[memory debugging] Extract frame information from inductor (#95753)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/95753
Approved by: https://github.com/Chillee

test/inductor/test_torchinductor.py

@@ -7065,6 +7065,28 @@ if HAS_CUDA and not TEST_WITH_ASAN:
 
             self.assertTrue(torch.allclose(module(input), traced(input)))
 
+        def test_memory_history_inductor(self):
+            def called_inside_compile(x, w, b):
+                a = x @ w + b
+                return torch.sigmoid(a)
+
+            @torch.compile
+            def fn(x, w, b):
+                x = called_inside_compile(x, w, b)
+                return called_inside_compile(x, w, b)
+
+            w = torch.rand(3, 3, device="cuda")
+            b = torch.rand(3, device="cuda")
+            x = torch.rand(3, device="cuda")
+            try:
+                torch.cuda.memory.empty_cache()
+                torch.cuda.memory._record_memory_history(True)
+                r = fn(x, w, b)
+            finally:
+                torch.cuda.memory._record_memory_history(False)
+            snapshot = str(torch.cuda.memory._snapshot())
+            self.assertTrue("called_inside_compile" in snapshot)
+
     copy_tests(CommonTemplate, CudaTests, "cuda")
 
     class CudaReproTests(TestCase):

torch/_inductor/codecache.py

@@ -15,6 +15,7 @@ import sys
 import sysconfig
 import tempfile
 import types
+from bisect import bisect_right
 from concurrent.futures import Future, ProcessPoolExecutor, ThreadPoolExecutor
 from ctypes import cdll
 from functools import partial
@@ -635,10 +636,11 @@ class CppCodeCache:
 
 class PyCodeCache:
     cache = dict()
+    linemaps = dict()
     clear = staticmethod(cache.clear)
 
     @classmethod
-    def load(cls, source_code, extra=""):
+    def load(cls, source_code, extra="", linemap=()):
         key, path = write(source_code, "py", extra)
         if key not in cls.cache:
             with open(path) as f:
@@ -654,8 +656,36 @@ class PyCodeCache:
                 exec(code, mod.__dict__, mod.__dict__)
                 # another thread might set this first
                 cls.cache.setdefault(key, mod)
+                cls.linemaps[path] = linemap
+
         return cls.cache[key]
 
+    @classmethod
+    @functools.lru_cache(None)
+    def stack_frames_for_code(cls, path, lineno):
+        if path not in cls.linemaps:
+            return None
+        # [(starting_line, <fx node>), ...]
+        linemap = cls.linemaps[path]
+        p = bisect_right(linemap, lineno, key=lambda x: x[0])
+        if p == 0:
+            return None
+        _, entry = linemap[p - 1]
+        if not entry:
+            return None
+
+        def parse_stack_trace(stack_trace):
+            # ideally fx stores stack traces as data rather than a string
+            # but this is not along a performance critical path
+            regex = r'File "(.+)", line (\d+), in (.+)\n'
+            matches = re.findall(regex, stack_trace)
+            return [
+                {"filename": f, "line": int(l), "name": n}
+                for f, l, n in reversed(matches)
+            ]
+
+        return parse_stack_trace(entry.stack_trace)
+
 
 class TritonCodeCache:
     @staticmethod

torch/_inductor/codegen/wrapper.py

@@ -17,6 +17,7 @@ from ..utils import (
     cache_on_self,
     get_benchmark_name,
     has_triton,
+    LineContext,
     sympy_dot,
     sympy_product,
     sympy_symbol,
@@ -562,7 +563,7 @@ class WrapperCodeGen(CodeGen):
 
         self.add_benchmark_harness(result)
 
-        return result.getvalue()
+        return result.getvaluewithlinemap()
 
     def codegen_inputs(self, code: IndentedBuffer, graph_inputs: Dict[str, ir.Buffer]):
         """Assign all symbolic shapes to locals"""
@@ -736,6 +737,9 @@ class WrapperCodeGen(CodeGen):
     def writeline(self, line):
         self.lines.append(line)
 
+    def enter_context(self, ctx):
+        self.lines.append(LineContext(ctx))
+
 
 class CppWrapperCodeGen(WrapperCodeGen):
     """

torch/_inductor/graph.py

@@ -619,11 +619,11 @@ class GraphLowering(torch.fx.Interpreter):
     def compile_to_module(self):
         from .codecache import PyCodeCache
 
-        code = self.codegen()
+        code, linemap = self.codegen()
         if config.debug:
             print(code)
 
-        mod = PyCodeCache.load(code)
+        mod = PyCodeCache.load(code, linemap=linemap)
         for name, value in self.constants.items():
             setattr(mod, name, value)
 

torch/_inductor/scheduler.py

@@ -674,6 +674,10 @@ class Scheduler:
         self.buffer_names_to_free = set()
         self.buffer_names_no_longer_needed = set()
 
+        # fx graph node to the position it appears in the graph
+        # for debug attribution
+        self.origin_to_index = {}
+
     def debug_draw_graph(self):
         """Generate an image of the graph for debugging"""
         if os.environ.get("INDUCTOR_WRITE_SCHEDULER_GRAPH", None) == "1":
@@ -1164,9 +1168,21 @@ class Scheduler:
             self.backends[device] = self.create_backend(device)
         return self.backends[device]
 
