Set sharktank package deps during package build reliably. (#440)

Progress on #294.

We could mirror requirements between `requirements.txt` and
`pyproject.toml` or keep them dynamic. For packaging, we probably don't
want to set explicit versions unless absolutely required. Developers may
want to pin to a specific version to make testing consistent though.

As part of this I added a dep on `iree-turbine` to sharktank's
`requirements.txt` and then switched workflows and user guides to
install `sharktank` first and _then_ install a newer/local version from
git if desired.

---------

Co-authored-by: Marius Brehler <marius.brehler@gmail.com>

.github/workflows/ci-llama-large-tests.yaml

@@ -62,9 +62,11 @@ jobs:
           # from non default locations first. Installing the PyTorch CPU
           # wheels saves multiple minutes and a lot of bandwidth on runner setup.
           pip install --no-compile -r pytorch-cpu-requirements.txt
+          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
+
+          # Install latest iree-tubrine.
           pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
             -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
-          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
 
           # Test with pinned nightly releases, not what iree-turbine uses.
           pip install -f https://iree.dev/pip-release-links.html --upgrade \

.github/workflows/ci-llama-quick-tests.yaml

@@ -63,9 +63,11 @@ jobs:
           # from non default locations first. Installing the PyTorch CPU
           # wheels saves multiple minutes and a lot of bandwidth on runner setup.
           pip install --no-compile -r pytorch-cpu-requirements.txt
+          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
+
+          # Install latest iree-tubrine.
           pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
             -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
-          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
 
           # Test with pinned nightly releases, not what iree-turbine uses.
           pip install -f https://iree.dev/pip-release-links.html --upgrade \

.github/workflows/ci-shark-platform.yml

@@ -58,11 +58,13 @@ jobs:
           # from non default locations first. Installing the PyTorch CPU
           # wheels saves multiple minutes and a lot of bandwidth on runner setup.
           pip install --no-compile -r pytorch-cpu-requirements.txt
+          pip install --no-compile -r requirements.txt -e sharktank/ shortfin/
+
+          # Install latest iree-tubrine.
           pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
             -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
-          pip install --no-compile -r requirements.txt -e sharktank/ shortfin/
 
-          # Try with the latest nightly releases, not what iree-turbine uses.
+          # Try with the latest IREE nightly releases, not what iree-turbine pins.
           # We could also pin to a known working or stable version.
           # This should eventually stabilize. Do the best we can for now.
           pip install -f https://iree.dev/pip-release-links.html --upgrade \

.github/workflows/ci-sharktank.yml

@@ -59,10 +59,13 @@ jobs:
           # from non default locations first. Installing the PyTorch CPU
           # wheels saves multiple minutes and a lot of bandwidth on runner setup.
           pip install --no-compile -r pytorch-cpu-requirements.txt
-          pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
-            -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
           pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
 
+          # Update to the latest iree packages.
+          pip install -f https://iree.dev/pip-release-links.html --upgrade \
+            iree-compiler iree-runtime --src deps \
+            -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
+
       - name: Run sharktank tests
         if: ${{ !cancelled() }}
         run: |

.github/workflows/ci_eval.yaml

@@ -60,17 +60,20 @@ jobs:
           # from non default locations first. Installing the PyTorch CPU
           # wheels saves multiple minutes and a lot of bandwidth on runner setup.
           pip install --no-compile -r pytorch-cpu-requirements.txt
+          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
+
+          # Install latest iree-tubrine.
           pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
             -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
-          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
 
-          # Try with the latest nightly releases, not what iree-turbine uses.
+          # Try with the latest IREE nightly releases, not what iree-turbine pins.
           # We could also pin to a known working or stable version.
           # This should eventually stabilize. Do the best we can for now.
           pip install -f https://iree.dev/pip-release-links.html --upgrade \
             iree-base-compiler \
             iree-base-runtime \
             "numpy<2.0"
+
       - name: Run perplexity test with vmfb
         run:  pytest -n 8 -v -s sharktank/tests/evaluate/perplexity_vmfb_test.py --longrun --iree-device='hip://7' --iree-hip-target=gfx942 --iree-hal-target-backends=rocm --llama3-8b-f16-model-path=/data/llama3.1/8b/llama8b_f16.irpa --llama3-8b-tokenizer-path=/data/llama3.1/8b/tokenizer_config.json
 
@@ -113,9 +116,11 @@ jobs:
           # from non default locations first. Installing the PyTorch CPU
           # wheels saves multiple minutes and a lot of bandwidth on runner setup.
           pip install --no-compile -r pytorch-cpu-requirements.txt
+          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
+
+          # Install latest iree-tubrine.
           pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
             -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
-          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
 
