[2/N][Fix] Fix typo in test folder (pytorch#166374)

Fix typo in test folder.

_typos.toml
```bash
[default.extend-words]
nd = "nd"
arange = "arange"
Nd = "Nd"
GLOBALs = "GLOBALs"
hte = "hte"
iy = "iy"
PN = "PN"
Dout = "Dout"
optin = "optin"
gam = "gam"
PTD = "PTD"
```

Pull Request resolved: pytorch#166374
Approved by: https://github.com/cyyever, https://github.com/ezyang

Author: lingebeng

test/conftest.py

@@ -238,7 +238,7 @@ def pytest_pycollect_makemodule(module_path, path, parent) -> Module:
 
 @pytest.hookimpl(hookwrapper=True)
 def pytest_report_teststatus(report, config):
-    # Add the test time to the verbose output, unforunately I don't think this
+    # Add the test time to the verbose output, unfortunately I don't think this
     # includes setup or teardown
     pluggy_result = yield
     if not isinstance(report, pytest.TestReport):

test/cpp/api/autograd.cpp

@@ -584,7 +584,7 @@ TEST(CustomAutogradTest, MarkDirty) {
     }
   };
 
-  // Clone here because modifying leafs inplace is not allowed
+  // Clone here because modifying leaves inplace is not allowed
   auto x = torch::randn({5, 5}, torch::requires_grad()).clone();
   auto version_before = x._version();
   auto out = MyFunction::apply(x);

test/cpp/api/parallel.cpp

@@ -264,7 +264,7 @@ TEST_F(ParallelTest, DataParallelNumericalEquivalence_MultiCUDA) {
     input += i;
     input_dp += i;
 
-    // non-prallel training
+    // non-parallel training
     torch::optim::SGD optim(model->parameters(), torch::optim::SGDOptions(0.1));
     auto output = model->forward(input);
     auto loss = torch::mse_loss(output, torch::zeros_like(output));

test/cpp_extensions/open_registration_extension/torch_openreg/README.md

@@ -188,7 +188,7 @@ Please refer to [this](https://docs.pytorch.org/docs/main/accelerator/index.html
   - Device-agnostic APIs
   - Memory Management
   - Generator
-  - Distrubuted
+  - Distributed
   - Custom Tensor&Storage
   - ...
 - **Improve Tests**: Add more test cases related to the integration mechanism.

test/distributed/checkpoint/test_planner.py

@@ -216,7 +216,7 @@ def create_data(rank):
         # Number of plans should remain unchanged
         self.assertEqual(len(all_plans), len(deduped_plans))
 
-        # Numer of items in the deduped plans should be less than the original plans
+        # Number of items in the deduped plans should be less than the original plans
         for new_plan, old_plan in zip(deduped_plans, all_plans):
             self.assertFalse(_compare_save_plans(new_plan, old_plan))
             self.assertTrue(len(new_plan.items) < len(old_plan.items))

test/distributed/tensor/test_attention.py

@@ -158,7 +158,7 @@ def _ring_attention_sdpa(
             # parameters because when require_grad is True, resize_ is not
             # allowed. But requires_grad of cp_q, cp_k, and cp_v are False
             # now. So we can just use context_parallel() to shard q, k, v.
-            # In reality, context_paralle() should be used to shard the input.
+            # In reality, context_parallel() should be used to shard the input.
             # In reality, context_parallel() should only be used to shard
             # the model inputs (batch).
 
