fix(test): replace jackfram llama with smollm

tests/conftest.py

@@ -47,9 +47,9 @@ def download_smollm2_135m_model():
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_llama_68m_random_model():
+def download_smollm2_135m_instruct_model():
     # download the model
-    snapshot_download_w_retry("JackFram/llama-68m")
+    snapshot_download_w_retry("HuggingFaceTB/SmolLM2-135M-Instruct")
 
 
 @pytest.fixture(scope="session", autouse=True)

tests/e2e/patched/test_4d_multipack_llama.py

@@ -28,7 +28,7 @@ class Test4dMultipackLlama(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
                 "flash_attention": False,
                 "sdp_attention": True,
                 "sample_packing": True,
@@ -72,7 +72,7 @@ class Test4dMultipackLlama(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
                 "flash_attention": False,
                 "sdp_attention": False,
                 "sample_packing": True,

tests/e2e/patched/test_fused_llama.py

@@ -32,7 +32,7 @@ class TestFusedLlama(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
                 "flash_attention": True,
                 "pad_to_sequence_len": True,
                 "flash_attn_fuse_qkv": True,

tests/e2e/patched/test_llama_s2_attention.py

@@ -31,8 +31,7 @@ class TestLlamaShiftedSparseAttention(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 16384,
                 "sample_packing": False,
                 "flash_attention": True,
@@ -77,8 +76,7 @@ class TestLlamaShiftedSparseAttention(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 16384,
                 "sample_packing": False,
                 "flash_attention": True,

tests/e2e/patched/test_lora_llama_multipack.py

@@ -31,8 +31,7 @@ class TestLoraLlama(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "sample_packing": True,
                 "flash_attention": True,
@@ -43,6 +42,7 @@ class TestLoraLlama(unittest.TestCase):
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "val_set_size": 0.2,
+                "lora_modules_to_save": ["lm_head", "embed_tokens"],
                 "special_tokens": {
                     "unk_token": "<unk>",
                     "bos_token": "<s>",

tests/e2e/test_dpo.py

@@ -31,8 +31,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -77,8 +76,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -124,8 +122,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -172,8 +169,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -218,8 +214,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -264,8 +259,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -314,8 +308,7 @@ class TestDPOLlamaLora(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",

tests/e2e/test_llama.py

@@ -26,8 +26,7 @@ class TestLlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "trust_remote_code": True,
                 "sequence_len": 512,
                 "val_set_size": 0.1,

tests/e2e/test_load_model.py

@@ -26,9 +26,8 @@ class TestLoadModelUtils:
         # load config
         self.cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
+                "tokenizer_config": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_config": "JackFram/llama-68m",
                 "sequence_len": 1024,
                 "load_in_8bit": False,
                 "adapter": "lora",

tests/e2e/test_lora_llama.py

@@ -28,8 +28,7 @@ class TestLoraLlama(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -37,6 +36,7 @@ class TestLoraLlama(unittest.TestCase):
                 "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
+                "lora_modules_to_save": ["lm_head", "embed_tokens"],
                 "val_set_size": 0.1,
                 "special_tokens": {
                     "unk_token": "<unk>",

tests/e2e/test_optimizers.py

@@ -28,8 +28,7 @@ class TestCustomOptimizers(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -74,8 +73,7 @@ class TestCustomOptimizers(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",

tests/test_normalize_config.py

@@ -16,9 +16,8 @@ class NormalizeConfigTestCase(unittest.TestCase):
     def _get_base_cfg(self):
         return DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "base_model_config": "JackFram/llama-68m",
+                "base_model_config": "HuggingFaceTB/SmolLM2-135M-Instruct",
-                "tokenizer_type": "LlamaTokenizer",
                 "num_epochs": 1,
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": 1,

tests/utils/test_models.py

@@ -18,9 +18,8 @@ class TestModelsUtils:
         # load config
         self.cfg = DictDefault(  # pylint: disable=attribute-defined-outside-init
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
                 "model_type": "LlamaForCausalLM",
-                "tokenizer_type": "LlamaTokenizer",
                 "load_in_8bit": True,
                 "load_in_4bit": False,
                 "adapter": "lora",