Add workflow, fixes for GPU test pipeline (run-llama#103)

* Add workflow for triggering BuildKite tests. Fixed test markers. Updated dev requirements.

* Unpin torch.

* Update requirements-dev.txt

Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>

* Rename GPU test workflow.

* Update .github/workflows/test_gpu.yml

Co-authored-by: Basile Dura <bdura@users.noreply.github.com>

---------

Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Co-authored-by: Basile Dura <bdura@users.noreply.github.com>

.github/workflows/test_gpu.yml

@@ -0,0 +1,19 @@
+name: GPU tests
+
+on:
+  schedule:
+    - cron: "0 0 * * *"
+  workflow_dispatch:
+
+jobs:
+  run:
+    if: github.repository_owner == 'explosion'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Trigger buildkite build
+        uses: buildkite/trigger-pipeline-action@v1.2.0
+        env:
+          PIPELINE: explosion-ai/spacy-llm-gpu-tests
+          BRANCH: main
+          MESSAGE: ":github: spacy-llm GPU tests - triggered from a GitHub Action"
+          BUILDKITE_API_ACCESS_TOKEN: ${{ secrets.BUILDKITE_SECRET }}

requirements-dev.txt

@@ -8,4 +8,7 @@ black==22.3.0
 types-requests==2.28.11.16
 # Prompting libraries needed for testing
 langchain>=0.0.144,<0.1; python_version>="3.9"
-minichain>=0.3,<0.4; python_version>="3.8" and python_version<"3.11"
+minichain>=0.3,<0.4; python_version>="3.8" and python_version<"3.11"
+# Necessary for running all local models on GPU.
+transformers>=4.0.0,<5.0.0
+torch

spacy_llm/tests/pipeline/test_llm.py

@@ -10,6 +10,8 @@
 from spacy_llm.pipeline import LLMWrapper
 from spacy_llm.registry import registry
 
+from ..compat import has_openai_key
+
 
 @pytest.fixture
 def noop_config() -> Dict[str, Any]:
@@ -47,7 +49,7 @@ def test_llm_pipe_empty(nlp):
 
 def test_llm_serialize_bytes():
     llm = LLMWrapper(
-        task=NoopTask,
+        task=NoopTask(),
         backend=None,  # type: ignore
         cache={"path": None, "batch_size": 0, "max_batches_in_mem": 0},
         vocab=None,  # type: ignore
@@ -57,7 +59,7 @@ def test_llm_serialize_bytes():
 
 def test_llm_serialize_disk():
     llm = LLMWrapper(
-        task=NoopTask,
+        task=NoopTask(),
         backend=None,  # type: ignore
         cache={"path": None, "batch_size": 0, "max_batches_in_mem": 0},
         vocab=None,  # type: ignore
@@ -68,6 +70,7 @@ def test_llm_serialize_disk():
         llm.from_disk(tmp_dir / "llm")
 
 
+@pytest.mark.skipif(has_openai_key is False, reason="OpenAI API key not available")
 @pytest.mark.external
 def test_type_checking_valid(noop_config) -> None:
     """Test type checking for consistency between functions."""

spacy_llm/tests/tasks/test_ner.py

@@ -73,6 +73,7 @@ def fewshot_cfg_string():
     """
 
 
+@pytest.mark.external
 @pytest.mark.skipif(has_openai_key is False, reason="OpenAI API key not available")
 @pytest.mark.parametrize("cfg_string", ["fewshot_cfg_string", "zeroshot_cfg_string"])
 def test_ner_config(cfg_string, request):
@@ -82,7 +83,8 @@ def test_ner_config(cfg_string, request):
     assert nlp.pipe_names == ["llm"]
 
 
-# @pytest.mark.external
+@pytest.mark.external
+@pytest.mark.skipif(has_openai_key is False, reason="OpenAI API key not available")
 @pytest.mark.parametrize("cfg_string", ["zeroshot_cfg_string", "fewshot_cfg_string"])
 def test_ner_predict(cfg_string, request):
     """Use OpenAI to get zero-shot NER results.