[2/N] Tiny manually fix for Ruff default ruleset and add to pre-commit (#992)

fzyzcjy · web-flow · commit f829eb2314a0 · 2025-12-01T20:37:19.000+08:00
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
@@ -17,6 +17,12 @@ repos:
         args: ['--maxkb=1000']
       - id: requirements-txt-fixer
 
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.14.7
+    hooks:
+      - id: ruff-check
+        args: [ --fix ]
+
   - repo: https://github.com/PyCQA/autoflake
     rev: v2.0.2
     hooks:
diff --git a/examples/eval/eval_delegate_rollout.py b/examples/eval/eval_delegate_rollout.py
@@ -10,7 +10,6 @@
 
 from slime.rollout.base_types import RolloutFnEvalOutput, RolloutFnTrainOutput
 from slime.rollout.sglang_rollout import generate_rollout as base_generate_rollout
-from slime.utils.metric_utils import compute_rollout_step
 
 logger = logging.getLogger(__name__)
 
@@ -83,7 +82,6 @@ def _log_delegate_metrics(args, rollout_id: int, metrics: dict | None, raw_respo
     if raw_response is not None:
         logger.info("External eval raw response for rollout %s: %s", rollout_id, raw_response)
     logger.info("eval %s (external): %s", rollout_id, flattened)
-    step = compute_rollout_step(args, rollout_id)
     return flattened
 
 
diff --git a/examples/fully_async/fully_async_rollout.py b/examples/fully_async/fully_async_rollout.py
@@ -1,4 +1,5 @@
 import asyncio
+import atexit
 import queue
 import threading
 import time
@@ -259,6 +260,5 @@ def generate_rollout_fully_async(args, rollout_id, data_buffer, evaluation=False
 
 
 # Register exit cleanup function
-import atexit
 
 atexit.register(stop_global_worker)
diff --git a/examples/search-r1/local_dense_retriever/retrieval_server.py b/examples/search-r1/local_dense_retriever/retrieval_server.py
@@ -360,7 +360,7 @@ def retrieve_endpoint(request: QueryRequest):
     scores = []
     try:
         results, scores = tmp
-    except:
+    except ValueError:
         results = tmp
 
     # Format response
diff --git a/pyproject.toml b/pyproject.toml
@@ -24,6 +24,9 @@ line_length = 119
 
 [tool.ruff]
 line-length = 119
+ignore = [
+    "E402",
+]
 
 [tool.pytest.ini_options]
 # durations=0 will display all tests execution time, sorted in ascending order starting from from the slowest one.
diff --git a/slime/backends/fsdp_utils/actor.py b/slime/backends/fsdp_utils/actor.py
@@ -217,7 +217,8 @@ def _get_init_weight_context_manager(self):
         # Check if model uses tied word embeddings (which doesn't work with meta tensors)
         use_meta_tensor = not self.hf_config.tie_word_embeddings
 
-        cpu_init_weights = lambda: torch.device("cpu")
+        def cpu_init_weights():
+            return torch.device("cpu")
 
         if use_meta_tensor:
             # Rank 0: CPU, others: meta device (memory efficient for large models)
diff --git a/slime/backends/megatron_utils/megatron_to_hf/deepseekv3.py b/slime/backends/megatron_utils/megatron_to_hf/deepseekv3.py
@@ -15,7 +15,7 @@ def convert_deepseekv3_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/megatron_to_hf/glm4.py b/slime/backends/megatron_utils/megatron_to_hf/glm4.py
@@ -12,7 +12,7 @@ def convert_glm4_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/megatron_to_hf/glm4moe.py b/slime/backends/megatron_utils/megatron_to_hf/glm4moe.py
@@ -15,7 +15,7 @@ def convert_glm4moe_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/megatron_to_hf/llama.py b/slime/backends/megatron_utils/megatron_to_hf/llama.py
@@ -12,7 +12,7 @@ def convert_llama_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/megatron_to_hf/qwen2.py b/slime/backends/megatron_utils/megatron_to_hf/qwen2.py
@@ -12,7 +12,7 @@ def convert_qwen2_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/megatron_to_hf/qwen3_next.py b/slime/backends/megatron_utils/megatron_to_hf/qwen3_next.py
@@ -15,7 +15,7 @@ def convert_qwen3_next_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/megatron_to_hf/qwen3moe.py b/slime/backends/megatron_utils/megatron_to_hf/qwen3moe.py
@@ -15,7 +15,7 @@ def convert_qwen3moe_to_hf(args, name, param):
 
     try:
         head_dim = args.kv_channels if args.kv_channels is not None else args.hidden_size // args.num_attention_heads
-    except:
+    except AttributeError:
         head_dim = args.hidden_size // args.num_attention_heads
     value_num_per_group = args.num_attention_heads // args.num_query_groups
 
diff --git a/slime/backends/megatron_utils/update_weight/hf_weight_iterator_direct.py b/slime/backends/megatron_utils/update_weight/hf_weight_iterator_direct.py
@@ -10,7 +10,7 @@
 
