[Draft] Fix blogpost code

.bazelignore

@@ -0,0 +1 @@
+.direnv

.gitignore

@@ -1,6 +1,5 @@
 .direnv/
 .pre-commit-config.yaml
-MODULE.bazel.lock
 bazel-*
 compile_commands.json
 lre.bazelrc

finetuning_on_cpu/.bazelrc

@@ -0,0 +1,35 @@
+# Pick up system specific configs.
+common --enable_platform_specific_config
+
+# Don't inherit PATH and LD_LIBRARY_PATH.
+build --incompatible_strict_action_env
+
+# Forbid network access unless explicitly enabled.
+build --sandbox_default_allow_network=false
+
+# Use correct runfile locations.
+build --nolegacy_external_runfiles
+
+# Enable sandboxing for exclusive tests like GPU performance tests.
+test --incompatible_exclusive_test_sandboxed
+
+# Simulate rules_python's initialization with a script and use a statically
+# linked musl Python toolchain.
+common --@rules_python//python/config_settings:bootstrap_impl=script
+common --@rules_python//python/config_settings:py_linux_libc=musl
+# common --@rules_python//python/config_settings:pip_whl_muslc_version=3.13
+
+# Remote optimizations.
+build --remote_download_minimal
+
+# Smaller profiling. Careful. Disabling this might explode remote cache usage.
+build --slim_profile
+build --experimental_profile_include_target_label
+build --noexperimental_profile_include_primary_output
+
+# Use the container image specified in `platforms/BUILD.bazel`.
+build --extra_execution_platforms=@//platforms:python-remote
+build --extra_toolchains=@zig_sdk//toolchain:linux_amd64_musl
+
+# Add user specific flags.
+try-import %workspace%/user.bazelrc

finetuning_on_cpu/.bazelversion

@@ -0,0 +1 @@
+8.2.1

finetuning_on_cpu/.gitignore

@@ -0,0 +1,6 @@
+.DS_Store # only for macs
+bazel-finetune-repo
+bazel-out
+bazel-bin
+bazel-testlogs
+.env

finetuning_on_cpu/BUILD.bazel

@@ -0,0 +1,4 @@
+exports_files(
+    ["python_provider.sh"],
+    visibility = ["//visibility:public"],  # Makes it visible to all packages
+)

finetuning_on_cpu/MODULE.bazel

@@ -0,0 +1,22 @@
+bazel_dep(name = "hermetic_cc_toolchain", version = "4.0.0")
+bazel_dep(name = "platforms", version = "0.0.11")
+bazel_dep(name = "rules_cc", version = "0.1.1")
+bazel_dep(name = "rules_python", version = "1.2.0")
+bazel_dep(name = "rules_shell", version = "0.4.1")
+
+# Used internally by rules_python.
+cc_toolchains = use_extension("@hermetic_cc_toolchain//toolchain:ext.bzl", "toolchains")
+use_repo(cc_toolchains, "zig_sdk")
+
+# Python configuration
+python = use_extension("@rules_python//python/extensions:python.bzl", "python")
+python.toolchain(python_version = "3.13")
+
+# Python packages
+pip = use_extension("@rules_python//python/extensions:pip.bzl", "pip")
+pip.parse(
+    hub_name = "pypi",
+    python_version = "3.13",
+    requirements_lock = "//:bazel_requirements_lock.txt",
+)
+use_repo(pip, "pypi")

