Taskfile.yml

version: '3'

# Taskfile.yml - reproducible entrypoints for Codegeist model workflows
#
# Why this exists:
# - Keeps the devcontainer image free of the optional 52-package inference
#   environment while making setup and execution repeatable outside one session.
# - Delegates all behavior to the locked training, inference, and GGUF projects;
#   package pins and model logic remain in their source files.
#
# Side effects:
# - `setup` creates the ignored `jobs/training/inference/.venv` on demand.
# - `infer` downloads the pinned public model and adapter on first execution.
# - `evidence` rewrites generated review artifacts under `docs/evidence/`.
# - `gguf-setup` refreshes the shared locked inference environment and creates
#   the ignored GGUF artifact root on demand.
# - `gguf-build` writes model intermediates and a public promotion tree below
#   the ignored `.artifacts/gguf/` root.
#
# Runtime contract:
# - `infer` runs only in the GPU-enabled project devcontainer configured by
#   `.codegeist/compose.local.yml`; `infer.py` validates loaded model placement
#   and dtype instead of repeating CUDA/BF16 environment availability checks.

vars:
  TRAINING_PROJECT: '{{.ROOT_DIR}}/jobs/training'
  INFERENCE_PROJECT: '{{.ROOT_DIR}}/jobs/training/inference'
  INFERENCE_ENVIRONMENT: '{{.ROOT_DIR}}/jobs/training/inference/.venv'
  TRAINING_PYTHON: '3.12.12'
  TEST_ENVIRONMENT: '/tmp/codegeist-training-test-venv'
  GGUF_PROJECT: '{{.ROOT_DIR}}/jobs/gguf'
  GGUF_CONVERSION_PROJECT: '{{.ROOT_DIR}}/jobs/gguf/conversion'
  GGUF_ENVIRONMENT: '{{.INFERENCE_ENVIRONMENT}}'
  GGUF_OUTPUT_ROOT: '{{.ROOT_DIR}}/.artifacts/gguf'

tasks:
  default:
    desc: List available workflows
    cmds:
      - task --list
    silent: true

  setup:
    desc: Install the locked GPU inference environment on demand
    env:
      UV_LINK_MODE: copy
    cmds:
      - >-
        uv sync
        --project "{{.INFERENCE_PROJECT}}"
        --frozen
        --no-dev
        --python "{{.TRAINING_PYTHON}}"
      - >-
        "{{.INFERENCE_ENVIRONMENT}}/bin/python" -c
        'import torch; assert torch.__version__.startswith("2.6.0"); assert torch.version.cuda == "12.4"; print(f"torch={torch.__version__} cuda={torch.version.cuda}")'

  test:
    desc: Run all weightless contracts and verify every lockfile
    env:
      UV_PROJECT_ENVIRONMENT: '{{.TEST_ENVIRONMENT}}'
    cmds:
      - uv lock --project "{{.TRAINING_PROJECT}}" --check
      - uv lock --project "{{.INFERENCE_PROJECT}}" --check
      - uv lock --project "{{.GGUF_PROJECT}}" --check
      - uv lock --project "{{.GGUF_CONVERSION_PROJECT}}" --check
      - >-
        uv run
        --project "{{.TRAINING_PROJECT}}"
        --frozen
        --only-group dev
        pytest "{{.TRAINING_PROJECT}}/tests"
      - >-
        uv run
        --project "{{.GGUF_PROJECT}}"
        --frozen
        --only-group dev
        pytest "{{.GGUF_PROJECT}}/tests"
      - git diff --check

  evidence:
    desc: Regenerate the curated evidence dashboard and Mermaid provenance
    cmds:
      - python3 "{{.TRAINING_PROJECT}}/render_training_evidence.py"

  infer:
    desc: Set up and run the strict public-adapter GPU verifier
    deps:
      - setup
    cmds:
      - >-
        "{{.INFERENCE_ENVIRONMENT}}/bin/python"
        "{{.TRAINING_PROJECT}}/infer.py"
        {{.CLI_ARGS}}

  gguf-setup:
    desc: Reuse the reviewed CUDA PEFT environment for GGUF merge and publication
    deps:
      - setup
    cmds:
      - mkdir -p "{{.GGUF_OUTPUT_ROOT}}"
      - >-
        "{{.GGUF_ENVIRONMENT}}/bin/python" -c
        'import torch; assert torch.__version__.startswith("2.6.0"); assert torch.version.cuda == "12.4"; print(f"torch={torch.__version__} cuda={torch.version.cuda}")'

  gguf-build:
    desc: Build one token-free merged Q4_K_M GGUF in the GPU devcontainer
    deps:
      - gguf-setup
    env:
      CODEGEIST_GGUF_OUTPUT_ROOT: '{{.GGUF_OUTPUT_ROOT}}'
      HF_HUB_DISABLE_IMPLICIT_TOKEN: '1'
    cmds:
      - >-
        env -u HF_TOKEN
        "{{.GGUF_ENVIRONMENT}}/bin/python"
        "{{.GGUF_PROJECT}}/build.py"
        {{.CLI_ARGS}}

  gguf-compare:
    desc: Require byte-identical Q4_K_M files from two clean builds
    requires:
      vars:
        - BUILD_A
        - BUILD_B
    cmds:
      - >-
        cmp
        "{{.BUILD_A}}/public/gguf/codegeist-llm-Q4_K_M.gguf"
        "{{.BUILD_B}}/public/gguf/codegeist-llm-Q4_K_M.gguf"
      - >-
        sha256sum
        "{{.BUILD_A}}/public/gguf/codegeist-llm-Q4_K_M.gguf"
        "{{.BUILD_B}}/public/gguf/codegeist-llm-Q4_K_M.gguf"

  gguf-publish:
    desc: Publish one reviewed GGUF tree with the runtime Hub credential
    deps:
      - gguf-setup
    env:
      CODEGEIST_GGUF_OUTPUT_ROOT: '{{.GGUF_OUTPUT_ROOT}}'
      HF_HUB_DISABLE_IMPLICIT_TOKEN: '1'
    cmds:
      - >-
        "{{.GGUF_ENVIRONMENT}}/bin/python"
        "{{.GGUF_PROJECT}}/publish.py"
        {{.CLI_ARGS}}

  all:
    desc: Regenerate evidence, run contracts, and execute GPU inference
    cmds:
      - task: evidence
      - task: test
      - task: infer