Update Test Generation Template to account for fp32 inputs

triton_kernel_agent/templates/test_generation.j2

@@ -77,6 +77,7 @@ CRITICAL TEST DATA REQUIREMENTS:
 - Only create variations if the problem description is vague about the input specifications
 - If original problem description specifies FP32, convert to BF16 problems instead of FP32 problems
 - AVOID FP32 for inputs and outputs - use BF16 instead when users specify FP32 problems
+  - EXCEPTION: If the user cast example inputs explicitly to FP32, test with both FP32 and BF16
 - GroupNorm: compute mean/var in fp32 over (C/G)*H*W per (N,group), eps=1e-5; apply affine in fp32, cast at end; no extra bf16 quantization before pool/norm.
 
 
@@ -127,10 +128,10 @@ def test_kernel():
         # Create test data using EXACT specifications from problem description
         # If problem specifies shape/dtype, use those exact values
         # input_tensor = torch.randn(EXACT_SHAPE, dtype=EXACT_DTYPE, device=device)
-        
+
         # Call kernel_function as a normal Python function
         # result = kernel_function(input_tensor, other_args...)
-        
+
         # Example device check (avoid comparing to literal 'cuda')
         # if isinstance(result, torch.Tensor) and result.device != input_tensor.device:
         #     return False
@@ -154,7 +155,7 @@ def test_kernel():
         #     print(f"Max absolute difference: {torch.max(torch.abs(result - expected))}")
         #     print(f"Relative error: {torch.max(torch.abs((result - expected) / (expected + 1e-8)))}")
         #     return False
-        
+
         return True  # if successful
     except Exception as e:
         # Surface undefined helper issues from kernel.py clearly
@@ -170,4 +171,4 @@ if __name__ == "__main__":
     sys.exit(0 if success else 1)
 ```
 
-Generate a complete test implementation: 
+Generate a complete test implementation:

utils/providers/available_models.py

@@ -64,4 +64,9 @@
         provider_classes=[RelayProvider],
         description="Claude 4.5 Opus (Released Nov 2025)",
     ),
+    ModelConfig(
+        name="gpt-5-2",
+        provider_classes=[RelayProvider],
+        description="GPT-5.2 flagship model (Dec 2025) - Note the name is different from the OpenAI model",
+    ),
 ]