Add typo checker in pre-commit (#6179)

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

python/sglang/srt/models/llama4.py

@@ -428,7 +428,7 @@ class Llama4DecoderLayer(nn.Module):
         # Fully Connected
         hidden_states = self.feed_forward(hidden_states, forward_batch)
 
-        # TODO(ch-wan): ues reduce-scatter in MLP to avoid this scatter
+        # TODO(ch-wan): use reduce-scatter in MLP to avoid this scatter
         # Scatter
         if self.dp_size != 1:
             # important: forward batch.gathered_buffer is used both after scatter and after gather.

python/sglang/srt/models/roberta.py

@@ -57,7 +57,7 @@ class RobertaEmbedding(nn.Module):
         input_shape = input_ids.size()
         inputs_embeds = self.word_embeddings(input_ids)
 
-        # adpated from vllm: https://github.com/vllm-project/vllm/commit/4a18fd14ba4a349291c798a16bf62fa8a9af0b6b/vllm/model_executor/models/roberta.py
+        # Adapted from vllm: https://github.com/vllm-project/vllm/commit/4a18fd14ba4a349291c798a16bf62fa8a9af0b6b/vllm/model_executor/models/roberta.py
 
         pos_list = []
         token_list = []

python/sglang/srt/models/torch_native_llama.py

@@ -37,7 +37,7 @@ $ python3 -m sglang.bench_one_batch --correct \
   --tensor-parallel-size 2 \
   --disable-cuda-graph
 ```
-We will eanble CUDA Graph support soon.
+We will enable CUDA Graph support soon.
 """
 
 import types

python/sglang/srt/openai_api/adapter.py

@@ -590,7 +590,7 @@ def v1_generate_response(
     echo = False
 
     if (not isinstance(request, list)) and request.echo:
-        # TODO: handle the case propmt is token ids
+        # TODO: handle the case prompt is token ids
         if isinstance(request.prompt, list) and isinstance(request.prompt[0], str):
             # for the case of multiple str prompts
             prompts = request.prompt
@@ -646,7 +646,7 @@ def v1_generate_response(
         finish_reason = ret_item["meta_info"]["finish_reason"]
 
         if to_file:
-            # to make the choise data json serializable
+            # to make the choice data json serializable
             choice_data = {
                 "index": 0,
                 "text": text,

python/sglang/srt/reasoning_parser.py

@@ -147,7 +147,7 @@ class ReasoningParser:
 
     Args:
         model_type (str): Type of model to parse reasoning from
-        stream_reasoning (bool): If Flase, accumulates reasoning content until complete.
+        stream_reasoning (bool): If False, accumulates reasoning content until complete.
             If True, streams reasoning content as it arrives.
     """
 

python/sglang/srt/sampling/sampling_batch_info.py

@@ -294,7 +294,7 @@ class SamplingBatchInfo:
             # Set the flag to True if any of the two has custom logit processor
             self.has_custom_logit_processor = True
 
-        # Note: becasue the __len()__ operator is defined on the temperatures tensor,
+        # Note: because the __len()__ operator is defined on the temperatures tensor,
         # please make sure any merge operation with len(self) or len(other) is done before
         # the merge operation of the temperatures tensor below.
         for item in [

python/sglang/srt/server_args.py

@@ -825,7 +825,7 @@ class ServerArgs:
         # Multi-node distributed serving
         parser.add_argument(
             "--dist-init-addr",
-            "--nccl-init-addr",  # For backward compatbility. This will be removed in the future.
+            "--nccl-init-addr",  # For backward compatibility. This will be removed in the future.
             type=str,
             help="The host address for initializing distributed backend (e.g., `192.168.0.2:25000`).",
         )
@@ -1096,7 +1096,7 @@ class ServerArgs:
         parser.add_argument(
             "--triton-attention-reduce-in-fp32",
             action="store_true",
-            help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16."
+            help="Cast the intermediate attention results to fp32 to avoid possible crashes related to fp16."
             "This only affects Triton attention kernels.",
         )
         parser.add_argument(
@@ -1188,7 +1188,7 @@ class ServerArgs:
             type=int,
             default=0,
             help="The number of shared_experts need to be replicated to fuse with normal experts in deepseek v3/r1, "
-            "set it to tp_size can get best optimized performace.",
+            "set it to tp_size can get best optimized performance.",
         )
         parser.add_argument(
             "--disable-chunked-prefix-cache",

python/sglang/srt/speculative/eagle_draft_cuda_graph_runner.py

@@ -82,12 +82,12 @@ class EAGLEDraftCudaGraphRunner:
             self.capture()
