fix: nonhybrid granite model id (#546)

* fix: granite for non-hybrid model id

* feat: granite4 nano model family (1b, 350m)

* fix: skip alora training test on CICD

Author: guicho271828

mellea/backends/model_ids.py

@@ -45,6 +45,18 @@ class ModelIdentifier:
     watsonx_name="ibm/granite-4-h-small",
 )
 
+IBM_GRANITE_4_HYBRID_1B = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.0-h-1b",
+    ollama_name="granite4:1b-h",
+    watsonx_name=None,
+)
+
+IBM_GRANITE_4_HYBRID_350m = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.0-h-350m",
+    ollama_name="granite4:350m-h",
+    watsonx_name=None,
+)
+
 
 # Deprecated Granite 3 models - kept for backward compatibility
 # These maintain their original model references (not upgraded to Granite 4)
@@ -65,9 +77,9 @@ class ModelIdentifier:
 # - Ollama/HF: Uses MICRO (fits in CI memory constraints)
 # - Watsonx: Uses SMALL (required for watsonx support)
 IBM_GRANITE_4_MICRO_3B = ModelIdentifier(
-    hf_model_name="ibm-granite/granite-4.0-h-micro",
-    ollama_name="granite4:micro-h",
-    watsonx_name="ibm/granite-4-h-small",
+    hf_model_name="ibm-granite/granite-4.0-micro",
+    ollama_name="granite4:micro",
+    watsonx_name="ibm/granite-4-small",
 )
 
 # Granite 3.3 Vision Model (2B)

test/cli/test_alora_train_integration.py

@@ -14,14 +14,24 @@
 import torch
 from transformers import AutoTokenizer
 
+pytestmark = [
+    pytest.mark.huggingface,
+    pytest.mark.llm,
+    pytest.mark.requires_gpu,
+    pytest.mark.requires_heavy_ram,
+    # Skip entire module in CI since 17/18 tests are qualitative
+    pytest.mark.skipif(
+        int(os.environ.get("CICD", 0)) == 1,
+        reason="Skipping alora training tests in CI - need gpus",
+    ),
+]
+
 # Check if MPS is available but PyTorch version is too old
 _mps_needs_cpu_fallback = torch.backends.mps.is_available() and tuple(
     int(x) for x in torch.__version__.split(".")[:2]
 ) < (2, 8)
 
 
-@pytest.mark.huggingface
-@pytest.mark.llm
 def test_alora_training_integration():
     """Integration test: Train a tiny aLoRA adapter and verify it works.
 
@@ -278,8 +288,6 @@ def test_alora_training_integration():
         )
 
 
-@pytest.mark.huggingface
-@pytest.mark.llm
 def test_lora_training_integration():
     """Integration test: Train a tiny standard LoRA adapter and verify it works.