Add typo checker in pre-commit (#6179)

Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>

python/sglang/compile_deep_gemm.py

@@ -129,7 +129,7 @@ def launch_server_process_and_send_one_request(
 
 
 def refine_server_args(server_args: ServerArgs, compile_args: CompileArgs):
-    # Disbale cuda graph and torch compile to save time
+    # Disable cuda graph and torch compile to save time
     server_args.disable_cuda_graph = True
     server_args.enable_torch_compile = False
     print(f"Disable CUDA Graph and Torch Compile to save time...")

python/sglang/lang/tracer.py

@@ -38,7 +38,7 @@ def extract_prefix_by_tracing(program, backend):
         with TracingScope(tracer):
             tracer.ret_value = program.func(tracer, **arguments)
     except (StopTracing, TypeError, AttributeError):
-        # Some exceptions may not be catched
+        # Some exceptions may not be caught
         pass
 
     # Run and cache prefix

python/sglang/srt/code_completion_parser.py

@@ -27,7 +27,7 @@ completion_template_name = None
 
 
 class FimPosition:
-    """Postion of fim middle token."""
+    """Position of fim middle token."""
 
     MIDDLE = auto()
     END = auto()

python/sglang/srt/configs/deepseekvl2.py

@@ -416,9 +416,9 @@ class DeepseekVLV2Processor(ProcessorMixin):
             h = w = math.ceil(
                 (self.image_size // self.patch_size) / self.downsample_ratio
             )
-            # global views tokens h * (w + 1), 1 is for line seperator
+            # global views tokens h * (w + 1), 1 is for line separator
             tokenized_image = [self.image_token_id] * h * (w + 1)
-            # add a seperator between global and local views
+            # add a separator between global and local views
             tokenized_image += [self.image_token_id]
             # local views tokens, (num_height_tiles * h) * (num_width_tiles * w + 1)
             tokenized_image += (

python/sglang/srt/disaggregation/decode.py

@@ -509,7 +509,7 @@ class SchedulerDisaggregationDecodeMixin:
     def event_loop_overlap_disagg_decode(self: Scheduler):
         result_queue = deque()
         self.last_batch: Optional[ScheduleBatch] = None
-        self.last_batch_in_queue = False  # last batch is modifed in-place, so we need another variable to track if it's extend
+        self.last_batch_in_queue = False  # last batch is modified in-place, so we need another variable to track if it's extend
 
         while True:
             recv_reqs = self.recv_requests()

python/sglang/srt/disaggregation/fake/conn.py

@@ -54,7 +54,7 @@ class FakeKVSender(BaseKVSender):
             logger.info(f"FakeKVSender send success")
         else:
             self.has_sent = False
-            logger.info(f"FakeKVSender send fake transfering")
+            logger.info(f"FakeKVSender send fake transferring")
 
     def failure_exception(self):
         raise Exception("Fake KVSender Exception")

python/sglang/srt/disaggregation/mooncake/conn.py

@@ -363,7 +363,7 @@ class MooncakeKVManager(BaseKVManager):
         self.request_status[bootstrap_room] = KVPoll.WaitingForInput
 
     def check_status(self, bootstrap_room: int):
-        # TOOD: do we really need the poll()?
+        # TODO: do we really need the poll()?
 
         return self.request_status[bootstrap_room]
 

python/sglang/srt/disaggregation/utils.py

@@ -112,7 +112,7 @@ def get_kv_class(transfer_backend: TransferBackend, class_type: KVClassType):
 
 
 def kv_to_page_indices(kv_indices: np.ndarray, page_size: int):
-    # 1. The page is guaruanteed to be full except the last page.
+    # 1. The page is guaranteed to be full except the last page.
     # 2. page index = kv_index // page_size
     # The return vector is kv_indices[::page_size] // page_size
     if page_size == 1:  # shortcut

python/sglang/srt/function_call_parser.py

@@ -86,8 +86,8 @@ class StructureInfo:
 
