[OPT][DNR] Virtual memory

triton_viz/core/config.py

@@ -26,6 +26,13 @@ def reset(self) -> None:
             os.getenv("REPORT_GRID_EXECUTION_PROGRESS", "0") == "1"
         )  # verify using setter
 
+        # --- Virtual memory flag ---
+        self._virtual_memory = os.getenv("TRITON_VIZ_VIRTUAL_MEMORY", "0") == "1"
+
+    @property
+    def virtual_memory(self) -> bool:
+        return self._virtual_memory
+
     # ---------- disable_sanitizer ----------
     @property
     def disable_sanitizer(self) -> bool:

triton_viz/core/patch.py

@@ -5,7 +5,9 @@
 from tqdm import tqdm
 
 from .config import config as cfg
+from dataclasses import dataclass
 from .callbacks import OpCallbacks, ForLoopCallbacks
+
 from .data import (
     Op,
     RawLoad,
@@ -53,6 +55,8 @@
 )
 from triton.runtime.interpreter import _patch_lang as triton_patch_lang
 from triton.runtime.interpreter import ASTTransformer as _OrigASTTransformer
+from triton.runtime.interpreter import _tuple_create, _unwrap_tensor, _rewrap_tensor
+from triton.tools.tensor_descriptor import TensorDescriptor
 from triton.runtime import JITFunction
 
 op_list = [
@@ -476,6 +480,69 @@ def unpatch_lang():
         importlib.reload(tl)
 
 
+@dataclass(frozen=True)
+class FakeTensor:
+    _data_ptr: int
+    dtype: str
+    shape: tuple[int, ...] = ()
+    _stride: tuple[int, ...] = ()
+    _is_contiguous: bool = True
+    _element_size: int = 1
+
+    def data_ptr(self) -> int:
+        return self._data_ptr
+
+    def stride(self) -> tuple[int, ...]:
+        return self._stride
+
+    def is_contiguous(self) -> bool:
+        return self._is_contiguous
+
+    def numel(self) -> int:
+        size = 1
+        for dim in self.shape:
+            size *= dim
+        return size
+
+    def element_size(self) -> int:
+        return self._element_size
+
+
+def _init_args_hst(args_dev, kwargs):
+    def _to_cpu(arg):
+        if isinstance(arg, tuple):
+            return _tuple_create(arg, map(_to_cpu, arg))
+        elif isinstance(arg, TensorDescriptor):
+            return TensorDescriptor(
+                _to_cpu(arg.base),
+                arg.shape,
+                arg.strides,
+                arg.block_shape,
+            )
+        elif not hasattr(arg, "data_ptr"):
+            return arg
+
+        unwrapped_arg = _unwrap_tensor(arg)
+        cpu_arg = FakeTensor(
+            _data_ptr=unwrapped_arg.data_ptr(),
+            dtype=unwrapped_arg.dtype,
+            shape=unwrapped_arg.shape,
+            _stride=unwrapped_arg.stride(),
+            _is_contiguous=unwrapped_arg.is_contiguous(),
+            _element_size=unwrapped_arg.element_size(),
+        )
+        cpu_arg = _rewrap_tensor(cpu_arg, original_tensor=arg)
+        return cpu_arg
+
+    args_hst = [_to_cpu(arg) for arg in args_dev]
+
+    # Process keyword arguments
+    kwargs_hst = {}
+    for key, value in kwargs.items():
+        kwargs_hst[key] = _to_cpu(value)
+    return args_hst, kwargs_hst
+
+
 def _grid_executor_call(self, *args_dev, **kwargs):
     if kwargs.pop("warmup", False):
         return
@@ -517,7 +584,10 @@ def run_grid_loops(grid):
     }
     client_manager = kwargs.pop("client_manager")
     kwargs.pop("jit_fn")
-    args_hst, kwargs_hst = self._init_args_hst(args_dev, kwargs)
+    if cfg.virtual_memory:
+        args_hst, kwargs_hst = _init_args_hst(args_dev, kwargs)
+    else:
+        args_hst, kwargs_hst = self._init_args_hst(args_dev, kwargs)
     # Prepare call arguments
     args = inspect.getcallargs(self.fn, *args_hst, **kwargs_hst)
     call_args = {}
@@ -547,7 +617,8 @@ def run_grid_loops(grid):
         name = self.fn.__name__
         print(f"Triton-Viz: execution time for {name}: {elapsed_time * 1000:.3f} ms")
     # Copy arguments back to propagate side-effects
-    self._restore_args_dev(args_dev, args_hst, kwargs, kwargs_hst)
+    if not cfg.virtual_memory:
+        self._restore_args_dev(args_dev, args_hst, kwargs, kwargs_hst)
 
 
 def _jit_function_call(self, *args, **kwargs):