documentlanguagemodel Public

Watch 0 Fork 0 Star 0
YAML · 5758 bytes Raw Blame History
1 name: Weekly chat-template drift (Sprint 12.6)
2
3 # Runs the closed-loop Go↔Jinja check every Sunday:
4 # 1. HF side — refresh-chat-template-goldens.py --check asserts no golden
5 # drifted since the last checked-in matrix. Fails fast if an upstream
6 # tokenizer revision silently changed its chat_template.
7 # 2. Go side — install Ollama, build a tiny-model GGUF via vendored
8 # llama.cpp, register it with `ollama create`, then run
9 # tests/integration/export/test_template_closed_loop.py with
10 # OLLAMA_NAME pointing at the registered model. Ollama's
11 # prompt_eval_count must equal the HF golden for every scenario.
12 #
13 # Trigger: weekly cron + workflow_dispatch for operators bumping bases.
14
15 on:
16 schedule:
17 # Sundays at 06:00 UTC — after quieter traffic windows for HF / Ollama pulls.
18 - cron: "0 6 * * 0"
19 workflow_dispatch: {}
20
21 concurrency:
22 group: weekly-template-drift
23 cancel-in-progress: false
24
25 env:
26 UV_VERSION: "0.11.6"
27 PYTHON_VERSION: "3.11"
28 # Pinned to BASE_MODELS["smollm2-135m"].revision (Sprint 06 registry).
29 # Same SHA as ci.yml — keep in sync when bumping.
30 TINY_MODEL_REVISION: "12fd25f77366fa6b3b4b768ec3050bf629380bac"
31
32 jobs:
33 check-hf-side:
34 # Cheap half: no Ollama, no llama.cpp, no quant. If this fails the
35 # whole workflow is done — nothing to verify against.
36 name: HF goldens unchanged
37 runs-on: ubuntu-latest
38 steps:
39 - uses: actions/checkout@v4
40
41 - name: Install uv
42 uses: astral-sh/setup-uv@v4
43 with:
44 version: ${{ env.UV_VERSION }}
45
46 - name: Sync dependencies
47 run: uv sync --all-extras --dev
48
49 - name: Restore HF cache
50 uses: actions/cache@v4
51 with:
52 path: ${{ github.workspace }}/.hf-cache
53 key: hf-tiny-${{ env.TINY_MODEL_REVISION }}-${{ hashFiles('pyproject.toml') }}
54 restore-keys: |
55 hf-tiny-${{ env.TINY_MODEL_REVISION }}-
56
57 - name: Pre-warm tiny model
58 env:
59 HF_HOME: ${{ github.workspace }}/.hf-cache
60 DLM_TINY_MODEL_REVISION: ${{ env.TINY_MODEL_REVISION }}
61 run: |
62 uv run python - <<'PY'
63 from tests.fixtures.tiny_model import tiny_model_path
64 print("tiny model at:", tiny_model_path())
65 PY
66
67 - name: Refresh goldens in --check mode (chatml only — only cached dialect)
68 env:
69 HF_HOME: ${{ github.workspace }}/.hf-cache
70 run: uv run python scripts/refresh-chat-template-goldens.py --check --dialect chatml
71
72 closed-loop:
73 # Expensive half: install Ollama, build a base GGUF, register it,
74 # and assert prompt_eval_count == HF golden.
75 name: Go↔Jinja closed loop (chatml)
76 needs: check-hf-side
77 runs-on: ubuntu-latest
78 steps:
79 - name: Checkout with llama.cpp submodule
80 uses: actions/checkout@v4
81 with:
82 submodules: recursive
83
84 - name: Install uv
85 uses: astral-sh/setup-uv@v4
86 with:
87 version: ${{ env.UV_VERSION }}
88
89 - name: Sync dependencies
90 run: uv sync --all-extras --dev
91
92 - name: Restore HF cache
93 uses: actions/cache@v4
94 with:
95 path: ${{ github.workspace }}/.hf-cache
96 key: hf-tiny-${{ env.TINY_MODEL_REVISION }}-${{ hashFiles('pyproject.toml') }}
97 restore-keys: |
98 hf-tiny-${{ env.TINY_MODEL_REVISION }}-
99
100 - name: Pre-warm tiny model
101 env:
102 HF_HOME: ${{ github.workspace }}/.hf-cache
103 DLM_TINY_MODEL_REVISION: ${{ env.TINY_MODEL_REVISION }}
104 run: |
105 uv run python - <<'PY'
106 from tests.fixtures.tiny_model import tiny_model_path
107 print("tiny model at:", tiny_model_path())
108 PY
109
110 - name: Restore llama.cpp build cache
111 id: llama-cpp-cache
112 uses: actions/cache@v4
113 with:
114 path: vendor/llama.cpp/build
115 key: llama-cpp-build-${{ hashFiles('.gitmodules', 'vendor/llama.cpp/VERSION') }}
116
117 - name: Build llama-quantize (if not cached)
118 if: steps.llama-cpp-cache.outputs.cache-hit != 'true'
119 run: |
120 set -euxo pipefail
121 command -v cmake >/dev/null 2>&1 || sudo apt-get install -y cmake
122 scripts/bump-llama-cpp.sh build
123
124 - name: Install Ollama
125 run: |
126 set -euxo pipefail
127 curl -fsSL https://ollama.com/install.sh | sh
128 # Start the ollama server in the background so `ollama create` /
129 # `ollama run` have something to talk to.
130 ollama serve >/tmp/ollama.log 2>&1 &
131 # Poll for readiness instead of a blind sleep.
132 for i in $(seq 1 30); do
133 if ollama list >/dev/null 2>&1; then
134 echo "ollama ready after ${i}s"
135 break
136 fi
137 sleep 1
138 done
139 ollama --version
140
141 - name: Export tiny model to Ollama (registers under dlm-test-chatml)
142 env:
143 HF_HOME: ${{ github.workspace }}/.hf-cache
144 DLM_TINY_MODEL_REVISION: ${{ env.TINY_MODEL_REVISION }}
145 run: |
146 set -euxo pipefail
147 # Placeholder: the tiny-model export pipeline lands via Sprint 14.5.
148 # Until then the closed-loop job exits 0 after the HF-side check
149 # — the scaffold is in place for the runner to fill.
150 echo "export pipeline TBD — see Sprint 14.5"
151
152 - name: Run closed-loop integration test
153 if: false # enabled once the export step above registers OLLAMA_NAME
154 env:
155 OLLAMA_NAME: "dlm-test-chatml:latest"
156 HF_HOME: ${{ github.workspace }}/.hf-cache
157 DLM_TINY_MODEL_REVISION: ${{ env.TINY_MODEL_REVISION }}
158 run: uv run pytest -m slow -v tests/integration/export/test_template_closed_loop.py
159