       - name: Run perplexity test in eager mode
         run:  pytest -n 8 -v -s sharktank/tests/evaluate/perplexity_torch_test.py --longrun --llama3-8b-f16-model-path=/data/llama3.1/8b/llama8b_f16.irpa --llama3-8b-tokenizer-path=/data/llama3.1/8b/tokenizer_config.json

README.md

@@ -69,7 +69,7 @@ We recommend setting up a virtual environment (venv). The project is configured
 to ignore `.venv` directories, and editors like VSCode pick them up by default.
 
 ```
-python -m venv --prompt sharktank .venv
+python -m venv .venv
 source .venv/bin/activate
 ```
 
@@ -94,12 +94,12 @@ pip install -r pytorch-rocm-requirements.txt
 ### Install development packages
 
 ```
-# Clone and install editable iree-turbine dep in deps/
-pip install -f https://iree.dev/pip-release-links.html --src deps \
-  -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
-
 # Install editable local projects.
 pip install -r requirements.txt -e sharktank/ shortfin/
+
+# Optionally clone and install editable iree-turbine dep in deps/
+pip install -f https://iree.dev/pip-release-links.html --src deps \
+  -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
 ```
 
 ### Running tests

docs/nightly_releases.md

@@ -69,11 +69,6 @@ source 3.11.venv/bin/activate
 # Install 'sharktank' package from nightly releases.
 python -m pip install sharktank -f https://github.com/nod-ai/SHARK-Platform/releases/expanded_assets/dev-wheels
 
-# Install some other dependencies.
-# TODO(#294): list these as "dependencies" in `pyproject.toml` or make optional?
-python -m pip install gguf numpy huggingface-hub transformers datasets \
-  sentencepiece
-
 # Test the installation.
 python -c "from sharktank import ops; print('Sanity check passed')"
 

sharktank/README.md

@@ -50,7 +50,15 @@ python -m sharktank.tools.dump_gguf --hf-dataset=open_llama_3b_v2_f16_gguf
 * To build wheels for Linux:
 
     ```bash
-    sudo ./build_tools/build_linux_package.sh
+    ./build_tools/build_linux_package.sh
+    ```
+
+    That should produce
+    `build_tools/wheelhouse/sharktank-{X.Y.Z}.dev0-py3-none-any.whl`, which can
+    then be installed with
+
+    ```bash
+    python3 -m pip install build_tools/wheelhouse/sharktank-{X.Y.Z}.dev0-py3-none-any.whl
     ```
 
 * To build a wheel for your host OS/arch manually:

sharktank/build_tools/build_linux_package.sh

@@ -8,7 +8,13 @@
 
 # build_linux_package.sh
 #
-# Builds sharktank Python package for Linux
+# Builds sharktank Python package for Linux.
+#
+# Note: requires a modern Python (3.12+ seems to work). Troubleshooting help:
+#   * https://stackoverflow.com/a/77284076
+#   * https://stackoverflow.com/a/77364602
+# Older versions like 3.10 don't include the package name and set as UNKNOWN?
+#   * Might just need some local packages updated?
 #
 # Usage:
 #   ./build_tools/build_linux_package.sh

sharktank/pyproject.toml

@@ -19,11 +19,9 @@ classifiers = [
   "Programming Language :: Python :: 3.13",
 ]
 requires-python = ">= 3.11"
-dependencies = ["iree-turbine"]
-dynamic = ["version"]  # the version is set via the `setup.py`
 
-[project.optional-dependencies]
-testing = ["pytest"]
+# Version is set via the `setup.py` and requirements are set via files below.
+dynamic = ["version", "dependencies", "optional-dependencies"]
 
 [project.urls]
 Repository = "https://github.com/nod-ai/SHARK-Platform"
@@ -36,8 +34,11 @@ namespaces = true
 [tool.setuptools.package-data]
 sharktank = ["py.typed", "kernels/templates/*.mlir"]
 
-[tool.setuptools.dynamic]
-dependencies = {file = ["requirements.txt"]}
+[tool.setuptools.dynamic.dependencies]
+file = ["requirements.txt"]
+
+[tool.setuptools.dynamic.optional-dependencies]
+testing = {file = ["requirements-tests.txt"]}
 
 [tool.pytest.ini_options]
 addopts = [

sharktank/requirements.txt

@@ -1,7 +1,12 @@
+iree-turbine
+
 # Runtime deps.
 gguf==0.6.0
 numpy==1.26.3
 
+# Needed for newer gguf versions (TODO: remove when gguf package includes this)
+# sentencepiece>=0.1.98,<=0.2.0
+
 # Model deps.
 huggingface-hub==0.22.2
 transformers==4.40.0