smoke-reference.yml

name: Smoke Reference

# Runs the three reference-model smoke tests tagged `@Tag("smoke-reference")`:
#  - Qwen3-1.7B Q8_0 GGUF        (kllama runner)
#  - Gemma-4 E2B SafeTensors     (kgemma runner)
#  - Gemma-4 E2B GGUF golden-token + timed smoke (kgemma runner, optional input)
#  - BERT + LEAF SafeTensors     (llm-test-java)
#
# By default each test self-skips when its model artifact is not resolvable
# through the standard env-var / `~/.lmstudio/models/` / `~/.cache/huggingface/hub/`
# fallback chain, so this workflow is **green with empty inputs** — it merely
# proves the wiring compiles and the JUnit filter resolves the smoke tier.
#
# To make a run actually exercise the models, trigger it with the matching
# workflow inputs (URLs / paths). The job downloads each artifact, sets the
# corresponding env var the test reads (see Qwen3ReferenceSmokeTest.kt,
# Gemma4ReferenceSmokeTest.kt, BertLeafReferenceSmokeTest.java), and runs
# the same `./gradlew test -PsmokeReference -PincludeIntegration` invocation
# that's documented in the repo's CHANGELOG and reference-smoke tests.
#
# Trigger pattern is manual (`workflow_dispatch`) — wiring this onto every
# push would silently consume Actions minutes without doing meaningful work
# until artifacts are available. A self-hosted runner with the three
# checkpoints pre-cached on disk is the natural place to flip this to
# `push: branches: [develop]` later.

on:
  workflow_dispatch:
    inputs:
      qwen3_gguf_url:
        description: "Direct URL to Qwen3-1.7B-Q8_0.gguf (~1.9 GB). Leave blank to skip the kllama test."
        required: false
        default: ""
      qwen25_gguf_url:
        description: "Direct URL to qwen2.5-0.5b-instruct-q8_0.gguf (~0.5 GB). Enables the Qwen2.5 golden-token parity gate (the #352 bias-path regression gate). Leave blank to skip."
        required: false
        default: ""
      apertus_gguf_url:
        description: "Direct URL to Apertus-8B-Instruct-2509-Q4_K_S.gguf (~4.6 GB). Enables the Apertus golden-token parity gate (QK-norm + xIELU + ungated FFN). Leave blank to skip."
        required: false
        default: ""
      gemma3n_gguf_url:
        description: "Direct URL to gemma-3n-E2B-it-Q4_K_M.gguf (~3.0 GB). Enables the Gemma 3n golden-token parity gate on the DSL lane (AltUp + Laurel + sparsity + PLE + shared KV; needs a large-memory runner: 20g test heap). Leave blank to skip."
        required: false
        default: ""
      gemma4_safetensors_dir_url:
        description: "Direct URL to a tar.gz containing the Gemma-4 E2B SafeTensors checkpoint directory. Leave blank to skip the kgemma test."
        required: false
        default: ""
      gemma4_gguf_url:
        description: "Direct URL to gemma-4-E2B-it-Q4_K_M.gguf (~3.1 GB). Enables the GGUF golden-token + timed smoke tests (needs a large-memory runner: the probe wants ~24g test heap). Leave blank to skip."
        required: false
        default: ""
      leaf_safetensors_dir_url:
        description: "Direct URL to a tar.gz containing the MongoDB mdbr-leaf-ir SafeTensors checkpoint directory. Leave blank to skip the BERT+LEAF test."
        required: false
        default: ""

concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

jobs:
  smoke-reference:
    runs-on: ubuntu-latest
    timeout-minutes: 90

    steps:
      - name: Checkout
        uses: actions/checkout@v7

      - name: Copy CI gradle.properties
        run: mkdir -p ~/.gradle ; cp .github/ci-gradle.properties ~/.gradle/gradle.properties

      - name: Set up JDK 25
        uses: actions/setup-java@v6.0.1
        with:
          distribution: 'zulu'
          java-version: 25

      - name: Disk space (before downloads)
        run: df -h

      - name: Stage Qwen3-1.7B Q8_0 GGUF
        if: inputs.qwen3_gguf_url != ''
        env:
          URL: ${{ inputs.qwen3_gguf_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/qwen3"
          curl -fsSL "$URL" -o "$RUNNER_TEMP/models/qwen3/Qwen3-1.7B-Q8_0.gguf"
          echo "QWEN3_1B7_MODEL_PATH=$RUNNER_TEMP/models/qwen3/Qwen3-1.7B-Q8_0.gguf" >> "$GITHUB_ENV"
          # Same file also feeds the Qwen3 golden-token parity gate.
          echo "QWEN3_17B_GGUF=$RUNNER_TEMP/models/qwen3/Qwen3-1.7B-Q8_0.gguf" >> "$GITHUB_ENV"

