Merge pull request #6 from haifeng-jin/rm

removed native pytorch code
haifeng-jin · May 20, 2024 · 4fe0fde · 4fe0fde
2 parents 4f1d20e + f9792f0
commit 4fe0fde
Show file tree

Hide file tree

Showing 38 changed files with 40 additions and 744 deletions.
diff --git a/.gitignore b/.gitignore
@@ -0,0 +1,4 @@
+.venv
+results.txt
+*.egg-info
+__pycache__
diff --git a/README.md b/README.md
@@ -16,23 +16,6 @@ Python package versions for each framework.
 
 ## Permission setups
 
-### HuggingFace setup
-
-On the [HuggingFace Gemma model page](https://huggingface.co/google/gemma-7b),
-make sure you have accepted the license near the top of the page.
-
-```shell
-pip install --upgrade huggingface_hub
-```
-
-```shell
-huggingface-cli login
-```
-
-It may require you to input a token.
-[More information about tokens.](https://huggingface.co/docs/hub/en/security-tokens)
-
-
 ### Kaggle setup
 
 On the [Kaggle Gemma model page](https://www.kaggle.com/models/keras/gemma),
@@ -82,7 +65,7 @@ run `shell/cleanup.sh`.
   structured as a Python package. I needs `pip install -e .` before using. Most
   of the settings are in `benchmark/__init__.py`. You can run a single benchmark
   by calling each script, for example,
-  `python benchmark/gemma/keras/predict.py results.txt`
+  `python benchmark/gemma/predict.py results.txt`
 * `shell` contains all the shell scripts for benchmarking.
 * `requirements` contains the version requirements for the PyPI packages in the
   dependencies.

diff --git a/benchmark/__init__.py b/benchmark/__init__.py
@@ -40,7 +40,7 @@ def append_to_file(file_path, content):
 
 def benchmark(run):
     if len(sys.argv) not in (2, 3):
-        print("Usage: python bert/keras/fit.py <file_path> [batch_size]")
+        print("Usage: python bert/fit.py <file_path> [batch_size]")
     else:
         if len(sys.argv) == 3:
             batch_size = int(sys.argv[2])

diff --git a/benchmark/bert/keras/fit.py → benchmark/bert/fit.py b/benchmark/bert/keras/fit.py → benchmark/bert/fit.py
@@ -2,14 +2,14 @@
 import keras_nlp
 
 import benchmark
-from benchmark import keras_utils
+from benchmark import utils
 
 
 def run(batch_size=benchmark.BERT_FIT_BATCH_SIZE):
     preprocessor = keras_nlp.models.BertPreprocessor.from_preset(
         "bert_base_en", sequence_length=benchmark.BERT_SEQ_LENGTH
     )
-    dataset = keras_utils.get_train_dataset_for_text_classification(
+    dataset = utils.get_train_dataset_for_text_classification(
         preprocessor=preprocessor,
         batch_size=batch_size,
         seq_len=benchmark.BERT_SEQ_LENGTH,
@@ -22,10 +22,10 @@ def run(batch_size=benchmark.BERT_FIT_BATCH_SIZE):
     model.compile(
         loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
         optimizer=keras.optimizers.AdamW(),
-        jit_compile=keras_utils.use_jit(),
+        jit_compile=utils.use_jit(),
     )
 
-    return keras_utils.fit(model, dataset)
+    return utils.fit(model, dataset)
 
 
 if __name__ == "__main__":

diff --git a/benchmark/bert/keras/__init__.py b/benchmark/bert/keras/__init__.py
diff --git a/benchmark/bert/keras/predict.py → benchmark/bert/predict.py b/benchmark/bert/keras/predict.py → benchmark/bert/predict.py
@@ -1,14 +1,14 @@
 import keras_nlp
 
 import benchmark
-from benchmark import keras_utils
+from benchmark import utils
 
 
 def run(batch_size=benchmark.BERT_BATCH_SIZE):
     preprocessor = keras_nlp.models.BertPreprocessor.from_preset(
         "bert_base_en", sequence_length=benchmark.BERT_SEQ_LENGTH
     )
-    dataset = keras_utils.get_train_dataset_for_text_classification(
+    dataset = utils.get_train_dataset_for_text_classification(
         preprocessor=preprocessor,
         batch_size=batch_size,
         seq_len=benchmark.BERT_SEQ_LENGTH,
@@ -19,10 +19,10 @@ def run(batch_size=benchmark.BERT_BATCH_SIZE):
         preprocessor=None,
     )
     model.compile(
-        jit_compile=keras_utils.use_jit(),
+        jit_compile=utils.use_jit(),
     )
 
