Revert "Enable possibly-undefined error code (pytorch#118533)"

This reverts commit 4f13f69.

Reverted pytorch#118533 on behalf of https://github.com/clee2000 due to sorry i'm trying to figure out a codev merge conflict, if this works i'll be back to rebase and merge ([comment](pytorch#118533 (comment)))

mypy.ini

@@ -13,7 +13,6 @@ show_column_numbers = True
 check_untyped_defs = True
 follow_imports = normal
 local_partial_types = True
-enable_error_code = possibly-undefined
 
 # do not reenable this:
 # https://github.com/pytorch/pytorch/pull/60006#issuecomment-866130657

test/typing/pass/creation_ops.py

@@ -1,4 +1,3 @@
-# mypy: disable-error-code="possibly-undefined"
 # flake8: noqa
 import torch
 from torch.testing._internal.common_utils import TEST_NUMPY

test/typing/reveal/tensor_constructors.py

@@ -1,4 +1,3 @@
-# mypy: disable-error-code="possibly-undefined"
 # flake8: noqa
 import torch
 from torch.testing._internal.common_utils import TEST_NUMPY

torch/__init__.py

@@ -515,7 +515,7 @@ def fn(a):
 sym_sqrt = current_module._sym_sqrt
 __all__.append("sym_sqrt")
 
-del fn, name, sym_name, current_module  # type: ignore[possibly-undefined]
+del fn, name, sym_name, current_module
 
 
 def sym_ite(b, t, f):

torch/_decomp/decompositions.py

@@ -2832,7 +2832,7 @@ def _rnn_helper(
             final_hiddens.append(bwd_hidden)
 
         if bidirectional:
-            input = torch.cat([fwd_inp, bwd_inp], fwd_inp.dim() - 1)  # type: ignore[possibly-undefined]
+            input = torch.cat([fwd_inp, bwd_inp], fwd_inp.dim() - 1)
         else:
             input = fwd_inp
 

torch/_dynamo/convert_frame.py

@@ -163,7 +163,7 @@ def _fn(*args, **kwargs):
             random.setstate(py_rng_state)
             torch.random.set_rng_state(torch_rng_state)
             if torch.cuda.is_available():
-                torch.cuda.set_rng_state(cuda_rng_state)  # type: ignore[possibly-undefined]
+                torch.cuda.set_rng_state(cuda_rng_state)
             torch.fx.graph_module._forward_from_src = prior_fwd_from_src
             assert (
                 guards.check()
@@ -568,7 +568,7 @@ def log_bytecode(prefix, name, filename, line_no, code):
             code.co_name,
             code.co_filename,
             code.co_firstlineno,
-            out_code,  # type: ignore[possibly-undefined]
+            out_code,
         )
 
         for hook in _bytecode_hooks.values():

torch/_dynamo/debug_utils.py

@@ -46,7 +46,7 @@
         "//deeplearning/fbgemm/fbgemm_gpu:sparse_ops_cpu",
         "//deeplearning/fbgemm/fbgemm_gpu:sparse_ops",
     ]
-    cur_target = libfb.py.build_info.BuildInfo.get_build_rule().replace("fbcode:", "//")  # type: ignore[possibly-undefined]
+    cur_target = libfb.py.build_info.BuildInfo.get_build_rule().replace("fbcode:", "//")
     extra_imports = "\n".join([f'torch.ops.load_library("{x}")' for x in extra_deps])
 
