Enable ruff's unused variable checking everywhere in pytorch (#136965)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/136965
Approved by: https://github.com/cyyever, https://github.com/albanD

benchmarks/transformer/score_mod.py

@@ -221,7 +221,7 @@ def run_single_backend_sdpa(
 ) -> ExperimentResults:
     backend_context = get_backend_context(backend)
     with backend_context:
-        device = torch.device("cuda")
+        _device = torch.device("cuda")
         eager_sdpa = generate_eager_sdpa(
             config.attn_type, config.shape, config.dtype, block_mask, score_mod
         )
@@ -372,8 +372,6 @@ def run_single_experiment(
         requires_grad=config.calculate_bwd_time,
         nested_tensors=config.attn_type == "document_mask",
     )
-    is_decoding = q_seq_len == 1
-
     score_mod = generate_score_mod(config.attn_type, config.shape)
     block_mask, mask_kwargs = generate_block_mask(config.attn_type, config.shape)
     kernel_options = get_kernel_options(config.attn_type, config.shape)

pyproject.toml

@@ -64,7 +64,6 @@ ignore = [
     "E741",
     "EXE001",
     "F405",
-    "F841",
     # these ignores are from flake8-logging-format; please fix!
     "G101",
     # these ignores are from ruff NPY; please fix!

test/dynamo/test_ctx_manager.py

@@ -2696,7 +2696,7 @@ class CPythonContextManagerTestCase(torch._dynamo.test_case.TestCase):
 
         @torch.compile(backend="eager", fullgraph=True)
         def fn(t):
-            y = t.sum()
+            _y = t.sum()
             with self.assertRaises(ZeroDivisionError):
                 with woohoo() as x:
                     self.assertEqual(state, [1])
@@ -2718,7 +2718,7 @@ class CPythonContextManagerTestCase(torch._dynamo.test_case.TestCase):
         @torch.compile(backend="eager", fullgraph=True)
         def fn(t):
             nonlocal frames
-            y = t.sum()
+            _y = t.sum()
             try:
                 with f():
                     1 / 0
@@ -2745,7 +2745,7 @@ class CPythonContextManagerTestCase(torch._dynamo.test_case.TestCase):
         @torch.compile(backend="eager", fullgraph=True)
         def fn(t):
             nonlocal frames
-            y = t.sum()
+            _y = t.sum()
             try:
                 with f():
                     raise RuntimeErrorSubclass(42)
@@ -2777,7 +2777,7 @@ class CPythonContextManagerTestCase(torch._dynamo.test_case.TestCase):
                 @torch.compile(backend="eager", fullgraph=True)
                 def fn(t):
                     nonlocal frames
-                    y = t.sum()
+                    _y = t.sum()
                     try:
                         with f():
                             raise stop_exc

test/higher_order_ops/test_invoke_subgraph.py

@@ -612,7 +612,7 @@ class GraphModule(torch.nn.Module):
 
         r1 = run(ones, train=False)
         r1.sum().backward()
-        g1 = weight.grad.clone()
+        weight.grad.clone()
 
     def test_dynamic(self):
         @mark_compile_region

test/inductor/test_unbacked_symints.py

@@ -376,7 +376,6 @@ class TestUnbackedSymints(InductorTestCase):
         sub = CustomSliceSubclass(t, slice_bounds=torch.tensor([2, 5], device=t.device))
         start = 2
         length = 3
-        ragged_idx = 1
         example_inputs = (sub, start, length)
 
         actual = torch.compile(fn, dynamic=dynamic, fullgraph=True)(*example_inputs)

test/test_torch.py

@@ -952,7 +952,7 @@ class TestTorchDeviceType(TestCase):
     def test_dtypetensor_warnings(self, device):
         msg = 'The torch.cuda.*DtypeTensor constructors are no longer recommended'
         with self.assertWarnsOnceRegex(UserWarning, msg):
-            t = torch.cuda.FloatTensor([0])
+            torch.cuda.FloatTensor([0])
 
         with self.assertWarnsOnceRegex(UserWarning, msg):
             torch.cuda.DoubleTensor([0])

test/test_transformers.py

@@ -326,7 +326,7 @@ class TestTransformers(NNTestCase):
                 e = None
                 try:
                     encoder(test, src_key_padding_mask=pad_mask.to(torch.uint8))
-                except AssertionError as e:
+                except AssertionError:
                     continue
                 self.assertFalse(e, "Failed to catch unsupported uint8 type exception")  # noqa: F821
 

tools/linter/adapters/docstring_linter.py

@@ -84,7 +84,7 @@ class DocstringLinter(_linter.FileLinter):
 
         def next_token(start: int, token_type: int, error: str) -> int:  # type: ignore[return]
             for i in range(start, len(tokens)):
-                if (t := tokens[i]).type == token_type:
+                if tokens[i].type == token_type:
                     return i
             _linter.ParseError.check(False, tokens[-1], error)
 