-    return keras_utils.predict(model, dataset)
+    return utils.predict(model, dataset)
 
 
 if __name__ == "__main__":

diff --git a/benchmark/bert/torch/__init__.py b/benchmark/bert/torch/__init__.py
diff --git a/benchmark/bert/torch/fit.py b/benchmark/bert/torch/fit.py
diff --git a/benchmark/bert/torch/predict.py b/benchmark/bert/torch/predict.py
diff --git a/benchmark/gemma/keras/fit.py → benchmark/gemma/fit.py b/benchmark/gemma/keras/fit.py → benchmark/gemma/fit.py
@@ -2,7 +2,7 @@
 import keras_nlp
 
 import benchmark
-from benchmark import keras_utils
+from benchmark import utils
 
 
 def get_model():
@@ -22,17 +22,17 @@ def run(batch_size=benchmark.GEMMA_FIT_BATCH_SIZE):
     preprocessor = keras_nlp.models.GemmaCausalLMPreprocessor.from_preset(
         "gemma_7b_en", sequence_length=benchmark.GEMMA_SEQ_LENGTH
     )
-    dataset = keras_utils.get_train_dataset_for_text_gen(
+    dataset = utils.get_train_dataset_for_text_gen(
         preprocessor, batch_size, seq_len=benchmark.GEMMA_SEQ_LENGTH
     )
     model = get_model()
 
     model.compile(
         loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
         optimizer=keras.optimizers.AdamW(),
-        jit_compile=keras_utils.use_jit(),
+        jit_compile=utils.use_jit(),
     )
-    return keras_utils.fit(model, dataset)
+    return utils.fit(model, dataset)
 
 
 if __name__ == "__main__":

diff --git a/benchmark/gemma/keras/__init__.py b/benchmark/gemma/keras/__init__.py
diff --git a/benchmark/gemma/keras/predict.py → benchmark/gemma/predict.py b/benchmark/gemma/keras/predict.py → benchmark/gemma/predict.py
@@ -2,7 +2,7 @@
 import keras_nlp
 
 import benchmark
-from benchmark import keras_utils
+from benchmark import utils
 
 
 def run(batch_size=benchmark.GEMMA_BATCH_SIZE):
@@ -12,7 +12,7 @@ def run(batch_size=benchmark.GEMMA_BATCH_SIZE):
         keras.mixed_precision.set_global_policy(benchmark.FLOAT_A100)
     model = keras_nlp.models.GemmaCausalLM.from_preset("gemma_7b_en")
     model.compile(sampler="greedy")
-    return keras_utils.generate(
+    return utils.generate(
         model=model,
         batch_size=batch_size,
         max_length=benchmark.GEMMA_MAX_LENGTH,

diff --git a/benchmark/gemma/torch/__init__.py b/benchmark/gemma/torch/__init__.py
diff --git a/benchmark/gemma/torch/fit.py b/benchmark/gemma/torch/fit.py
diff --git a/benchmark/gemma/torch/predict.py b/benchmark/gemma/torch/predict.py
diff --git a/benchmark/mistral/keras/fit.py → benchmark/mistral/fit.py b/benchmark/mistral/keras/fit.py → benchmark/mistral/fit.py
@@ -2,7 +2,7 @@
 import keras_nlp
 
 import benchmark
-from benchmark import keras_utils
+from benchmark import utils
 
 
 def get_model():
@@ -23,17 +23,17 @@ def run(batch_size=benchmark.MISTRAL_FIT_BATCH_SIZE):
         "mistral_7b_en",
         sequence_length=benchmark.MISTRAL_SEQ_LENGTH,
     )
-    dataset = keras_utils.get_train_dataset_for_text_gen(
+    dataset = utils.get_train_dataset_for_text_gen(
         preprocessor, batch_size, seq_len=benchmark.MISTRAL_SEQ_LENGTH
     )
     model = get_model()
 
     model.compile(
         loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
         optimizer=keras.optimizers.AdamW(),
-        jit_compile=keras_utils.use_jit(),
+        jit_compile=utils.use_jit(),
     )
-    return keras_utils.fit(model, dataset)
+    return utils.fit(model, dataset)
 
 
 if __name__ == "__main__":

diff --git a/benchmark/mistral/keras/__init__.py b/benchmark/mistral/keras/__init__.py
diff --git a/benchmark/mistral/keras/predict.py → benchmark/mistral/predict.py b/benchmark/mistral/keras/predict.py → benchmark/mistral/predict.py
@@ -2,7 +2,7 @@
 import keras_nlp
 
 import benchmark
-from benchmark import keras_utils
+from benchmark import utils
 
 
 def run(batch_size=benchmark.MISTRAL_BATCH_SIZE):
@@ -12,7 +12,7 @@ def run(batch_size=benchmark.MISTRAL_BATCH_SIZE):
         keras.mixed_precision.set_global_policy(benchmark.FLOAT_A100)
     model = keras_nlp.models.MistralCausalLM.from_preset("mistral_7b_en")
     model.compile(sampler="greedy")
-    return keras_utils.generate(
+    return utils.generate(
         model=model,
         batch_size=batch_size,
         max_length=benchmark.MISTRAL_MAX_LENGTH,

diff --git a/benchmark/mistral/torch/__init__.py b/benchmark/mistral/torch/__init__.py