 

torch/_dynamo/eval_frame.py

@@ -1430,7 +1430,7 @@ def graph_with_interpreter(*args):
                 example_fake_inputs,
                 graph_captured_input,
                 graph_captured_result,
-                result_traced,  # type: ignore[possibly-undefined]
+                result_traced,
                 flat_args_dynamic_dims,
             )
         # Store constraints and inputs as metadata for user passes, e.g. turn constraints to runtime check

torch/_dynamo/symbolic_convert.py

@@ -1115,7 +1115,7 @@ def POP_FINALLY(self, inst):
             tos = self.pop()
         _ = self.pop()
         if preserve_tos:
-            self.push(tos)  # type: ignore[possibly-undefined]
+            self.push(tos)
 
     def FOR_ITER(self, inst):
         it = self.pop().realize()

torch/_dynamo/test_minifier_common.py

@@ -118,7 +118,7 @@ def _maybe_subprocess_run(self, args, *, isolate, cwd=None):
                 finally:
                     log.removeHandler(log_handler)
                     if cwd is not None:
-                        os.chdir(prev_cwd)  # type: ignore[possibly-undefined]
+                        os.chdir(prev_cwd)
                     # Make sure we don't leave buggy compiled frames lying
                     # around
                     torch._dynamo.reset()

torch/_dynamo/utils.py

@@ -773,7 +773,7 @@ def preserve_rng_state():
         with torch.utils._python_dispatch._disable_current_modes():
             torch.random.set_rng_state(rng_state)
             if torch.cuda.is_available():
-                torch.cuda.set_rng_state(cuda_rng_state)  # type: ignore[possibly-undefined]
+                torch.cuda.set_rng_state(cuda_rng_state)
 
 
 def is_jit_model(model0):
@@ -892,7 +892,7 @@ def timed(model, example_inputs, times=1):
         result = model(*example_inputs)
         synchronize()
     t1 = time.perf_counter()
-    return result, t1 - t0  # type: ignore[possibly-undefined]
+    return result, t1 - t0
 
 
 def check_is_cuda(gm, example_inputs):

torch/_functorch/autograd_function.py

@@ -199,7 +199,7 @@ def wrap_outputs_maintaining_identity(
             result.append(unwrapped_input_to_orig_input[id(output)])
             continue
         if out_dims_specified:
-            result.append(wrap_fn(output, flat_out_dims[i]))  # type: ignore[possibly-undefined, index]
+            result.append(wrap_fn(output, flat_out_dims[i]))  # type: ignore[index]
         else:
             result.append(wrap_fn(output))
 

torch/_higher_order_ops/triton_kernel_wrap.py

@@ -163,7 +163,7 @@ def parse_ttir(ttir, kwargs):
         return None
 
     try:
-        import lark  # type: ignore[import-not-found]
+        import lark
         from lark import Lark, Transformer, v_args
     except ModuleNotFoundError:
         warnings.warn(

torch/_inductor/autotune_process.py

@@ -440,25 +440,25 @@ def benchmark(
             output_tensor = self.output_tensor_meta.to_tensor()
 
         if debug:
-            create_tensor_elapse = time.time() - start_ts  # type: ignore[possibly-undefined]
+            create_tensor_elapse = time.time() - start_ts
             start_ts = time.time()
 
         fn = self.make_run_fn(*input_tensors, output_tensor=output_tensor)
 
         if debug:
-            load_elapse = time.time() - start_ts  # type: ignore[possibly-undefined]
+            load_elapse = time.time() - start_ts
             start_ts = time.time()
 
         out = do_bench(fn)
         torch.cuda.synchronize()  # shake out any CUDA errors
 
         if debug:
-            bench_elapse = time.time() - start_ts  # type: ignore[possibly-undefined]
+            bench_elapse = time.time() - start_ts
             log.debug(
                 "InChildProcess %s: load %f, create tensor %f, bench %f",
                 str(self),
-                load_elapse,  # type: ignore[possibly-undefined]
-                create_tensor_elapse,  # type: ignore[possibly-undefined]
+                load_elapse,
+                create_tensor_elapse,
                 bench_elapse,
             )
         self.cleanup_run_fn()

torch/_inductor/codegen/cuda/cutlass_epilogue_gen.py

@@ -99,7 +99,7 @@ def ir_to_evt_string(
                 result = pnode.inner_fn(index)
                 # each epilogue node results in a single "using" statement and may refer to the previous steps by name
                 formatter.aliases[node.name] = result
-            res = formatter.getvalue(result)  # type: ignore[possibly-undefined]
+            res = formatter.getvalue(result)
             if _MAGIC_SYMPY_ERROR_STRING in res:
                 raise CUTLASSEVTOpNotImplementedError(
                     "sympy / indexing expressions not yet supported in EVT fusion"
@@ -266,7 +266,7 @@ def ir_to_evt_argument_string(
                 if node.name is not None:
                     formatter.aliases[node.name] = result
 