@@ -137,12 +137,12 @@ class DocstringLinter(_linter.FileLinter):
                 print()
             top = sorted(v, reverse=True)[:REPORT_TOP_RESULTS]
             if len(top) == 1:
-                s = "s"
-                t = ""
-            else:
                 s = ""
                 t = f"{len(top)} "
-            print(f"Top {t}undocumented {k}s:")
+            else:
+                s = "es" if k.endswith("s") else "s"
+                t = ""
+            print(f"Top {t}undocumented {k}{s}:")
             for lines, path, tname in top:
                 print(f"    {lines} lines: {path}:{tname}")
 

tools/test/docstring_linter_testdata/python_code.py.txt.lintrunner

@@ -46,11 +46,11 @@ tools/test/docstring_linter_testdata/python_code.py.txt:84: No docstring found f
    85 |     def short1(self):
    86 |         pass
 
-Top 3 undocumented classs:
+Top undocumented classes:
     12 lines: tools/test/docstring_linter_testdata/python_code.py.txt:NotDocstring
     6 lines: tools/test/docstring_linter_testdata/python_code.py.txt:LongWithShortDocstring
     6 lines: tools/test/docstring_linter_testdata/python_code.py.txt:Long
 
-Top 2 undocumented functions:
+Top undocumented functions:
     12 lines: tools/test/docstring_linter_testdata/python_code.py.txt:needs_docs
     11 lines: tools/test/docstring_linter_testdata/python_code.py.txt:not_short

torch/_dynamo/symbolic_convert.py

@@ -3229,7 +3229,6 @@ class InliningInstructionTranslator(InstructionTranslatorBase):
     def inline_call_(self):
         parent = self.parent
         code = self.f_code
-        func = self.funcvar
 
         strict_ctx: Any = contextlib.nullcontext()
         if parent.strict_checks_fn:

torch/_inductor/codegen/cuda/cuda_template.py

@@ -90,9 +90,7 @@ class CUDATemplate(KernelTemplate):
             call_args,
             expected_args,
         )
-        extra_args = V.graph.sizevars.size_hints(
-            map(sympy.expand, call_args[len(expected_args) :])
-        )
+        V.graph.sizevars.size_hints(map(sympy.expand, call_args[len(expected_args) :]))
         size_args = V.graph.sizevars.size_hints(kernel.get_layout_args())
 
         kernel_hash_name = f"cuda_{self.name}_{next(self.index_counter)}"

torch/_inductor/codegen/wrapper.py

@@ -1270,7 +1270,8 @@ class PythonWrapperCodegen(CodeGen):
 
         # conservatively use the sum of all allocated buffer sizes
         # in potentially nested scopes as the total allocated size
-        total_allocated_buffer_size = sum(
+        # FIXME(rec): not used
+        _total_allocated_buffer_size = sum(
             s.total_allocated_buffer_size for s in past_planning_states
         )
 

torch/_inductor/compile_worker/subproc_pool.py

@@ -189,7 +189,7 @@ class SubprocPool:
             data = b""
             try:
                 job_id, data = _recv_msg(self.read_pipe)
-            except Exception as e:
+            except Exception:
                 # Something went wrong during the read. There's no way we have a
                 # valid job_id.
                 log.exception("failure in subproc_pool._recv_msg")

torch/_inductor/kernel/mm_scaled.py

@@ -438,7 +438,6 @@ def scaled_mm_options_device_tma(  # type: ignore[no-untyped-def]
         "Expect scale_a and scale_b to be either both scalars (including single-element tensors) "
         f"or 1-dimensional tensors with the same size. Got scale_a: {len(size_a)} and scale_b: {len(size_b)}."
     )
-    NUM_SMS = torch.cuda.get_device_properties("cuda").multi_processor_count
     return dict(
         GROUP_M=8,
         EVEN_K=even_k_symbolic,

torch/_inductor/select_algorithm.py

@@ -695,8 +695,6 @@ class TritonTemplateKernel(TritonKernel):
             lengths = [V.graph.sizevars.simplify(s) for s in input_node.get_size()]
             assert len(indices) == len(lengths)
 
-            stride = self.named_input_nodes[input_name].get_stride()
-
             index_symbols = [sympy.Symbol(x, integer=True) for x in indices]
             assert len(indices) == len(lengths)
 
@@ -722,7 +720,6 @@ class TritonTemplateKernel(TritonKernel):
                 sympy.Integer(1), sympy_product(lengths)
             )
             xindex_range_root.set_name("xindex")
-            xindex_expr = xindex_range_root.expr
 
             # Note - ["None" override_mask]
             # MM Templates work by taking out of bounds index values and wrapping them around to 0

torch/ao/quantization/fx/convert.py

@@ -1123,10 +1123,6 @@ def convert(
     # for dynamic quant ops or weight only quant ops
     _run_weight_observers(model, backend_config)
 
