[Typing] Pyrefly: --remove-unused-ignores during pre-commit (#3010)

Hey there 👋 

I've noticed that pyrefly can do cleanup of unused ignores with:

```bash
pyrefly check --remove-unused-ignores
```

It can be integrated as a part of `pre-commit` step via adding this flag
as an arg.

> [!Important]
> Ensure `pass_filenames: false` is set. 
> Pyrefly is designed to check the entire project state, and running it
only on a subset of staged files can lead to inconsistent results where
it "thinks" an ignore is unused because it can't see the rest of the
project.

Author: Andrei-Aksionov

.pre-commit-config.yaml

@@ -71,16 +71,17 @@ repos:
       name: Pyrefly (type checking)
       pass_filenames: false
       language: system
+      args: ["--remove-unused-ignores"]
 
 
 - repo: local
   hooks:
     # Check hyperlinks and external links
     - id: lychee-link-checker
       name: lychee-link-checker
-      verbose: true
       language: system
       require_serial: true
+      verbose: true
       entry: |
         /bin/bash -c '
         if ! command -v lychee >/dev/null 2>&1; then
@@ -103,6 +104,7 @@ repos:
         - "--accept=200,403,429,503"
         - "--exclude=http://localhost:.*"
         - "--exclude=http://127.0.0.1:.*"
+        - "--exclude=tcp://localhost:.*"
         - "--max-retries=5"
         - "--retry-wait-time=10"
         - "--user-agent=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36"

scripts/checkpoint_conversion/convert_from_hf.py

@@ -25,7 +25,6 @@ def convert_from_hf(input_dir, output_dir, model_name, model_flavor):
         model = model_config.build()
     model = ModelWrapper(model)
 
-    # pyrefly: ignore[bad-instantiation, not-callable]
     sd_adapter = model_spec.state_dict_adapter(model_config, None)
     assert (
         sd_adapter is not None

scripts/checkpoint_conversion/convert_to_hf.py

@@ -33,7 +33,6 @@ def convert_to_hf(
         model = model_config.build()
     model = ModelWrapper(model)
 
-    # pyrefly: ignore[bad-instantiation, not-callable]
     sd_adapter = model_spec.state_dict_adapter(model_config, hf_assets_path)
     assert (
         sd_adapter is not None

scripts/checkpoint_conversion/numerical_tests_qwen3_vl.py

@@ -258,9 +258,7 @@ def run_tt(model_flavor, checkpoint_path, tt_inputs, device):
 
     # Replace FlexAttention with SDPA for single-process inference
     # (unfused FlexAttention without torch.compile has poor fp16 numerics).
-    from torchtitan.models.common.attention import (
-        ScaledDotProductAttention,  # pyrefly: ignore [missing-module-attribute]
-    )
+    from torchtitan.models.common.attention import ScaledDotProductAttention
 
     for layer in model.layers.values():
         layer.attention.attn_backend = "sdpa"

torchtitan/components/tokenizer.py

@@ -235,13 +235,10 @@ def _load_tokenizer_from_path(self, tokenizer_path: str) -> Tokenizer:
                 tokenizer = Tokenizer(bpe_model)
 
                 # Configure GPT-2 style components for proper space handling
-                # pyrefly: ignore [read-only]
                 tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(
                     add_prefix_space=False
                 )
-                # pyrefly: ignore [read-only]
                 tokenizer.decoder = decoders.ByteLevel()
-                # pyrefly: ignore [read-only]
                 tokenizer.post_processor = processors.ByteLevel(trim_offsets=True)
 
                 return tokenizer

torchtitan/components/validate.py

@@ -116,7 +116,6 @@ def __init__(
         self.tokenizer = tokenizer
         self.parallel_dims = parallel_dims
         self.loss_fn = loss_fn
-        # pyrefly: ignore [unexpected-keyword]
         self.dl_config = replace(config.dataloader, infinite=config.steps != -1)
         self.dp_world_size = dp_world_size
         self.dp_rank = dp_rank

torchtitan/config/manager.py

@@ -161,7 +161,6 @@ def _merge_configs(base, custom) -> type:
             DeprecationWarning,
             stacklevel=2,
         )
-        # pyrefly: ignore [unknown-name]
         result: list[str | tuple[str, Any] | tuple[str, Any, Any]] = []
         b_map = {f.name: f for f in fields(base)}
         c_map = {f.name: f for f in fields(custom)}
@@ -196,7 +195,6 @@ def _validate_config(self) -> None:
         # pyrefly: ignore [missing-attribute]
         if not os.path.exists(self.config.hf_assets_path):
             logger.warning(
-                # pyrefly: ignore [missing-attribute]
                 f"HF assets path {self.config.hf_assets_path} does not exist!"
             )
             old_tokenizer_path = (
@@ -254,10 +252,8 @@ def list_str_rule(type_info: tyro.constructors.PrimitiveTypeInfo):
 
     try:
 
-        # pyrefly: ignore[missing-import]
         from rich import print as rprint
 
-        # pyrefly: ignore[missing-import]
         from rich.pretty import Pretty
 
         config_manager = ConfigManager()

torchtitan/distributed/deepep/deepep.py

@@ -31,7 +31,6 @@
 
 
 # Global buffer (single buffer per process, recreated if group changes)
-# pyrefly: ignore [bad-assignment]
 _buffer: Buffer = None
 
 # Global cache for dispatch handles, keyed by handle_id
@@ -118,7 +117,6 @@ def _dispatch_op_impl(
     recv_num_tokens_per_expert = torch.tensor(
         recv_num_tokens_per_expert_list, dtype=torch.int32, device="cpu"
     )
-    # pyrefly: ignore [bad-return]
     return recv_x, recv_indices, recv_scores, recv_num_tokens_per_expert, handle_id
 
 

torchtitan/distributed/deepep/hybridep.py

@@ -181,7 +181,7 @@ def _dispatch_impl(
             _buffer.group_size,
             num_local_experts,
             topk_idx.shape[1],
-            moe_expert_capacity_factor,  # pyrefly: ignore [bad-argument-type]
+            moe_expert_capacity_factor,
             pad_multiple=pad_multiple,
         )
 

torchtitan/distributed/parallel_dims.py

@@ -146,7 +146,6 @@ def unflatten_mesh(
                 0,
                 dim_degrees,
                 dim_names,
-                # pyrefly: ignore [bad-argument-type]
                 backend_override=backend_override,
             )