-            res: str = formatter.getvalue(result)  # type: ignore[possibly-undefined]
+            res: str = formatter.getvalue(result)
             if _MAGIC_SYMPY_ERROR_STRING in res:
                 raise CUTLASSEVTOpNotImplementedError(
                     "sympy / indexing expressions not yet supported in EVT fusion"

torch/_inductor/codegen/wrapper.py

@@ -155,7 +155,7 @@ def get_cpp_op_schema(kernel: torch._ops.OpOverload) -> str:
         cpp_return_value = f"std::tuple<{tuple_returns}>"
 
     cpp_arg_type = [f"{convert_arg_type(arg)} {arg.name}" for arg in args]
-    return f"{cpp_return_value}({', '.join(cpp_arg_type)})"  # type: ignore[possibly-undefined]
+    return f"{cpp_return_value}({', '.join(cpp_arg_type)})"
 
 
 # TODO: Move to a well known place

torch/_inductor/comm_analysis.py

@@ -209,7 +209,7 @@ def estimate_nccl_collective_runtime(snode: "BaseSchedulerNode") -> float:
         nsteps = nRanks - 1
 
     # Convert bus BW to algorithm BW (tensor bytes / algoBW = actual execution time)
-    ratio = (1.0 * nRanks) / nsteps  # type: ignore[possibly-undefined]
+    ratio = (1.0 * nRanks) / nsteps
     bandwidth = busBw * ratio
     # Convert GB/s to GB/ns
     bandwidth_GB_per_ns = bandwidth / 1e9
@@ -236,7 +236,7 @@ def estimate_nccl_collective_runtime(snode: "BaseSchedulerNode") -> float:
     if nNodes > 1:
         netOverhead = 1.0  # getNetOverhead(comm);
     intraLat = max(intraLat, netOverhead)
-    latency += (nsteps - nInterSteps) * intraLat + nInterSteps * interLat  # type: ignore[possibly-undefined]
+    latency += (nsteps - nInterSteps) * intraLat + nInterSteps * interLat
     # Convert us to ns
     latency_ns = latency * 1e3
 

torch/_inductor/fx_passes/group_batch_fusion.py

@@ -170,9 +170,9 @@ def fuse(self, graph: torch.fx.GraphModule, subset: List[torch.fx.Node]):
                 input, weight = node.args
                 bias = None
             batch_nodes.append(node)
-            batch_inputs.append(input)  # type: ignore[possibly-undefined]
-            batch_weights.append(weight)  # type: ignore[possibly-undefined]
-            batch_biases.append(bias)  # type: ignore[possibly-undefined]
+            batch_inputs.append(input)
+            batch_weights.append(weight)
+            batch_biases.append(bias)
 
         with graph.inserting_before(subset[-1]):
             fused_inputs = decompose_stack(graph, batch_inputs)
@@ -191,7 +191,7 @@ def fuse(self, graph: torch.fx.GraphModule, subset: List[torch.fx.Node]):
                     new_bias_add = graph.call_function(
                         aten.add, args=((batch_biases[i], new_mm))
                     )
-            new_mm_cont = new_bias_add if has_bias else new_mm  # type: ignore[possibly-undefined]
+            new_mm_cont = new_bias_add if has_bias else new_mm
             original_mm.replace_all_uses_with(new_mm_cont)
             new_mm_cont.meta.update(original_mm.meta)
             graph.erase_node(original_mm)

