f80a24303294a02c86b574529cdc420aac5
   patch -p1 < {path/to/intel-pytorch-extension}/torch_patches/models/dlrm.patch
 ```
@@ -82,7 +82,7 @@
   cd transformers
   git checkout 1a779ad7ecb9e5215b6bd1cfa0153469d37e4274
   patch -p1 < {path/to/intel-pytorch-extension}/torch_patches/models/bert.patch
-  pip install -r ./examples/requirement.txt
+  pip install -r ./examples/requirements.txt
   pip install --editable .
 ```
 
@@ -95,6 +95,7 @@
   git clone https://github.com/pytorch/vision
   cd vision
   # add ResNext-101-32x4d model
+  git checkout 883f1fb01a8ba0e1b1cdc16c16f2e6e0ef87e3bd
   patch -p1 < {path/to/intel-pytorch-extension}/torch_patches/models/vision.patch
   python setup.py install
 ```
@@ -124,7 +125,7 @@ export KMP_SETTINGS=1
   --arch-mlp-bot=13-512-256-128 --arch-mlp-top=1024-1024-512-256-1 \
   --arch-sparse-feature-size=128 --max-ind-range=4000000 \
   --numpy-rand-seed=727 \
-  --print-freq=1024 --print-time --mini-batch-size=2048 --num-batches=4096 
+  --print-freq=1024 --print-time --mini-batch-size=2048 --num-batches=10240 
 ```
 ```
     # run DLRM bf16 trainining
@@ -136,7 +137,7 @@ export KMP_SETTINGS=1
   --arch-mlp-bot=13-512-256-128 --arch-mlp-top=1024-1024-512-256-1 \
   --arch-sparse-feature-size=128 --max-ind-range=4000000 \
   --numpy-rand-seed=727 \
-  --print-freq=1024 --print-time --mini-batch-size=2048 --num-batches=4096 \
+  --print-freq=1024 --print-time --mini-batch-size=2048 --num-batches=10240 \
   --use-ipex --mix-precision 
 ```
 ```
@@ -149,7 +150,7 @@ export KMP_SETTINGS=1
   --arch-mlp-bot=13-512-256-128 --arch-mlp-top=1024-1024-512-256-1 \
   --arch-sparse-feature-size=128 --max-ind-range=4000000 \
   --numpy-rand-seed=727 \
-  --print-freq=1024 --print-time --mini-batch-size=16 --num-batches=4096 \
+  --print-freq=1024 --print-time --mini-batch-size=16  \
   --inference-only --share-weight --num-instance=24
 ```
 ```
@@ -162,7 +163,7 @@ export KMP_SETTINGS=1
   --arch-mlp-bot=13-512-256-128 --arch-mlp-top=1024-1024-512-256-1 \
   --arch-sparse-feature-size=128 --max-ind-range=4000000 \
   --numpy-rand-seed=727 \
-  --print-freq=1024 --print-time --mini-batch-size=16 --num-batches=4096 \
+  --print-freq=1024 --print-time --mini-batch-size=16  \
   --use-ipex --mix-precision --inference-only --share-weight --num-instance=24
 ```
 
@@ -189,33 +190,33 @@ examples/language-modeling/run_language_modeling.py \
 ```
 ```
   # run Bert fp32 inference
-for i in $(seq 0 $LAST_INSTANCE); do
+for i in $(seq 0 23); do
     LOG_i=cpufp32_bs1_ins${i}.txt
     echo "### running on instance $i, numa node 0, core  $i"
     numactl --physcpubind=$i --membind=0  python -u \
     examples/language-modeling/run_language_modeling.py \
     --output_dir=output_$i --per_gpu_eval_batch_size=1 \
-    --model_type=bert_large  --do_eval --eval_data_file=$HOME/wikitext-2-raw/wiki.train.raw \
+    --model_type=bert_large  --do_eval --eval_data_file=$DATASET_PATH/wiki.train.raw \
     --overwrite_output_dir --mlm --seed=42 --max_step=30  2>&1 | tee $LOG_i &
 done
 ```
 ```
   # run Bert bf16 inference
-for i in $(seq 0 $LAST_INSTANCE); do
+for i in $(seq 0 23); do
     LOG_i=cpufp32_bs1_ins${i}.txt
     echo "### running on instance $i, numa node 0, core  $i"
     numactl --physcpubind=$i --membind=0  python -u \
     examples/language-modeling/run_language_modeling.py \
     --output_dir=output_$i --per_gpu_eval_batch_size=1 \