 _GetInfoFunc = Callable[[str], StructureInfo]
 """
-helper alias of function
-ususally it is a function that takes a name string and returns a StructureInfo object,
+Helper alias of function
+Usually it is a function that takes a name string and returns a StructureInfo object,
 which can be used to construct a structural_tag object
 """
 

python/sglang/srt/layers/attention/flashattention_backend.py

@@ -308,7 +308,7 @@ class FlashAttentionBackend(AttentionBackend):
         ), "Sliding window and cross attention are not supported together"
 
         self.forward_metadata: FlashAttentionMetadata = None
-        # extra metdata for handling speculative decoding topk > 1, extended draft decode and verify
+        # extra metadata for handling speculative decoding topk > 1, extended draft decode and verify
         self.forward_metadata_spec_decode_expand: FlashAttentionMetadata = None
         self.max_context_len = model_runner.model_config.context_len
         self.device = model_runner.device

python/sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py

@@ -919,7 +919,7 @@ def _fwd_kernel(
 
         e_max = n_e_max
 
-    # stage 2: compute the trianlge part
+    # stage 2: compute the triangle part
 
     cur_block_m_end = tl.minimum(cur_seq_len_extend, (cur_block_m + 1) * BLOCK_M)
     for start_n in range(0, cur_block_m_end, BLOCK_N):

python/sglang/srt/layers/dp_attention.py

@@ -201,7 +201,7 @@ def _dp_gather(
             global_tokens, local_tokens, 0, local_start_pos, local_num_tokens, False
         )
 
-    # Input IDs are in int 32. We should use inplace_all_reduce for local case becaues of custom all reduce.
+    # Input IDs are in int 32. We should use inplace_all_reduce for local case because of custom all reduce.
     NUM_GPUS_PER_NODE = 8
     if (
         not local_tokens.dtype.is_floating_point

python/sglang/srt/layers/layernorm.py

@@ -76,7 +76,7 @@ class RMSNorm(CustomOp):
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
         if not x.is_contiguous():
-            # NOTE: Romove this if aiter kernel supports discontinuous input
+            # NOTE: Remove this if aiter kernel supports discontinuous input
             x = x.contiguous()
         if residual is not None:
             fused_add_rms_norm(x, residual, self.weight.data, self.variance_epsilon)

python/sglang/srt/layers/moe/ep_moe/kernels.py

@@ -116,7 +116,7 @@ def deepep_run_moe_deep_preprocess(topk_ids: torch.Tensor, num_experts: int):
     seg_indptr = torch.empty(num_experts + 1, device=topk_ids.device, dtype=torch.int64)
     src2dst = torch.empty(topk_ids.numel(), device=topk_ids.device, dtype=torch.int64)
 
-    # Find offet
+    # Find offset
     expert_ids = torch.arange(
         num_experts + 1, device=topk_ids.device, dtype=reorder_topk_ids.dtype
     )

python/sglang/srt/layers/moe/ep_moe/layer.py

@@ -611,7 +611,7 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
                 self.quant_config.weight_block_size[1],
             )
             # NOTE(HandH1998): To ensure proper alignment of the block-wise quantization scales, the output_size of the weights for both the gate and up layers must be divisible by block_n.
-            # Required by collum parallel or enabling merged weights
+            # Required by column parallel or enabling merged weights
             if intermediate_size % block_n != 0:
                 raise ValueError(
                     f"The output_size of gate's and up's weight = "

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -994,7 +994,7 @@ def get_default_config(
                     "num_stages": 2 if _is_hip else 4,
                 }
         else:
-            # Block-wise quant: BLOCK_SIZE_K must be divisable by block_shape[1]
+            # Block-wise quant: BLOCK_SIZE_K must be divisible by block_shape[1]
             config = {
                 "BLOCK_SIZE_M": 64,
                 "BLOCK_SIZE_N": block_shape[0],