torch/_inductor/fx_passes/mkldnn_fusion.py

@@ -283,7 +283,7 @@ def fn(match, *args, **kwargs):
                     L[aten.mul](out, negative_slope),
                 )
                 if lowp_dtype:
-                    out = L[prims.convert_element_type.default](out, dtype=dtype2)  # type: ignore[possibly-undefined]
+                    out = L[prims.convert_element_type.default](out, dtype=dtype2)
                 return out
 
         return fn
@@ -324,7 +324,7 @@ def fn(match, *args, **kwargs):
                     out = L[prims.convert_element_type.default](out, dtype=torch.float)
                 out = L[aten.clamp_max](L[aten.clamp_min](out, min_value), max_value)
                 if lowp_dtype:
-                    out = L[prims.convert_element_type.default](out, dtype=dtype2)  # type: ignore[possibly-undefined]
+                    out = L[prims.convert_element_type.default](out, dtype=dtype2)
                 return out
 
         return fn

torch/_inductor/fx_passes/pre_grad.py

@@ -105,7 +105,7 @@ def pre_grad_passes(gm: torch.fx.GraphModule, example_inputs):
 
         gm_after_fx_passes = gm.__copy__()
         numeric_check_if_enabled(
-            gm_before_fx_passes,  # type: ignore[possibly-undefined]
+            gm_before_fx_passes,
             gm_after_fx_passes,
             example_inputs,
             config.fx_passes_numeric_check.get("num_iterations", 1),

torch/_inductor/fx_passes/quantization.py

@@ -1360,7 +1360,7 @@ def qconv_weight_prepack(match: Match, *args, **kwargs):
             graph.erase_node(conv_node)
             # Erase the dequant pattern
             if dtype == torch.bfloat16:
-                graph.erase_node(convert_to_bf16)  # type: ignore[possibly-undefined]
+                graph.erase_node(convert_to_bf16)
             # Erase the dequant pattern
             graph.erase_node(mul_node)
             graph.erase_node(sub_node)
@@ -1369,7 +1369,7 @@ def qconv_weight_prepack(match: Match, *args, **kwargs):
             if clone_node is not None:
                 graph.erase_node(clone_node)
             if dtype == torch.bfloat16:
-                graph.erase_node(weight_to_bf16_node)  # type: ignore[possibly-undefined]
+                graph.erase_node(weight_to_bf16_node)
             graph.erase_node(dequant_per_channel)
             counters["inductor"]["qconv2d_weight_prepack_matcher_count"] += 1
             counters["inductor"]["qconv2d_weight_prepack_matcher_nodes"] += len(
@@ -1697,14 +1697,14 @@ def qlinear_weight_prepack(match: Match, *args, **kwargs):
                 if input_contiguous:
                     graph.erase_node(output_reshape_node)
                 elif not input_contiguous and bias:
-                    graph.erase_node(output_add_node_for_bias)  # type: ignore[possibly-undefined]
+                    graph.erase_node(output_add_node_for_bias)
             graph.erase_node(linear_node)
             if input_dim_exceeds_two:
                 if input_contiguous:
                     graph.erase_node(act_reshape_node)
                 else:
                     graph.erase_node(act_expand_node)
-                    graph.erase_node(wgt_expand_node)  # type: ignore[possibly-undefined]
+                    graph.erase_node(wgt_expand_node)
             if dtype == torch.bfloat16:
                 graph.erase_node(activation_to_bf16_node)
             # Erase the dequant pattern
@@ -1714,7 +1714,7 @@ def qlinear_weight_prepack(match: Match, *args, **kwargs):
             # Erase the dequant per channel pattern
             graph.erase_node(t_node)
             if dtype == torch.bfloat16:
-                graph.erase_node(weight_to_bf16_node)  # type: ignore[possibly-undefined]
+                graph.erase_node(weight_to_bf16_node)
             graph.erase_node(dequant_per_channel)
 