finetuning_on_cpu/bazel_requirements_lock.txt

@@ -0,0 +1,173 @@
+
+aiohappyeyeballs==2.6.1
+    # via aiohttp
+aiohttp==3.11.14
+    # via
+    #   datasets
+    #   fsspec
+aiosignal==1.3.2
+    # via aiohttp
+async-timeout==5.0.1
+    # via aiohttp
+attrs==25.3.0
+    # via aiohttp
+certifi==2025.1.31
+    # via requests
+charset-normalizer==3.4.1
+    # via requests
+datasets==3.4.1
+    # via -r requirements.txt
+dill==0.3.7
+    # via
+    #   datasets
+    #   multiprocess
+filelock==3.18.0
+    # via
+    #   huggingface-hub
+    #   torch
+    #   transformers
+frozenlist==1.5.0
+    # via
+    #   aiohttp
+    #   aiosignal
+fsspec[http]==2024.12.0
+    # via
+    #   datasets
+    #   huggingface-hub
+    #   torch
+huggingface-hub==0.29.3
+    # via
+    #   datasets
+    #   tokenizers
+    #   transformers
+idna==3.10
+    # via
+    #   requests
+    #   yarl
+jinja2==3.1.6
+    # via torch
+markupsafe==3.0.2
+    # via jinja2
+mpmath==1.3.0
+    # via sympy
+multidict==6.2.0
+    # via
+    #   aiohttp
+    #   yarl
+multiprocess==0.70.15
+    # via datasets
+networkx==3.2.1
+    # via torch
+numpy==1.26.4
+    # via
+    #   datasets
+    #   pandas
+    #   transformers
+packaging==24.2
+    # via
+    #   datasets
+    #   huggingface-hub
+    #   transformers
+pandas==2.2.3
+    # via datasets
+propcache==0.3.0
+    # via
+    #   aiohttp
+    #   yarl
+pyarrow==19.0.1
+    # via datasets
+python-dateutil==2.9.0.post0
+    # via pandas
+pytz==2025.2
+    # via pandas
+pyyaml==6.0.2
+    # via
+    #   datasets
+    #   huggingface-hub
+    #   transformers
+regex==2024.11.6
+    # via transformers
+requests==2.32.3
+    # via
+    #   datasets
+    #   huggingface-hub
+    #   transformers
+safetensors==0.5.3
+    # via transformers
+six==1.17.0
+    # via python-dateutil
+sympy==1.13.1
+    # via torch
+tokenizers==0.21.1
+    # via transformers
+setuptools==69.0.2
+    # via torch
+torch==2.6.0
+    # via -r requirements.txt
+tqdm==4.67.1
+    # via
+    #   datasets
+    #   huggingface-hub
+    #   transformers
+transformers==4.50.1
+    # via -r requirements.txt
+typing-extensions==4.12.2
+    # via
+    #   huggingface-hub
+    #   multidict
+    #   torch
+tzdata==2025.2
+    # via pandas
+urllib3==2.3.0
+    # via requests
+xxhash==3.5.0
+    # via datasets
+yarl==1.18.3
+    # via aiohttp
+accelerate==1.5.2
+psutil==5.9.8
+    # for accelerate
+anyio==4.9.0
+    # for transformers
+distro==1.9.0
+    # for transformers
+httpx==0.28.1
+    # for transformers
+jiter==0.9.0
+    # for transformers
+pydantic-core==2.33.0
+    # for transformers
+pydantic==2.11.0
+    # for transformers
+sniffio==1.3.1
+    # for httpx
+httpcore==1.0.7
+    # for httpx
+h11==0.14.0
+    # for httpx
+annotated-types==0.7.0
+    # for transformers
+typing-inspection==0.4.0
+    # for transformers
+exceptiongroup==1.2.2
+    # for transformers
+python-dotenv==1.1.0
+    # for loading env variables
+scipy==1.15.2
+    # for torch to get embeddings
+
+# Manual adjusted below.
+nvidia-cublas-cu12==12.9.0.13
+nvidia-cuda-cupti-cu12==12.9.19
+nvidia-cuda-nvrtc-cu12==12.9.41
+nvidia-cuda-runtime-cu12==12.9.37
+nvidia-cudnn-cu12==9.10.0.56
+nvidia-cufft-cu12==11.4.0.6
+nvidia-curand-cu12==10.3.10.19
+nvidia-cusolver-cu12==11.7.4.40
+nvidia-cusparse-cu12==12.5.9.5
+nvidia-cusparselt-cu12==0.7.1
+nvidia-nccl-cu12==2.26.5
+nvidia-nvjitlink-cu12==12.9.41
+nvidia-nvtx-cu12==12.9.19
+triton==3.3.0

finetuning_on_cpu/platforms/BUILD.bazel

@@ -0,0 +1,12 @@
+platform(
+    name = "python-remote",
+    constraint_values = [
+        "@platforms//os:linux",
+        "@platforms//cpu:x86_64",
+    ],
+    exec_properties = {
+        "OSFamily": "Linux",
+        "ISA": "x86-64",
+        "container-image": "docker.io/evaanahmed2001/python-bazel-env@sha256:8de13199d587964b218c0b671272b42031cf4944b2f426e6eee7d7542802bf7c",
+    },
+)

finetuning_on_cpu/python_provider.sh

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+# Generic Python test wrapper that handles cross-architecture execution
+
+# Get the directory where this script is located
+DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+# First argument is the Python module or script to run
+PYTHON_SCRIPT="$1"
+# Remaining arguments are passed to the Python script
+shift
+
+echo "Running Python test: $PYTHON_SCRIPT"
+echo "Python version:"
+python3 --version
+
+# Check if the argument is a file or a module
+if [[ -f "$PYTHON_SCRIPT" ]]; then
+    # Run as a script if it's a file
+    echo "Running as script"
+    python3 "$PYTHON_SCRIPT" "$@"
+else
+    # Run as a module if it's not a file
+    echo "Running as module"
+    python3 -m "$PYTHON_SCRIPT" "$@"
+fi
+
+# Capture and return the exit code
+exit_code=$?
+echo "Python test completed with exit code: $exit_code"
+exit $exit_code

finetuning_on_cpu/src/training/BUILD.bazel

@@ -0,0 +1,27 @@
+load("@rules_python//python:defs.bzl", "py_test")
+
+exports_files(
+    ["train_model.py"],
+    visibility = ["//visibility:public"],  # Makes it visible to all packages
+)
+
+py_test(
+    name = "training_test",
+    size = "large",
+    srcs = ["train_model.py"],
+    main = "train_model.py",
+    # IMPORTANT: Model training is irreproducible. We can use NativeLink to
+    #            speed up the build, but we can't "reuse" a cached model as
+    #            every training run will yield different results. The `no-cache`
+    #            flag here prevents any sort of caching while still allowing for
+    #            remote execution in Nativelink's distributed infrastructure.
+    tags = ["no-cache"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "@pypi//accelerate",
+        "@pypi//datasets",
+        "@pypi//psutil",
+        "@pypi//torch",
+        "@pypi//transformers",
+    ],
+)