         except RuntimeError as e:
             raise Exception(
-                f"Capture cuda graph failed: {e}\n"
+                f"Capture CUDA graph failed: {e}\n"
                 "Possible solutions:\n"
                 "1. set --mem-fraction-static to a smaller value (e.g., 0.8 or 0.7)\n"
                 "2. set --cuda-graph-max-bs to a smaller value (e.g., 16)\n"
                 "3. disable torch compile by not using --enable-torch-compile\n"
-                "4. disable cuda graph by --disable-cuda-graph. (Not recommonded. Huge perf loss)\n"
+                "4. disable CUDA graph by --disable-cuda-graph. (Not recommended. Huge performance loss)\n"
                 "Open an issue on GitHub https://github.com/sgl-project/sglang/issues/new/choose \n"
             )
 
@@ -149,7 +149,7 @@ class EAGLEDraftCudaGraphRunner:
 
         # Run and capture
         def run_once():
-            # Backup two fileds, which will be modified in-place in `draft_forward`.
+            # Backup two fields, which will be modified in-place in `draft_forward`.
             output_cache_loc_backup = forward_batch.out_cache_loc
             hidden_states_backup = forward_batch.spec_info.hidden_states
 

python/sglang/srt/speculative/eagle_utils.py

@@ -167,12 +167,12 @@ class EagleVerifyOutput:
     draft_input: EagleDraftInput
     # Logit outputs from target worker
     logits_output: LogitsProcessorOutput
-    # Accepeted token ids including the bonus token
+    # Accepted token ids including the bonus token
     verified_id: torch.Tensor
-    # Accepeted token length per sequence in a batch in CPU.
+    # Accepted token length per sequence in a batch in CPU.
     accept_length_per_req_cpu: List[int]
-    # Accepeted indices from logits_output.next_token_logits
-    accepeted_indices: torch.Tensor
+    # Accepted indices from logits_output.next_token_logits
+    accepted_indices: torch.Tensor
 
 
 @dataclass
@@ -316,7 +316,7 @@ class EagleVerifyInput:
 
         This API updates values inside logits_output based on the accepted
         tokens. I.e., logits_output.next_token_logits only contains
-        accepeted token logits.
+        accepted token logits.
         """
         bs = self.retrive_index.shape[0]
         candidates = self.draft_token.reshape(bs, self.draft_token_num)
@@ -493,7 +493,7 @@ class EagleVerifyInput:
                 logits_output=logits_output,
                 verified_id=verified_id,
                 accept_length_per_req_cpu=accept_length_cpu,
-                accepeted_indices=accept_index,
+                accepted_indices=accept_index,
             )
         else:
             assign_req_to_token_pool[(bs,)](
@@ -539,7 +539,7 @@ class EagleVerifyInput:
                 logits_output=logits_output,
                 verified_id=verified_id,
                 accept_length_per_req_cpu=accept_length_cpu,
-                accepeted_indices=accept_index,
+                accepted_indices=accept_index,
             )
 
 

python/sglang/srt/speculative/eagle_worker.py

@@ -201,7 +201,7 @@ class EAGLEWorker(TpModelWorker):
             self.has_prefill_wrapper_verify = False
         else:
             raise ValueError(
-                f"EAGLE is not supportted in attention backend {self.server_args.attention_backend}"
+                f"EAGLE is not supported in attention backend {self.server_args.attention_backend}"
             )
 
         self.draft_model_runner.draft_attn_backend = self.draft_attn_backend
@@ -245,8 +245,8 @@ class EAGLEWorker(TpModelWorker):
         Args:
             batch: The batch to run forward. The state of the batch is modified as it runs.
         Returns:
-            A tuple of the final logit output of the target model, next tokens accepeted,
-            the batch id (used for overlap schedule), and number of accepeted tokens.
+            A tuple of the final logit output of the target model, next tokens accepted,
+            the batch id (used for overlap schedule), and number of accepted tokens.
         """
         if batch.forward_mode.is_decode():
             with self.draft_tp_context(self.draft_model_runner.tp_group):
@@ -491,11 +491,11 @@ class EAGLEWorker(TpModelWorker):
         )
 
         # Post process based on verified outputs.
-        # Pick indices that we care (accepeted)
+        # Pick indices that we care (accepted)
         logits_output.next_token_logits = logits_output.next_token_logits[
-            res.accepeted_indices
+            res.accepted_indices
         ]
-        logits_output.hidden_states = logits_output.hidden_states[res.accepeted_indices]
+        logits_output.hidden_states = logits_output.hidden_states[res.accepted_indices]
 
         # Prepare the batch for the next draft forwards.
