make sure to download fixtures for kd test (axolotl-ai-cloud#2541)

* make sure to download fixtures for kd test

* use same alpaca dataset

Author: winglian

tests/conftest.py

@@ -193,6 +193,14 @@ def download_tiny_shakespeare_dataset():
     snapshot_download_w_retry("winglian/tiny-shakespeare", repo_type="dataset")
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_evolkit_kd_sample_dataset():
+    # download the dataset
+    snapshot_download_w_retry(
+        "axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample", repo_type="dataset"
+    )
+
+
 @pytest.fixture(scope="session", autouse=True)
 def download_deepseek_model_fixture():
     snapshot_download_w_retry("axolotl-ai-co/DeepSeek-V3-11M", repo_type="model")
@@ -208,6 +216,16 @@ def download_huggyllama_model_fixture():
     )
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_llama33_70b_model_fixture():
+    # download the tokenizer only
+    snapshot_download_w_retry(
+        "axolotl-ai-co/Llama-3.3-70B-Instruct-tokenizer",
+        repo_type="model",
+        allow_patterns=["*token*", "config.json"],
+    )
+
+
 @pytest.fixture(scope="session", autouse=True)
 def download_llama_1b_model_fixture():
     # download the tokenizer only
@@ -315,6 +333,14 @@ def download_llama2_model_fixture():
     )
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_llama32_1b_model_fixture():
+    snapshot_download_w_retry(
+        "osllmai-community/Llama-3.2-1B",
+        repo_type="model",
+    )
+
+
 @pytest.fixture
 @enable_hf_offline
 def tokenizer_huggyllama(

tests/e2e/multigpu/solo/__init__.py

@@ -0,0 +1,2 @@
+# Tests under this directory should get run "solo" on their own as they
+# seem to cause issues when run in the same batch as other tests.

tests/e2e/multigpu/solo/test_flex.py

@@ -49,8 +49,9 @@ def test_loss_llama(self, temp_dir):
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 1,

tests/e2e/patched/test_resume.py

@@ -46,8 +46,9 @@ def test_resume_lora_packed(self, temp_dir):
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 2,

tests/e2e/solo/test_flex.py

@@ -41,8 +41,9 @@ def test_loss_llama(self, temp_dir):
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 1,

tests/e2e/test_packing_loss.py

@@ -40,8 +40,9 @@ def test_loss_packed(self, temp_dir):
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 1,