-    graph_inputs: List[str] = [
-        node.name for node in model.graph.nodes if node.op == "placeholder"
-    ]
-
     # additional state to override inputs to be quantized, if specified
     # by the user
     placeholder_node_seen_cnt = 0

torch/distributed/elastic/timer/file_based_local_timer.py

@@ -41,7 +41,7 @@ def _retry(max_retries: int, sleep_time: float) -> Callable:
             for i in range(max_retries):
                 try:
                     return func(*args, **kwargs)
-                except Exception as e:
+                except Exception:
                     logger.exception("Error running %s. Retrying...", func.__name__)
                     if i < max_retries - 1:
                         time.sleep(sleep_time)
@@ -277,7 +277,7 @@ class FileTimerServer:
         #     thread, which will not block the process to stop.
         try:
             fd = open(self._file_path)
-        except Exception as e:
+        except Exception:
             logger.exception("Could not open the FileTimerServer pipe")
             raise
 

torch/distributed/tensor/experimental/_attention.py

@@ -319,7 +319,6 @@ class _AllGatherRotater(_RingRotater):
             )
 
     def next_buffer(self) -> torch.Tensor:
-        size = dist.get_world_size(self._pg)
         rank = dist.get_rank(self._pg)
         idx = rank - self._idx
 

torch/export/_trace.py

@@ -1548,10 +1548,6 @@ def _export_to_aten_ir_make_fx(
         strict=True,
         stack_weights=True,
     ), _ignore_backend_decomps(), _compiling_state_context():  # type: ignore[attr-defined]
-        param_len = len(dict(mod.named_parameters(remove_duplicate=False)))
-        buffer_len = len(dict(mod.named_buffers(remove_duplicate=False)))
-        params_len = param_len + buffer_len
-
         gm, graph_signature = transform(_make_fx_helper)(
             mod,
             fake_args,

torch/fx/experimental/symbolic_shapes.py

@@ -2900,7 +2900,7 @@ class DimConstraints:
                 assert op == "==", t
                 try:
                     results[left]["eq"] = sympy.sympify(right)
-                except TypeError as e:  # rhs source is not linked to Dim name
+                except TypeError:  # rhs source is not linked to Dim name
                     pass
 
         # order forced specializations based on name

torch/nested/_internal/ops.py

@@ -917,8 +917,6 @@ def chunk_default(func, *args, **kwargs):
 
     if operating_on_batch:
         chunks = new_kwargs["chunks"]
-        dim0_size = inp._size[0]
-        chunk_size = math.ceil(dim0_size / chunks)
 
         # get _offsets of the chunks
         lengths = inp._offsets.diff()

torch/testing/_internal/common_methods_invocations.py

@@ -8752,7 +8752,6 @@ def sample_inputs_scaled_mm(op_info, device, dtype, requires_grad, **kwargs):
 def sample_inputs_scaled_dot_product_attention(op_info, device, dtype, requires_grad, **kwargs):
     make = partial(make_tensor, device=device, dtype=dtype, requires_grad=requires_grad)
     batch, seq_q, seq_kv, num_heads, head_dim = 4, 3, 6, 4, 8
-    num_heads_q_gqa, num_heads_kv_gqa = 32, 8
 
     dim_3_q_shape = (batch, seq_q, head_dim)
     dim_3_kv_shape = (batch, seq_kv, head_dim)

torch/testing/_internal/distributed/rpc/rpc_test.py

@@ -1452,7 +1452,7 @@ class RpcTest(RpcAgentTestFixture, RpcTestCommon):
         model = torch.nn.parallel.DistributedDataParallel(model)
 
         with self.assertRaisesRegex(RuntimeError, 'Current RPC agent is not set! Did you initialize the RPC framework'):
-            params = [RRef(param) for param in model.parameters()]
+            [RRef(param) for param in model.parameters()]
 
     def test_world_size_one(self):
         self._world_size_one(

torch/testing/_internal/opinfo/definitions/nested.py

@@ -447,7 +447,6 @@ def reduction_reference(op, sample):
     assert op._extra_op_data.dim_args is not None
     single_dim_argname, dimlist_argname = op._extra_op_data.get_dim_argnames()
     assert single_dim_argname is not None
-    supports_dimlist = dimlist_argname is not None
 
     dim = sample.kwargs.get(
         dimlist_argname, sample.kwargs.get(single_dim_argname, None)
@@ -814,7 +813,6 @@ def sample_inputs_unary_dimwise(
 
 def batchwise_reference_chunk(op, sample):
     # reference for chunk() over dim=0
-    kwargs = sample.kwargs
     B = sample.input.size(0)
     num_chunks = sample.kwargs["chunks"]
     chunk_size = math.ceil(B / num_chunks)