improve installation experience (#666)

MANIFEST.in

@@ -1,3 +1,4 @@
 recursive-include exllamav2 *
+global-include *.typed
 global-exclude *.pyc
 global-exclude dni_*

setup.py

@@ -1,15 +1,23 @@
 from setuptools import setup, Extension
-from torch.utils import cpp_extension
-from torch import version as torch_version
+import importlib.util
 import os
 
+if torch := importlib.util.find_spec("torch") is not None:
+    from torch.utils import cpp_extension
+    from torch import version as torch_version
+
 extension_name = "exllamav2_ext"
-verbose = False
-ext_debug = False
+precompile = "EXLLAMA_NOCOMPILE" not in os.environ
+verbose = "EXLLAMA_VERBOSE" in os.environ
+ext_debug = "EXLLAMA_EXT_DEBUG" in os.environ
 
-precompile = 'EXLLAMA_NOCOMPILE' not in os.environ
+if precompile and not torch:
+    print(
+        "cannot precompile unless torch is installed \
+To explicitly JIT install run EXLLAMA_NOCOMPILE= pip install <xyz>"
+    )
 
-windows = (os.name == "nt")
+windows = os.name == "nt"
 
 extra_cflags = ["/Ox"] if windows else ["-O3"]
 
@@ -18,7 +26,7 @@ if ext_debug:
 
 extra_cuda_cflags = ["-lineinfo", "-O3"]
 
-if torch_version.hip:
+if torch and torch_version.hip:
     extra_cuda_cflags += ["-DHIPBLAS_USE_HIP_HALF"]
 
 extra_compile_args = {
@@ -26,7 +34,8 @@ extra_compile_args = {
     "nvcc": extra_cuda_cflags,
 }
 
-setup_kwargs = {
+setup_kwargs = (
+    {
         "ext_modules": [
             cpp_extension.CUDAExtension(
                 extension_name,
@@ -81,9 +90,13 @@ setup_kwargs = {
                 ],
                 extra_compile_args=extra_compile_args,
                 libraries=["cublas"] if windows else [],
-        )],
-    "cmdclass": {"build_ext": cpp_extension.BuildExtension}
-} if precompile else {}
+            )
+        ],
+        "cmdclass": {"build_ext": cpp_extension.BuildExtension},
+    }
+    if precompile and torch
+    else {}
+)
 
 version_py = {}
 with open("exllamav2/version.py", encoding="utf8") as fp:
@@ -120,9 +133,12 @@ setup(
         "websockets",
         "regex",
         "numpy",
-        "rich"
+        "rich",
     ],
     include_package_data=True,
+    package_data={
+        "": ["py.typed"],
+    },
     verbose=verbose,
     **setup_kwargs,
 )