      - name: Stage Qwen2.5-0.5B-Instruct Q8_0 GGUF
        if: inputs.qwen25_gguf_url != ''
        env:
          URL: ${{ inputs.qwen25_gguf_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/qwen25"
          curl -fsSL "$URL" -o "$RUNNER_TEMP/models/qwen25/qwen2.5-0.5b-instruct-q8_0.gguf"
          echo "QWEN25_05B_GGUF=$RUNNER_TEMP/models/qwen25/qwen2.5-0.5b-instruct-q8_0.gguf" >> "$GITHUB_ENV"

      - name: Stage Apertus-8B-Instruct Q4_K_S GGUF
        if: inputs.apertus_gguf_url != ''
        env:
          URL: ${{ inputs.apertus_gguf_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/apertus"
          curl -fsSL "$URL" -o "$RUNNER_TEMP/models/apertus/Apertus-8B-Instruct-2509-Q4_K_S.gguf"
          echo "APERTUS_GGUF_PATH=$RUNNER_TEMP/models/apertus/Apertus-8B-Instruct-2509-Q4_K_S.gguf" >> "$GITHUB_ENV"
          # The 8B parity gate needs more than the module's 6g default test heap.
          echo "APERTUS_HEAP_ARG=-PapertusTestMaxHeap=12g" >> "$GITHUB_ENV"

      - name: Stage Gemma 3n E2B GGUF
        if: inputs.gemma3n_gguf_url != ''
        env:
          URL: ${{ inputs.gemma3n_gguf_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/gemma3n"
          curl -fsSL "$URL" -o "$RUNNER_TEMP/models/gemma3n/gemma-3n-E2B-it-Q4_K_M.gguf"
          echo "GEMMA3N_E2B_GGUF=$RUNNER_TEMP/models/gemma3n/gemma-3n-E2B-it-Q4_K_M.gguf" >> "$GITHUB_ENV"
          # The E2B parity gate self-skips below 16 GB test heap.
          echo "GEMMA3N_HEAP_ARG=-PgemmaTestMaxHeap=20g" >> "$GITHUB_ENV"

      - name: Stage Gemma-4 E2B SafeTensors
        if: inputs.gemma4_safetensors_dir_url != ''
        env:
          URL: ${{ inputs.gemma4_safetensors_dir_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/gemma4"
          curl -fsSL "$URL" | tar -xz -C "$RUNNER_TEMP/models/gemma4"
          echo "GEMMA4_E2B_SAFETENSORS_PATH=$RUNNER_TEMP/models/gemma4" >> "$GITHUB_ENV"

      - name: Stage Gemma-4 E2B GGUF
        if: inputs.gemma4_gguf_url != ''
        env:
          URL: ${{ inputs.gemma4_gguf_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/gemma4-gguf"
          curl -fsSL "$URL" -o "$RUNNER_TEMP/models/gemma4-gguf/gemma-4-E2B-it-Q4_K_M.gguf"
          echo "GEMMA4_E2B_GGUF_PATH=$RUNNER_TEMP/models/gemma4-gguf/gemma-4-E2B-it-Q4_K_M.gguf" >> "$GITHUB_ENV"
          echo "GEMMA4_PROBE=1" >> "$GITHUB_ENV"

      - name: Stage MongoDB LEAF SafeTensors
        if: inputs.leaf_safetensors_dir_url != ''
        env:
          URL: ${{ inputs.leaf_safetensors_dir_url }}
        run: |
          set -euo pipefail
          mkdir -p "$RUNNER_TEMP/models/leaf"
          curl -fsSL "$URL" | tar -xz -C "$RUNNER_TEMP/models/leaf"
          echo "LEAF_MODEL_DIR=$RUNNER_TEMP/models/leaf" >> "$GITHUB_ENV"

      - name: Run smoke-reference tier
        env:
          GRADLE_OPTS: -Dorg.gradle.jvmargs="-Xmx4g -Dfile.encoding=UTF-8"
        run: |
          ./gradlew --no-daemon --stacktrace \
            -Dorg.gradle.caching=true \
            -Dorg.gradle.configuration-cache=true \
            -PsmokeReference -PincludeIntegration \
            ${APERTUS_HEAP_ARG:-} \
            ${GEMMA3N_HEAP_ARG:-} \
            test

      - name: Disk space (after run)
        if: always()
        run: df -h || true

      - name: Memory info (on failure)
        if: failure()
        run: |
          free -h || true
          cat /proc/meminfo | head -n 50 || true

      - name: Upload smoke-reference test reports
        if: always()
        uses: actions/upload-artifact@v7
        with:
          name: smoke-reference-reports
          path: |
            **/build/reports/tests/**
            **/build/test-results/**
          retention-days: 14