convert thunder types to torch types (#792)

thunder/__init__.py

@@ -47,6 +47,7 @@
     _create_callable,
     trace,
     transform_for_execution,
+    transform_to_torch_types,
 )
 import thunder.extend as extend
 from thunder.extend import Executor, add_default_executor
@@ -76,6 +77,7 @@
 import torch as pytorch
 
 import thunder.clang as clang
+from thunder.core.pytree import tree_flatten, tree_unflatten, tree_map
 
 # Imports executors (to populate default executors and make them accessible)
 import thunder.executors.pythonex
@@ -624,6 +626,7 @@ def get_computation_and_inputs(*args, **kwargs):
                     backward_trc = cudagraphex.fusion_pass(backward_trc, num_static_inputs=len(backward_trc.args[0][0]))
                     backward_traces.append(backward_trc)
 
+            computation_trc = transform_to_torch_types(computation_trc)
             comp = computation_trc.python_callable()
 
             if backward_trc is not None:

thunder/common.py

@@ -44,6 +44,7 @@
 
 import torch as torch
 import numpy as np
+import thunder
 
 #
 # Datastructures for compiled functions
@@ -858,3 +859,20 @@ def _fn(*args, **kwargs) -> tuple[Any, list[TraceCtx]]:
     _fn._using_grad_transform = _using_grad_transform
 
     return _fn
+
+
+def transform_to_torch_types(trace: TraceCtx):
+    # convert the thunder types to torch types if any
+    def map_to_torch(x: Any) -> Any:
+        if isinstance(x, thunder.dtypes.dtype):
+            return thunder.dtypes.to_torch_dtype(x)
+        elif isinstance(x, thunder.devices.Device):
+            return thunder.devices.to_torch_device(x)
+        return x
+
+    last = trace.bound_symbols[-1]
+    assert last.sym.id == prims.PrimIDs.RETURN
+    new_args = tree_map(map_to_torch, last.args)
+    new_bsym = prims.python_return.bind(*new_args, output=())
+    trace.bound_symbols[-1] = new_bsym
+    return trace

thunder/tests/test_core.py

@@ -390,7 +390,7 @@ def foo(a, dev=dev):
     x, y = lc_result
 
     assert x == 1
-    assert y == dev
+    assert y == thunder.devices.to_torch_device(dev)
 
 
 @instantiate(dtypes=(thunder.float32,))
@@ -2969,12 +2969,13 @@ def fn(x):
     jfn = thunder.jit(fn)
     actual_dtype = jfn(x)
 
-    assert actual_dtype == thunder.dtypes.float32
+    assert fn(x) == jfn(x)
+    assert actual_dtype == torch.float32
 
     # Check with a different default dtype.
     with set_default_dtype_ctx(torch.float16):
         actual_dtype = jfn(x)
-        assert actual_dtype == thunder.dtypes.float16
+        assert actual_dtype == torch.float16
 
     assert thunder.cache_misses(jfn) == 2
 
@@ -3002,10 +3003,12 @@ def fn():
         return torch.arange(start=1, end=2, step=0.5).dtype
 
     jfn = thunder.jit(fn)
-    assert jfn() == thunder.dtypes.float32
+    assert fn() == jfn()
+    assert jfn() == torch.float32
 
     def fn():
         return torch.arange(start=1, end=3, step=1).dtype
 
     jfn = thunder.jit(fn)
-    assert jfn() == thunder.dtypes.int64
+    assert fn() == jfn()
+    assert jfn() == torch.int64