 try:
     from sglang.srt.utils.patch_torch import monkey_patch_torch_reductions
-except:
+except ImportError:
     from sglang.srt.patch_torch import monkey_patch_torch_reductions
 
 from tqdm import tqdm
diff --git a/slime/backends/sglang_utils/sglang_engine.py b/slime/backends/sglang_utils/sglang_engine.py
@@ -306,7 +306,7 @@ def destroy_weights_update_group(self, group_name):
                     "group_name": group_name,
                 },
             )
-        except:
+        except requests.exceptions.RequestException:
             # catch the case there the engine is just created and does not have the group.
             pass
 
diff --git a/slime/rollout/rm_hub/math_utils.py b/slime/rollout/rm_hub/math_utils.py
@@ -204,7 +204,7 @@ def _is_float(num: str) -> bool:
 def _is_int(x: float) -> bool:
     try:
         return abs(x - int(round(x))) <= 1e-7
-    except:
+    except Exception:
         return False
 
 
@@ -217,7 +217,7 @@ def _str_is_int(x: str) -> bool:
         x = _strip_properly_formatted_commas(x)
         x = float(x)
         return abs(x - int(round(x))) <= 1e-7
-    except:
+    except ValueError:
         return False
 
 
@@ -299,7 +299,7 @@ def _normalize(expr: str) -> str:
     if "\\" in expr:
         try:
             expr = _parse_latex(expr)
-        except:
+        except Exception:
             pass
 
     # edge case with mixed numbers and negative signs
@@ -353,7 +353,7 @@ def are_equal_under_sympy(ground_truth_normalized: str, given_normalized: str):
             simplified = sympy.simplify(sympy_diff)
             if simplified == 0:
                 are_equal = True
-    except:
+    except Exception:
         pass
     return are_equal
 
@@ -397,7 +397,7 @@ def last_boxed_only_string(string):
                 break
         i += 1
 
-    if right_brace_idx == None:
+    if right_brace_idx is None:
         retval = None
     else:
         retval = string[idx : right_brace_idx + 1]
@@ -411,7 +411,7 @@ def remove_boxed(s):
         assert s[: len(left)] == left
         assert s[-1] == "}"
         return s[len(left) : -1]
-    except:
+    except AssertionError:
         return None
 
 
diff --git a/slime/utils/arguments.py b/slime/utils/arguments.py
@@ -1207,7 +1207,7 @@ def add_sglang_tp_size():
                 help="Path to the YAML config for custom function arguments.",
             )
             parser.add_argument("--padded-vocab-size", type=int, default=None)
-        except:
+        except argparse.ArgumentError:
             pass
 
         return parser
@@ -1529,7 +1529,8 @@ def slime_validate_args(args):
 
 
 def hf_validate_args(args, hf_config):
-    equal = lambda x, y: x == y
+    def equal(x, y):
+        return x == y
 
     errors = []
 
diff --git a/slime/utils/data.py b/slime/utils/data.py
@@ -136,13 +136,13 @@ def __len__(self):
 def get_minimum_num_micro_batch_size(total_lengths, max_tokens_per_gpu):
     # use first fit to get the number of micro batches
     batches = []
-    for l in total_lengths:
+    for length in total_lengths:
         for i in range(len(batches)):
-            if batches[i] + l <= max_tokens_per_gpu:
-                batches[i] += l
+            if batches[i] + length <= max_tokens_per_gpu:
+                batches[i] += length
                 break
         else:
-            batches.append(l)
+            batches.append(length)
 
     return len(batches)
 
diff --git a/slime/utils/fp8_kernel.py b/slime/utils/fp8_kernel.py
@@ -67,7 +67,10 @@ def blockwise_cast_to_fp8_triton(x: torch.Tensor, block_size=None) -> Tuple[torc
     M, N = x.shape
     y = torch.empty(M, N, device=x.device, dtype=torch.float8_e4m3fn)
     s = torch.empty(ceil_div(M, BLOCK_M), ceil_div(N, BLOCK_N), dtype=torch.float32, device=x.device)
-    grid = lambda meta: (triton.cdiv(M, meta["BLOCK_M"]), triton.cdiv(N, meta["BLOCK_N"]))
+
+    def grid(meta):
+        return (triton.cdiv(M, meta["BLOCK_M"]), triton.cdiv(N, meta["BLOCK_N"]))
+
     if x.is_contiguous():
         kwargs = {"BLOCK_M": BLOCK_M, "BLOCK_N": BLOCK_N, "num_warps": 8, "num_stages": 2}
     else:
diff --git a/slime/utils/http_utils.py b/slime/utils/http_utils.py
@@ -1,5 +1,6 @@
 import asyncio
 import ipaddress
+import json
 import logging
 import multiprocessing
 import os
@@ -140,7 +141,7 @@ async def _post(client, url, payload, max_retries=60):
             response.raise_for_status()
             try:
                 output = response.json()
-            except:
+            except json.JSONDecodeError:
                 output = response.text
         except Exception as e:
             retry_count += 1
diff --git a/slime/utils/ppo_utils.py b/slime/utils/ppo_utils.py
@@ -160,7 +160,6 @@ def get_reinforce_plus_plus_returns(
     from megatron.core import mpu
 