@@ -701,7 +701,7 @@ def test_cp_flex_attention_document_mask(self) -> None:
         )
 
         # TODO: change this for-loop to run_subtests
-        # Use a for-loop instead of run_subtests because we need to intialize the mask
+        # Use a for-loop instead of run_subtests because we need to initialize the mask
         # for each subtest. This can be baked into self._test_cp_flex_attention as
         # a str argument denoting mask type.
         for batch_size, max_seq_len, lb_type in itertools.product(

test/distributed/test_c10d_nccl.py

@@ -4901,7 +4901,7 @@ def test_batched_send_recv(self, op_sizes_per_coalesce, timing_enabled):
             for p2p_op_idx, input_sizes in zip(
                 range(first_op, coalesced_op, 1), op_sizes_per_coalesce
             ):
-                # the indivudal ops inside the coalescing group the individual op metadata,
+                # the individual ops inside the coalescing group the individual op metadata,
                 # but not the timing info coming from the actual coalesced kernel
                 profiling_name = (
                     "nccl:recv 0<-1" if self.rank == 0 else "nccl:send 1->0"

test/distributed/test_nvshmem.py

@@ -398,7 +398,7 @@ def test_all_to_all_vdev_2d(self, align: int) -> None:
             nsplits, dtype=torch.int64, device=self.device
         ).copy_(inp_splits)
         # 2 rows: output splits, output offsets
-        # Initiallizing all values to -1 to check if they are updated
+        # Initializing all values to -1 to check if they are updated
         out_splits_offsets = symm_mem.empty(
             (2, nsplits), dtype=torch.int64, device=self.device
         ).fill_(-1)
@@ -503,7 +503,7 @@ def test_all_to_all_vdev_2d_offset(self) -> None:
             (2, nsplits), dtype=torch.int64, device=self.device
         )
         # 2 rows: output splits, output offsets
-        # Initiallizing all values to -1 to check if they are updated
+        # Initializing all values to -1 to check if they are updated
         out_splits_offsets = symm_mem.empty(
             (2, nsplits), dtype=torch.int64, device=self.device
         ).fill_(-1)
@@ -617,15 +617,15 @@ def dispatch_then_combine(device, align: int, group) -> None:
         inp_splits
     )
     # 2 rows: output splits, output offsets
-    # Initiallizing all values to -1 to check if they are updated
+    # Initializing all values to -1 to check if they are updated
     out_splits_offsets = symm_mem.empty(
         (2, nsplits), dtype=torch.int64, device=device
     ).fill_(-1)
 
     # Buffers for combine
     combine_out = symm_mem.empty(max_out_numel, dtype=dtype, device=device).fill_(-1)
     # 2 rows: output splits, output offsets
-    # Initiallizing all values to -1 to check if they are updated
+    # Initializing all values to -1 to check if they are updated
     combine_out_splits_offsets = symm_mem.empty(
         (2, nsplits), dtype=torch.int64, device=device
     ).fill_(-1)

test/distributed/test_symmetric_memory.py

@@ -274,7 +274,7 @@ def test_subgroup(self) -> None:
             self.assertTrue(buf.eq(peer_rank + world.size() // 2).all())
 
 
-# We move AsyncTP tests to a seperate test suite because 1) Async TP ops are not
+# We move AsyncTP tests to a separate test suite because 1) Async TP ops are not
 # the core symmetric memory APIs, they are more like applications, 2)
 # MultiProcContinuousTest will skip all the following tests if a test fails (
 # we should fix this too). We still want to get the test signals for the core
@@ -621,7 +621,7 @@ def test_optimal_layout(self, dim: int) -> None:
 
 # [READ ME FIRST]
 # The `SymmMemEmptySetDeviceTest` suite parameterizes whether user sets the
-# device before calling symm_mem.emtpy.  Either way should work.
+# device before calling symm_mem.empty.  Either way should work.
 # However, since `set_device` is persistent, we cannot use the
 # `MultiProcContinuousTest` template because the next function will be
 # "contaminated", leading to flaky tests (e.g. hang). Therefore, we use

test/dynamo/cpython/3_13/mathdata/ieee754.txt

@@ -51,7 +51,7 @@ nan
 >>> INF / INF
 nan
 
-However unambigous operations with inf return inf:
+However unambiguous operations with inf return inf:
 >>> INF * INF
 inf
 >>> 1.5 * INF