-    --model_type=bert_large  --do_eval --eval_data_file=$HOME/wikitext-2-raw/wiki.train.raw \
+    --model_type=bert_large  --do_eval --eval_data_file=$DATASET_PATH/wiki.train.raw \
     --overwrite_output_dir --mlm --seed=42 --max_step=30 --ipex --dnnl --mix_precision  2>&1 | tee $LOG_i &
 done
 ```
 
 3. ResNext-101-32x4d
 ```
   cd {path/to/examples}/imagenet/
-  export DATA_PATH={path/to/dataset}
+  export DATASET_PATH={path/to/dataset}
 ```
 
 ```
@@ -224,13 +225,13 @@ done
 ```
 ```
   # run Bert bf16 traininig
-  bash run_training_cpu_ipex.sh resnext101_32x4d $DATA_PATH
+  bash run_training_cpu_ipex.sh resnext101_32x4d $DATASET_PATH dnnl bf16
 ```
 ```
   # run Bert fp32 inference
   bash run_inference_cpu_latency.sh resnext101_32x4d 
 ```
 ```
   # run Bert bf16 inference
-  bash run_inference_cpu_latency_ipex.sh resnext101_32x4d $DATA_PATH
+  bash run_inference_cpu_latency_ipex.sh resnext101_32x4d $DATASET_PATH dnnl bf16 jit
 ```

torch_patches/models/bert.patch

@@ -1,5 +1,5 @@
 diff --git a/examples/language-modeling/run_language_modeling.py b/examples/language-modeling/run_language_modeling.py
-index 740cb636..6fdf1393 100644
+index 740cb636..ce3b74fb 100644
 --- a/examples/language-modeling/run_language_modeling.py
 +++ b/examples/language-modeling/run_language_modeling.py
 @@ -19,7 +19,7 @@ GPT, GPT-2 and CTRL are fine-tuned using a causal language modeling (CLM) loss.
@@ -37,7 +37,7 @@ index 740cb636..6fdf1393 100644
 +        else:
 +            ipex.core.disable_auto_dnnl()
 +        if model_args.mix_precision:
-+            ipex.enable_auto_mix_precision(mixed_dtype=torch.bfloat16, train=True)
++            ipex.enable_auto_mixed_precision(mixed_dtype=torch.bfloat16, train=True)
      if (
          os.path.exists(training_args.output_dir)
          and os.listdir(training_args.output_dir)
@@ -61,10 +61,32 @@ index 740cb636..6fdf1393 100644
      else:
          raise ValueError(
              "You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
+diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
+index a03ac23f..4754b2a0 100755
+--- a/src/transformers/trainer.py
++++ b/src/transformers/trainer.py
+@@ -1023,6 +1023,8 @@ class Trainer:
+             loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only)
+             batch_size = inputs[list(inputs.keys())[0]].shape[0]
+             samples_count += batch_size
++            if samples_count == self.args.max_steps * batch_size:
++                break
+             if loss is not None:
+                 eval_losses.append(loss * batch_size)
+             if logits is not None:
 diff --git a/src/transformers/training_args.py b/src/transformers/training_args.py
-index 64900136..aae71ac6 100644
+index 64900136..5c183f97 100644
 --- a/src/transformers/training_args.py
 +++ b/src/transformers/training_args.py
+@@ -6,7 +6,7 @@ from dataclasses import dataclass, field
+ from typing import Any, Dict, Optional, Tuple
+
+ from .file_utils import cached_property, is_torch_available, is_torch_tpu_available, torch_required
+-
++import intel_pytorch_extension as ipex
+
+ if is_torch_available():
+     import torch
 @@ -234,6 +234,10 @@ class TrainingArguments:
          default=None, metadata={"help": "An optional descriptor for the run. Notably used for wandb logging."}
      )
@@ -81,21 +103,8 @@ index 64900136..aae71ac6 100644
              device = torch.device("cpu")
              n_gpu = 0
 +        elif self.ipex:
