Use libdevice file hash rather than path (#3224)

We've been trying to share cached compilations across machines, but
libdevice isn't always in the same place (we actually package it into
python archive files so we don't rely on the system CUDA installation).
By hashing the file instead of its path we can make this hash
consistent; and we really do care more about the contents of libdevice
than its path, anyways.

third_party/nvidia/backend/compiler.py

@@ -53,6 +53,12 @@ def ptx_get_version(cuda_version) -> int:
     raise RuntimeError("Triton only support CUDA 10.0 or higher")
 
 
+@functools.lru_cache(None)
+def file_hash(path):
+    with open(path, "rb") as f:
+        return hashlib.sha256(f.read()).hexdigest()
+
+
 @dataclass(frozen=True)
 class CUDAOptions:
     num_warps: int = 4
@@ -76,8 +82,10 @@ def __post_init__(self):
                "num_warps must be a power of 2"
 
     def hash(self):
-        key = '_'.join([f'{name}-{val}' for name, val in self.__dict__.items()])
-        return hashlib.md5(key.encode("utf-8")).hexdigest()
+        hash_dict = dict(self.__dict__)
+        hash_dict["extern_libs"] = tuple((k, file_hash(v)) for k, v in sorted(hash_dict["extern_libs"]))
+        key = "_".join([f"{name}-{val}" for name, val in sorted(hash_dict.items())])
+        return hashlib.sha256(key.encode("utf-8")).hexdigest()
 
 
 class CUDABackend(BaseBackend):