Fix ty type checker errors and warnings (#606)

Suppress false-positive torch.distributed warnings, remove redundant
casts, and suppress platform-specific os.sched_getaffinity error.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: arcticfly

src/art/megatron/train.py

@@ -63,8 +63,8 @@ def freeze_model(model_chunks: list[MegatronModule]) -> list[MegatronModule]:
     data_parallel_random_init=False,
 )
 
-rank = torch.distributed.get_rank()
-world_size = torch.distributed.get_world_size()
+rank = torch.distributed.get_rank()  # ty:ignore[possibly-missing-attribute]
+world_size = torch.distributed.get_world_size()  # ty:ignore[possibly-missing-attribute]
 
 if rank == 0:
     print("TORCHINDUCTOR_CACHE_DIR:", os.environ["TORCHINDUCTOR_CACHE_DIR"])
@@ -141,7 +141,7 @@ def print0(*values: Any) -> None:
 offload_to_cpu(model, optimizer, rank, offload_state)
 
 while True:
-    torch.distributed.barrier()
+    torch.distributed.barrier()  # ty:ignore[possibly-missing-attribute]
     jobs_dir = "/tmp/megatron_training_jobs"
     os.makedirs(jobs_dir, exist_ok=True)
     job_names = sorted(
@@ -259,9 +259,9 @@ def print0(*values: Any) -> None:
             for param in chunk.parameters():
                 if param.grad is None:
                     continue
-                torch.distributed.all_reduce(
+                torch.distributed.all_reduce(  # ty:ignore[possibly-missing-attribute]
                     param.grad,
-                    op=torch.distributed.ReduceOp.AVG,
+                    op=torch.distributed.ReduceOp.AVG,  # ty:ignore[possibly-missing-attribute]
                     group=ps.get_data_parallel_group(),
                 )
                 num_grads += 1
@@ -276,7 +276,7 @@ def print0(*values: Any) -> None:
         optimizer.zero_grad()
 
         # Mean reduce loss across all ranks for logging
-        torch.distributed.all_reduce(loss, op=torch.distributed.ReduceOp.AVG)
+        torch.distributed.all_reduce(loss, op=torch.distributed.ReduceOp.AVG)  # ty:ignore[possibly-missing-attribute]
 
         if rank == 0:
             with open("/tmp/megatron_training_log.jsonl", "a+") as log_file:
@@ -322,7 +322,7 @@ def print0(*values: Any) -> None:
     gc.collect()
     torch.cuda.empty_cache()
     # Ensure all ranks have finished saving before signaling completion
-    torch.distributed.barrier()
+    torch.distributed.barrier()  # ty:ignore[possibly-missing-attribute]
     if rank == 0:
         os.remove(job_path)
         with open("/tmp/megatron_training_log.jsonl", "a+") as log_file:

src/art/preprocessing/tokenize.py

@@ -347,7 +347,7 @@ def tokenize_trajectory(
     return TokenizedResult(
         advantage=advantage,
         chat=chat,
-        tokens=[cast(str, tokenizer.decode(token_id)) for token_id in token_ids],
+        tokens=[tokenizer.decode(token_id) for token_id in token_ids],
         token_ids=token_ids,
         input_pos=list(range(len(token_ids))),
         assistant_mask=assistant_mask,

src/art/serverless/client.py

@@ -250,7 +250,7 @@ async def create(
 
     @cached_property
     def events(self) -> "TrainingJobEvents":
-        return TrainingJobEvents(cast(AsyncOpenAI, self._client))
+        return TrainingJobEvents(cast(AsyncOpenAI, self._client))  # ty:ignore[redundant-cast]
 
 
 class Client(AsyncAPIClient):

src/art/tinker/cookbook_v/renderers/qwen3.py

@@ -448,9 +448,9 @@ def _preprocess_message_parts(
             base_parts: list[ImagePart | TextPart] = []
             for p in content:
                 if p["type"] == "text":
-                    base_parts.append(cast(TextPart, p))
+                    base_parts.append(p)
                 elif p["type"] == "image":
-                    base_parts.append(cast(ImagePart, p))
+                    base_parts.append(p)
                 elif p["type"] == "thinking":
                     if not strip_thinking:
                         # Render thinking as <think>...</think> text

src/art/tinker/server.py

@@ -296,7 +296,7 @@ async def chat_completions(
 
     def _default_num_workers(self) -> int:
         try:
-            return max(1, len(os.sched_getaffinity(0)))
+            return max(1, len(os.sched_getaffinity(0)))  # ty:ignore[unresolved-attribute]
         except (AttributeError, OSError):
             return os.cpu_count() or 1