         batch.forward_mode = ForwardMode.DECODE
@@ -597,7 +597,7 @@ class EAGLEWorker(TpModelWorker):
         self.capture_for_decode(logits_output, forward_batch.spec_info)
 
     def forward_draft_extend_after_decode(self, batch: ScheduleBatch):
-        # Backup fileds that will be modified in-place
+        # Backup fields that will be modified in-place
         seq_lens_backup = batch.seq_lens.clone()
         req_pool_indices_backup = batch.req_pool_indices
         accept_length_backup = batch.spec_info.accept_length

python/sglang/test/simple_eval_common.py

@@ -140,7 +140,7 @@ class ChatCompletionSampler(SamplerBase):
                     max_tokens=self.max_tokens,
                 )
                 return response.choices[0].message.content
-            # NOTE: BadRequestError is triggered once for MMMU, please uncomment if you are reruning MMMU
+            # NOTE: BadRequestError is triggered once for MMMU, please uncomment if you are rerunning MMMU
             except openai.BadRequestError as e:
                 print("Bad Request Error", e)
                 return ""

python/sglang/test/simple_eval_humaneval.py

@@ -121,7 +121,7 @@ class HumanEval(Eval):
                 convo=convo,
                 metrics={
                     f"pass@{k}": estimate_pass_at_k([total], [correct], k)
-                    # this will be aggrated so no need of .mean()
+                    # this will be aggregated so no need of .mean()
                     for k in self._ks_passes
                     if total >= k
                 },

python/sglang/test/test_programs.py

@@ -370,7 +370,7 @@ def test_dtype_gen():
     @sgl.function
     def dtype_gen(s):
         s += "Q: What is the full name of DNS?\n"
-        s += "A: The full nams is " + sgl.gen("str_res", dtype=str, stop="\n") + "\n"
+        s += "A: The full names is " + sgl.gen("str_res", dtype=str, stop="\n") + "\n"
         s += "Q: Which year was DNS invented?\n"
         s += "A: " + sgl.gen("int_res", dtype=int) + "\n"
         s += "Q: What is the value of pi?\n"

python/sglang/utils.py

@@ -278,7 +278,7 @@ def graceful_registry(sub_module_name: str):
             f"{sub_module_name} Received signal to shutdown. Performing graceful shutdown..."
         )
         if signum == signal.SIGTERM:
-            logger.info(f"{sub_module_name} recive sigterm")
+            logger.info(f"{sub_module_name} receive sigterm")
 
     signal.signal(signal.SIGTERM, graceful_shutdown)
 

scripts/ci_install_dependency.sh

@@ -25,7 +25,7 @@ pip install -e "python[all]"
 pip install torch_memory_saver
 pip install transformers==4.51.0 sentence_transformers accelerate peft pandas datasets timm torchaudio==2.6.0
 
-# For compling xgrammar kernels
+# For compiling xgrammar kernels
 pip install cuda-python nvidia-cuda-nvrtc-cu12
 
 # For lmms_evals evaluating MMMU

scripts/ci_install_dependency_8_gpu.sh

@@ -43,7 +43,7 @@ pip install -e "python[all]"
 pip install torch_memory_saver
 pip install transformers==4.51.0 sentence_transformers accelerate peft pandas datasets timm torchaudio==2.6.0
 
-# For compling xgrammar kernels
+# For compiling xgrammar kernels
 pip install cuda-python nvidia-cuda-nvrtc-cu12
 
 # For lmms_evals evaluating MMMU

scripts/deprecated/convert_yi_vl.py

 """
-Convert Yi-VL config into a format useable with SGLang
+Convert Yi-VL config into a format usable with SGLang
 
 Usage: python3 scripts/convert_yi_vl.py --model-path <path-to-model>
 """

scripts/export_deepseek_nextn.py

@@ -90,7 +90,7 @@ def export_nextn_layer_parameters(input_dir, output_dir, nextn_layer_id):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="Export NextN layer paramerters for DeepSeek-V3/R1"
+        description="Export NextN layer parameters for DeepSeek-V3/R1"
     )
     parser.add_argument(
         "--input-dir",

sgl-kernel/CMakeLists.txt

@@ -114,7 +114,7 @@ set(SGL_KERNEL_CUDA_FLAGS
     "--expt-extended-lambda"
     "--threads=32"
 
-    # Supress warnings
+    # Suppress warnings
     "-Xcompiler=-Wconversion"
     "-Xcompiler=-fno-strict-aliasing"
 

sgl-kernel/README.md

@@ -87,7 +87,7 @@ Third-party libraries:
 
   The main different Between sm80/sm87 and sm86/sm89 is the shared memory size. you can follow the link below for more information https://docs.nvidia.com/cuda/cuda-c-programming-guide/#shared-memory-8-x.
 
-  And for sgl-kernel right now, we can build fa3 on sm80/sm86/sm89/sm90a. Thats mean if you use **A100(tested)**/A*0/**L20(tested)**/L40/L40s/**3090(tested)** you can use fa3.
+  And for sgl-kernel right now, we can build fa3 on sm80/sm86/sm89/sm90a. That means if you use **A100(tested)**/A*0/**L20(tested)**/L40/L40s/**3090(tested)** you can use fa3.
 
 ### Kernel Development