             counters["inductor"]["qlinear_weight_prepack_matcher_count"] += 1

torch/_inductor/graph.py

@@ -845,7 +845,7 @@ def debug(msg):
             ):
                 debug("fallback_handler")
                 result = fallback_handler(n.target, add_to_fallback_set=False)(
-                    *args, **kwargs  # type: ignore[possibly-undefined]
+                    *args, **kwargs
                 )
             elif n.op == "call_function" and n.target in layout_constraints:
                 debug("layout_constraints")

torch/_inductor/lowering.py

@@ -607,7 +607,7 @@ def register_pointwise(
         fn,
         override_return_dtype=override_return_dtype,
         override_fn_when_input_bool=override_fn_when_input_bool,
-        override_fn_when_cuda_float64=fn_libdevice if use_libdevice_for_f64 else None,  # type: ignore[possibly-undefined]
+        override_fn_when_cuda_float64=fn_libdevice if use_libdevice_for_f64 else None,
         allow_alpha=allow_alpha,
     )
     fn = register_lowering(
@@ -3630,8 +3630,8 @@ def accumulate(grad, out, index_ranges):
                     out = right_reflect[i]
                     index_range = (xyz[i], dhw[i] - padding_right[i], dhw[i] - 1)
 
-                outs.append(out)  # type: ignore[possibly-undefined]
-                index_ranges.append(index_range)  # type: ignore[possibly-undefined]
+                outs.append(out)
+                index_ranges.append(index_range)
 
             grad = accumulate(grad, outs, index_ranges)
 

torch/_inductor/pattern_matcher.py

@@ -1196,7 +1196,7 @@ def apply(self, graph: torch.fx.GraphModule) -> int:
                     if (
                         self.prevent_match_across_mutations
                         and is_match(m)
-                        and len(set(map(get_mutation_region_id_partial, m.nodes))) != 1  # type: ignore[possibly-undefined]
+                        and len(set(map(get_mutation_region_id_partial, m.nodes))) != 1
                     ):
                         continue
                     if os.environ.get("TORCHINDUCTOR_PATTERN_MATCH_DEBUG") == node.name:

torch/_inductor/scheduler.py

@@ -1038,7 +1038,7 @@ def fuse(cls, producer, consumer):
                 else:
                     fused_nodes.append(node)
 
-        return cls(producer.scheduler, fused_nodes, prev_node_1, prev_node_2)  # type: ignore[possibly-undefined]
+        return cls(producer.scheduler, fused_nodes, prev_node_1, prev_node_2)
 
     def __init__(
         self,
@@ -2256,13 +2256,13 @@ def codegen(self):
 
             if node.is_template():
                 node, *epilogue = node.get_nodes()
-                self.get_backend(device).codegen_template(node, epilogue)  # type: ignore[possibly-undefined]
+                self.get_backend(device).codegen_template(node, epilogue)
             elif node.is_extern():
                 self.codegen_extern_call(node)
             elif node.is_foreach():
-                self.get_backend(device).codegen_foreach(node)  # type: ignore[possibly-undefined]
+                self.get_backend(device).codegen_foreach(node)
             elif isinstance(node, (FusedSchedulerNode, SchedulerNode)):
-                self.get_backend(device).codegen_nodes(node.get_nodes())  # type: ignore[possibly-undefined]
+                self.get_backend(device).codegen_nodes(node.get_nodes())
             else:
                 assert isinstance(node, NopKernelSchedulerNode)
                 node.allocate()
@@ -2271,7 +2271,7 @@ def codegen(self):
                 V.graph.wrapper_code.generate_inf_and_nan_checker(node)
 
             if config.triton.debug_sync_kernel:
-                self.get_backend(device).codegen_sync()  # type: ignore[possibly-undefined]
+                self.get_backend(device).codegen_sync()
 
             self.available_buffer_names.update(node.get_names())
 

torch/_inductor/utils.py

@@ -331,7 +331,7 @@ def timed(
         synchronize(device)
     t1 = time.perf_counter()
     # GC the result after timing
-    assert result is not None  # type: ignore[possibly-undefined]
+    assert result is not None
     return t1 - t0