     cp_size = mpu.get_context_parallel_world_size()
-    cp_rank = mpu.get_context_parallel_rank()
 
     final_returns_chunks = []
     for i in range(len(rewards)):
diff --git a/slime/utils/seqlen_balancing.py b/slime/utils/seqlen_balancing.py
@@ -53,9 +53,6 @@ def __init__(self, items: List[Tuple[int, int]], k: int) -> None:
                 self.sets[i].add(idx=idx, val=seqlen)
             self.sets = sorted(self.sets, reverse=True)
 
-        def spread(self):
-            return self.sets[0].sum - self.sets[-1].sum
-
         def get_partitions(self):
             partitions = []
             for i in range(len(self.sets)):
diff --git a/slime/utils/tensorboard_utils.py b/slime/utils/tensorboard_utils.py
@@ -5,7 +5,7 @@
 
 try:
     from torch.utils.tensorboard import SummaryWriter
-except:
+except ImportError:
     SummaryWriter = None
 
 logger = logging.getLogger(__name__)
diff --git a/tools/convert_hf_to_fp8.py b/tools/convert_hf_to_fp8.py
@@ -37,7 +37,6 @@ def ceildiv(a, b):
 
 
 def block_fp8(weight, block_size):
-    dtype = weight.dtype
 
     # per block quant
     block_n, block_k = block_size[0], block_size[1]
@@ -72,7 +71,6 @@ def block_fp8(weight, block_size):
 
 
 def channel_fp8(weight):
-    dtype = weight.dtype
     channel_max = torch.max(weight.abs(), dim=-1, keepdim=True)[0]
     scale = channel_max.clamp(min=1e-12).to(torch.float32) / FP8_MAX
     qweight = (weight / scale).clamp(min=FP8_MIN, max=FP8_MAX)
@@ -81,7 +79,6 @@ def channel_fp8(weight):
 
 
 def tensor_fp8(weight):
-    dtype = weight.dtype
     scale = weight.abs().max().clamp(min=1e-12).to(torch.float32) / FP8_MAX
     qweight = (weight / scale).clamp(min=FP8_MIN, max=FP8_MAX)
     qweight = qweight.to(torch.float8_e4m3fn)
diff --git a/tools/convert_hf_to_torch_dist.py b/tools/convert_hf_to_torch_dist.py
@@ -23,7 +23,7 @@ def add_convertion_args(parser):
     parser.add_argument("--hf-checkpoint", type=str, required=True, help="HuggingFace model path")
     try:
         parser.add_argument("--padded-vocab-size", type=int, default=None)
-    except:
+    except Exception:
         pass
     return parser
 
@@ -43,7 +43,8 @@ def get_args():
         "You are using too many GPUs for this conversion."
     )
 
-    ceildiv = lambda a, b: -(a // -b)  # Ceiling division
+    def ceildiv(a, b):
+        return -(a // -b)
 
     if args.pipeline_model_parallel_size == 1 and world_size > 1:
         pp_size = world_size
diff --git a/tools/convert_torch_dist_to_hf.py b/tools/convert_torch_dist_to_hf.py
@@ -5,6 +5,7 @@
 import re
 import shutil
 import time
+
 from typing import Optional
 
 import safetensors.torch
@@ -38,7 +39,7 @@ def read_metadata(self):
         with self.fs.create_stream(path, "rb") as metadata_file:
             metadata = UnpicklerWrapper(metadata_file).load()
         if getattr(metadata, "storage_meta", None) is None:
-            metadata.storage_meta = StorageMeta()
+            metadata.storage_meta = dist_cp.StorageMeta()
         metadata.storage_meta.load_id = self.load_id
         if metadata.planner_data is None:
             metadata.planner_data = {}
diff --git a/tools/fp8_cast_bf16.py b/tools/fp8_cast_bf16.py
@@ -31,7 +31,10 @@ def weight_dequant(x: torch.Tensor, s: torch.Tensor, block_size: int = 128) -> t
     assert x.dim() == 2 and s.dim() == 2
     M, N = x.size()
     y = torch.empty_like(x, dtype=torch.get_default_dtype())
-    grid = lambda meta: (triton.cdiv(M, meta["BLOCK_SIZE"]), triton.cdiv(N, meta["BLOCK_SIZE"]))
+
+    def grid(meta):
+        return (triton.cdiv(M, meta["BLOCK_SIZE"]), triton.cdiv(N, meta["BLOCK_SIZE"]))
+
     weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE=block_size)
     return y
 