-+            device = torch.device("dpcpp")
++            device = torch.device(ipex.DEVICE)
 +            n_gpu = 0
          elif is_torch_tpu_available():
              device = xm.xla_device()
              n_gpu = 0
-diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
-index 1e09e03d..d254d977 100755
---- a/src/transformers/trainer.py
-+++ b/src/transformers/trainer.py
-@@ -1037,6 +1039,8 @@ class Trainer:
-             loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only)
-             batch_size = inputs[list(inputs.keys())[0]].shape[0]
-             samples_count += batch_size
-+            if samples_count == self.args.max_steps * batch_size:
-+                break
-             if loss is not None:
-                 eval_losses.append(loss * batch_size)
-             if logits is not None:

torch_patches/models/dlrm.patch

@@ -1,3 +1,16 @@
+diff --git a/dlrm_data_pytorch.py b/dlrm_data_pytorch.py
+index 736b6dd..e6b0bc3 100644
+--- a/dlrm_data_pytorch.py
++++ b/dlrm_data_pytorch.py
+@@ -382,7 +382,7 @@ def ensure_dataset_preprocessed(args, d_path):
+
+ def make_criteo_data_and_loaders(args):
+
+-    if args.mlperf_logging and args.memory_map and args.data_set == "terabyte":
++    if args.memory_map and args.data_set == "terabyte":
+         # more efficient for larger batches
+         data_directory = path.dirname(args.raw_data_file)
+
 diff --git a/dlrm_s_pytorch.py b/dlrm_s_pytorch.py
 index 3c5c85f..660f6eb 100644
 --- a/dlrm_s_pytorch.py
@@ -131,8 +144,8 @@ index 3c5c85f..660f6eb 100644
 +        import intel_pytorch_extension as ipex
 +        ipex.core.enable_auto_dnnl()
 +        if args.mix_precision:
-+            ipex.enable_auto_mix_precision(mixed_dtype=torch.bfloat16, train=not args.inference_only)
-+        device = torch.device("dpcpp")
++            ipex.enable_auto_mixed_precision(mixed_dtype=torch.bfloat16, train=not args.inference_only)
++        device = torch.device(ipex.DEVICE)
 +        print("Runining with IPEX...")       
      else:
          device = torch.device("cpu")
@@ -218,9 +231,6 @@ index 3c5c85f..660f6eb 100644
 +    if args.share_weight and args.inference_only:
 +            for j, (X, lS_o, lS_i, T) in enumerate(train_ld):
 +                traced_model = torch.jit.trace(dlrm, wrap_input(X, lS_o, lS_i, use_gpu, device), check_trace=False)
-+                # g=traced_model.graph
-+                # torch._C._jit_pass_inline(g)
-+                # print(g)
 +                break
 +            bench = ThroughputBenchmark(traced_model)
 +            j = 0
@@ -231,12 +241,12 @@ index 3c5c85f..660f6eb 100644
 +                if j == 1: break
 +            if args.use_ipex and args.mix_precision:
 +                ipex.core.disable_mix_bf16_fp32()
-+                stats = bench.benchmark(
-+                    num_calling_threads=args.num_instance,
-+                    num_warmup_iters=100 * args.num_instance,
-+                    num_iters=1000 * args.num_instance,
-+                )
-+                print(stats)
++            stats = bench.benchmark(
++                num_calling_threads=args.num_instance,
++                num_warmup_iters=100 * args.num_instance,
++                num_iters=1000 * args.num_instance,
++            )
++            print(stats)
 +            if args.enable_profiling:
 +                print(prof.key_averages().table(sort_by="self_cpu_time_total"))
 +            sys.exit()