Group dependencies and remove requirements.txt (stanford-crfm#1681)

Author: yifanmai

.github/workflows/test.yml

@@ -20,7 +20,7 @@ jobs:
       - uses: actions/cache@v2
         with:
           path: ~/.cache/pip
-          key: pip-${{ hashFiles('requirements.txt') }}-${{ matrix.python-version }}
+          key: pip-${{ hashFiles('requirements-freeze.txt') }}-${{ matrix.python-version }}
           restore-keys: |
             pip-
       - run: pip install -e .
@@ -43,7 +43,7 @@ jobs:
       - uses: actions/cache@v2
         with:
           path: ~/.cache/pip
-          key: pip-${{ hashFiles('requirements.txt') }}
+          key: pip-${{ hashFiles('requirements-freeze.txt') }}
           restore-keys: |
             pip-
       # Installs dependencies and performs static code checks

MANIFEST.in

@@ -1,4 +1,3 @@
 recursive-include src/helm/proxy/clients/ *.sp
 recursive-include src/helm/benchmark/ *.json
 recursive-include src/helm/benchmark/static/ *.css *.html *.js *.png *.yaml
-include requirements.txt

requirements.txt

@@ -23,68 +23,79 @@ include_package_data = True
 
 install_requires=
     # Common
-    zstandard~=0.18.0
-    tqdm~=4.64.1
-    pyhocon~=0.3.59
+    cattrs~=22.2.0
     dacite~=1.6.0
-    simple-slurm~=0.2.6
-
-    # Proxy
-    aleph-alpha-client~=2.14.0
-    anthropic~=0.2.5
-    bottle~=0.12.23
-    gunicorn~=20.1.0
+    importlib-resources~=5.10.0
     Mako~=1.2.3
+    numpy~=1.23.3
+    pyhocon~=0.3.59
+    retrying~=1.3.4
+    simple-slurm~=0.2.6  # For slurm_jobs
+    spacy~=3.5.3
+    tqdm~=4.64.1
+    zstandard~=0.18.0
     # sqlitedict==2.0.0 is slow! https://github.com/RaRe-Technologies/sqlitedict/issues/152
-    # Keep the version at 1.7.0.
+    # Keep sqlitedict version at 1.7.0.
     sqlitedict~=1.7.0
-    pymongo~=4.2.0
-    retrying~=1.3.4
-    websocket-client~=1.3.2 # For Anthropic (Legacy stanford-online-all-v4-s3)
-    openai~=0.27.8
-    tiktoken~=0.3.3
-    transformers~=4.28.1
-    tokenizers~=0.13.3
-    icetk~=0.0.4
+    # TODO: Remove these from common
     protobuf~=3.20.2  # Can't use 4.21.0 due to backward incompatibility
-    google-api-python-client~=2.64.0
-    revChatGPT~=0.1.1
+    pymongo~=4.2.0
 
-    # Scenarios
-    datasets~=2.5.2  # For math scenarios
+    # Basic Scenarios
+    datasets~=2.5.2
     pyarrow~=11.0.0  # Pinned transitive dependency for datasets; workaround for #1026
-    gdown~=4.4.0
     jsonlines~=3.1.0  # Not really needed
-    sympy~=1.11.1  # For math scenarios
-    sentencepiece~=0.1.97
-    numba~=0.56.4
-    cattrs~=22.2.0
-    xlrd~=2.0.1  # Used by pandas.read_excel in ice_scenario
 
-    # Metrics
-    importlib-resources~=5.10.0
+    # Basic metrics
     nltk~=3.7
+    pyext~=0.7
+    rouge-score~=0.1.2
     scipy~=1.9.1
     uncertainty-calibration~=0.1.3
-    rouge-score~=0.1.2
-    pyext~=0.7
-    pytrec_eval==0.5
-    sacrebleu~=2.2.1
-    scaleapi~=2.13.0
     # Work around https://github.com/p-lambda/verified_calibration/issues/11
     # TODO: Remove after this issue is resolved
     scikit-learn~=1.1.2
-    spacy~=3.5.3
-    summ-eval~=0.892
+
+    # Server Extras
+    bottle~=0.12.23
+    gunicorn~=20.1.0
+
+    # Scenario Extras
+    gdown~=4.4.0  # For opinions_qa_scenario
+    sympy~=1.11.1  # For numeracy_scenario
+    xlrd~=2.0.1  # For ice_scenario: used by pandas.read_excel
+
+    # Model Extras
+    aleph-alpha-client~=2.14.0
+    anthropic~=0.2.5
+    icetk~=0.0.4  # for ice_tokenizer_client
+    openai~=0.27.8
+    revChatGPT~=0.1.1
+    sentencepiece~=0.1.97  # For palmyra_client and yalm_tokenizer
+    tiktoken~=0.3.3  # for openai_client
+    tokenizers~=0.13.3  # for aleph_alpha_client
+    websocket-client~=1.3.2  # For Anthropic (Legacy stanford-online-all-v4-s3)
+
+    # Models and Metrics Extras
+    transformers~=4.28.1  # For anthropic_client, huggingface_client, huggingface_tokenizer, test_openai_token_cost_estimator, model_summac (via summarization_metrics)
+    # TODO: Upgrade torch
+    torch~=1.12.1  # For huggingface_client, yalm_tokenizer, model_summac (via summarization_metrics)
+    torchvision~=0.13.1  # For huggingface_client, yalm_tokenizer, model_summac (via summarization_metrics)
+
+    # Metrics Extras
+    google-api-python-client~=2.64.0  # For perspective_api_client via toxicity_metrics
+    numba~=0.56.4  # For copyright_metrics
+    pytrec_eval==0.5  # For ranking_metrics
+    sacrebleu~=2.2.1  # For disinformation_metrics, machine_translation_metrics
+    summ-eval~=0.892  # For summarization_metrics
+
+    # Human Evaluation Extras
+    scaleapi~=2.13.0
     surge-api~=1.1.0
-    # End users should install a CUDA version of PyTorch manually if needed
-    torch~=1.12.1  # Summarization metrics
-    torchvision~=0.13.1
 
-    # plotting
+    # Plots Extras
     colorcet~=3.0.1
     matplotlib~=3.6.0
-    numpy~=1.23.3
     seaborn~=0.11.0
 
 [options.entry_points]