python/sglang/srt/layers/moe/topk.py

@@ -270,7 +270,7 @@ def select_experts(
     routed_scaling_factor: Optional[float] = None,
 ):
     n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
-    # DeekSeek V2/V3/R1 serices models uses grouped_top_k
+    # DeepSeek V2/V3/R1 series models use grouped_top_k
     if use_grouped_topk:
         assert topk_group is not None
         assert num_expert_group is not None

python/sglang/srt/layers/quantization/__init__.py

@@ -109,7 +109,7 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
     if quantization in VLLM_QUANTIZATION_METHODS and not VLLM_AVAILABLE:
         raise ValueError(
             f"{quantization} quantization requires some operators from vllm. "
-            "Pleaes install vllm by `pip install vllm==0.8.4`"
+            "Please install vllm by `pip install vllm==0.8.4`"
         )
 
     return QUANTIZATION_METHODS[quantization]

python/sglang/srt/layers/quantization/blockwise_int8.py

@@ -152,7 +152,7 @@ class BlockInt8LinearMethod(LinearMethodBase):
                     f"{input_size_per_partition} is not divisible by "
                     f"weight quantization block_k = {block_k}."
                 )
-        # Required by collum parallel or enabling merged weights
+        # Required by column parallel or enabling merged weights
         if (tp_size > 1 and output_size // output_size_per_partition == tp_size) or len(
             output_partition_sizes
         ) > 1:
@@ -285,7 +285,7 @@ class BlockInt8MoEMethod:
             self.quant_config.weight_block_size[1],
         )
         # NOTE(HandH1998): To ensure proper alignment of the block-wise quantization scales, the output_size of the weights for both the gate and up layers must be divisible by block_n.
-        # Required by collum parallel or enabling merged weights
+        # Required by column parallel or enabling merged weights
         if intermediate_size % block_n != 0:
             raise ValueError(
                 f"The output_size of gate's and up's weight = "

python/sglang/srt/layers/quantization/deep_gemm.py

@@ -103,10 +103,10 @@ _INITIALIZATION_DICT: Dict[Tuple[DeepGemmKernelType, int, int, int], bool] = dic
 def _compile_warning_1():
     if not _IN_PRECOMPILE_STAGE and _IS_FIRST_RANK_ON_NODE:
         logger.warning(
-            "Entering DeepGEMM JIT Pre-Complie session. "
+            "Entering DeepGEMM JIT Pre-Compile session. "
             "And it may takes a long time(Typically 10-20 mins) "
             "if you have not run `sglang.compile_deep_gemm`. "
-            "Recommand to run `sglang.compile_deep_gemm` with same args as `sglang.launch_server`"
+            "It is recommended to run `sglang.compile_deep_gemm` with same args as `sglang.launch_server`"
             " for pre-compilation to reduce the overhead if you have not run it before. "
             "For example: "
             "`python3 -m sglang.compile_deep_gemm --model deepseek-ai/DeepSeek-V3 --tp 8 --trust-remote-code`"
@@ -115,7 +115,7 @@ def _compile_warning_1():
 
 def _compile_warning_2():
     logger.warning(
-        "Entering DeepGEMM JIT Single Kernel Complie session. "
+        "Entering DeepGEMM JIT Single Kernel Compile session. "
         "And it will makes inference throughput becomes flaky. "
         "Please run `sglang.compile_deep_gemm` with same args as `sglang.launch_server`"
         " for pre-compilation to solve this issue. "
@@ -298,7 +298,7 @@ def _maybe_compile_deep_gemm_one_type_all(
         logger.info(
             f"Try DeepGEMM JIT Compiling for "
             f"<{kernel_helper.name}> N={n}, K={k}, num_groups={num_groups} with all Ms."
-            f"{' It only takes a litte time (typically 1 sec) if you have run `python3 -m sglang.compile_deep_gemm`. ' if not _IN_PRECOMPILE_STAGE else ''}"
+            f"{' It only takes a little time (typically 1 sec) if you have run `python3 -m sglang.compile_deep_gemm`. ' if not _IN_PRECOMPILE_STAGE else ''}"
         )
 
         # NOTE(alcanderian): get_num_sms should be change when 2-batch-overlap is introduced