+    def enter_context(self, node):
+        def get_order(n):
+            if n not in self.origin_to_index:
+                self.origin_to_index.update({n: i for i, n in enumerate(n.graph.nodes)})
+            return self.origin_to_index[n]
+
+        origins = [(get_order(e), e) for n in node.get_nodes() for e in n.node.origins]
+        if origins:
+            _, last = max(origins)
+            V.graph.wrapper_code.enter_context(last)
+
     @dynamo_timed
     def codegen(self):
         for node in self.nodes:
+            self.enter_context(node)
             self.buffer_names_no_longer_needed.update(node.last_usage)
 
             if not isinstance(node, NopKernelSchedulerNode):

torch/_inductor/utils.py

@@ -12,7 +12,7 @@ import tempfile
 import textwrap
 import time
 from io import StringIO
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, List, NamedTuple, Optional, Union
 from unittest import mock
 
 import sympy
@@ -420,6 +420,10 @@ def get_dtype_size(dtype):
     return torch.empty((), dtype=dtype).element_size()
 
 
+class LineContext(NamedTuple):
+    context: Any
+
+
 class IndentedBuffer:
     tabwidth = 4
 
@@ -427,19 +431,27 @@ class IndentedBuffer:
         self._lines = []
         self._indent = initial_indent
 
-    def getvalue(
-        self,
-    ):
+    def getvaluewithlinemap(self):
         buf = StringIO()
+        p = 1
+        linemap = []
         for line in self._lines:
             if isinstance(line, DeferredLineBase):
                 line = line()
                 if line is None:
                     continue
+            elif isinstance(line, LineContext):
+                linemap.append((p, line.context))
+                continue
             assert isinstance(line, str)
             buf.write(line)
             buf.write("\n")
-        return buf.getvalue()
+            p += 1 + line.count("\n")
+        return buf.getvalue(), linemap
+
+    def getvalue(self):
+        v, _ = self.getvaluewithlinemap()
+        return v
 
     def getrawvalue(self):
         buf = StringIO()
@@ -448,6 +460,8 @@ class IndentedBuffer:
                 line = line()
                 if line is None:
                     continue
+            elif isinstance(line, LineContext):
+                continue
             assert isinstance(line, str)
             # backslash implies line continuation
             if line.endswith("\\"):
@@ -467,7 +481,9 @@ class IndentedBuffer:
         return " " * (self._indent * self.tabwidth)
 
     def writeline(self, line):
-        if isinstance(line, DeferredLineBase):
+        if isinstance(line, LineContext):
+            self._lines.append(line)
+        elif isinstance(line, DeferredLineBase):
             self._lines.append(line.with_prefix(self.prefix()))
         elif line.strip():
             self._lines.append(f"{self.prefix()}{line}")
@@ -493,12 +509,15 @@ class IndentedBuffer:
         if isinstance(other_code, IndentedBuffer):
             dedent = float("inf")
             for line in other_code._lines:
-                if line:
+                if not isinstance(line, LineContext) and line:
                     dedent = min(dedent, len(line) - len(line.lstrip()))
             if math.isinf(dedent):
                 dedent = 0
             for line in other_code._lines:
-                IndentedBuffer.writeline(self, line[dedent:])
+                if isinstance(line, LineContext):
+                    self._lines.append(line)
+                else:
+                    IndentedBuffer.writeline(self, line[dedent:])
         else:
             other_code = textwrap.dedent(other_code)
             if strip:

torch/csrc/profiler/python/combined_traceback.cpp

@@ -56,6 +56,16 @@ struct PythonTraceback : public CapturedTraceback::Python {
     py::str name_s = "name";
     py::str filename_s = "filename";
 
+    auto torch = py::module::import("torch");
+    py::object stack_frames_for_code;
+    if (py::hasattr(torch, "_inductor")) {
+      py::object inductor = torch.attr("_inductor");
+      if (py::hasattr(inductor, "codecache")) {
+        stack_frames_for_code = inductor.attr("codecache")
+                                    .attr("PyCodeCache")
+                                    .attr("stack_frames_for_code");
+      }
+    }
     for (const auto& f : to_symbolize) {
       auto f_code = (PyCodeObject*)f.code;
       py::handle filename = f_code->co_filename;
@@ -67,6 +77,20 @@ struct PythonTraceback : public CapturedTraceback::Python {
           py::cast<std::string>(filename),
           py::cast<std::string>(funcname),
           (uint64_t)lineno});
+      // find all the additional frames associated with inductor generated
+      // code
+      if (stack_frames_for_code.ptr()) {
+        py::object extra = stack_frames_for_code(filename, lineno);
+        if (!extra.is_none()) {
+          for (py::handle h : extra) {
+            result.tracebacks.back().push_back(result.all_frames.size());
+            result.all_frames.emplace_back(unwind::Frame{
+                py::cast<std::string>(h[filename_s]),
+                py::cast<std::string>(h[name_s]),
+                py::cast<uint64_t>(h[line_s])});
+          }
+        }
+      }
     }
   }
 };