Compile improvements - auto cache size, fix fullgraph setting, fix triton detection (#899)

* Compile improvements - auto cache size, fix fullgraph setting, fix triton detection

* remove forced torchao no longer needed

toolkit/config_modules.py

@@ -712,14 +712,11 @@ class ModelConfig:
 
         if self.compile and self.quantize:
             print("Quantized model detected - allowing torch.compile (experimental)")
-            # make it torchao instead of quantio for compatibility with torch compile
-            if self.qtype == "qfloat8":
-                self.qtype = "float8"
         self.block_compile = kwargs.get("block_compile", False)
         self.compile_mode = kwargs.get("compile_mode", "default")
         self.compile_fullgraph = kwargs.get("compile_fullgraph", False)
         self.compile_dynamic = kwargs.get("compile_dynamic", True)
-        self.cache_size_limit = kwargs.get("cache_size_limit", 8)
+        self.cache_size_limit = kwargs.get("cache_size_limit", None)
         
         # kwargs to pass to the model
         self.model_kwargs = kwargs.get("model_kwargs", {})