[Model] Enable Llama xfail inference on CI

- Disable TVM param conversion in order to reduce DRAM memory usage
- This behaiviour disables TVM const prop and verificaiton, but itn't a blocker for functional model
- Compile requirement is reduced to around 25 GB, which is enough to fit below 32 GB host memory on CI

Closes #247

forge/test/mlir/llama/test_llama_inference.py

@@ -10,7 +10,7 @@
 from test.mlir.llama.utils.utils import load_model
 
 
-@pytest.mark.xfail(reason="Tile broadcast op is not supported on MLIR.")
+@pytest.mark.xfail()
 def test_llama_inference():
     # Load Llama 3B model and tokenizer
     model_path = "openlm-research/open_llama_3b"
@@ -27,6 +27,7 @@ def test_llama_inference():
     # Compile the model
     compiled_model = forge.compile(framework_model, input_ids)
 
+
 @pytest.mark.skip(reason="No need to run in CI, this is PoC that should be mapped to work on device.")
 def test_llama_inference_no_cache_cpu():
     """

forge/test/mlir/llama/utils/utils.py

@@ -9,7 +9,12 @@
 def load_model(model_path="openlm-research/open_llama_3b", use_cache=False):
     # Compiler configurations
     compiler_cfg = forge.config._get_global_compiler_config()
+
+    # Disable CPU fallback, we want to run whole model on device
     compiler_cfg.enable_tvm_cpu_fallback = False
+    # Reduce compile memory usage, but disables TVM verification
+    # and TVM constant evaluation (Forge const eval is enabled)
+    compiler_cfg.convert_framework_params_to_tvm = False
 
     # Load Llama 3B model
     config = LlamaConfig()
@@ -23,5 +28,5 @@ def load_model(model_path="openlm-research/open_llama_3b", use_cache=False):
         model_path, device_map="auto", config=config
     )
     framework_model.eval()
-    
+
     return framework_model