misc: add pre-commit config (#637)

python/sglang/srt/models/qwen2_moe.py

@@ -401,9 +401,11 @@ class Qwen2MoeForCausalLM(nn.Module):
             # These are the weights for the experts
             # (param_name, weight_name, expert_id, shard_id)
             (
-                "experts.w13_weight"
-                if weight_name in ["gate_proj", "up_proj"]
-                else "experts.w2_weight",
+                (
+                    "experts.w13_weight"
+                    if weight_name in ["gate_proj", "up_proj"]
+                    else "experts.w2_weight"
+                ),
                 f"experts.{expert_id}.{weight_name}.weight",
                 expert_id,
                 shard_id,
@@ -418,7 +420,7 @@ class Qwen2MoeForCausalLM(nn.Module):
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
-            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+            for param_name, weight_name, shard_id in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name:
                     continue

python/sglang/srt/server.py

@@ -32,8 +32,8 @@ from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.managers.controller.manager_multi import (
     start_controller_process as start_controller_process_multi,
 )
+from sglang.srt.managers.controller.manager_single import launch_tp_servers
 from sglang.srt.managers.controller.manager_single import (
-    launch_tp_servers,
     start_controller_process as start_controller_process_single,
 )
 from sglang.srt.managers.detokenizer_manager import start_detokenizer_process
@@ -198,11 +198,22 @@ def launch_server(server_args: ServerArgs, pipe_finish_writer, model_overide_arg
 
         if server_args.node_rank != 0:
             tp_size_local = server_args.tp_size // server_args.nnodes
-            gpu_ids = [i for _ in range(server_args.nnodes) for i in range(tp_size_local)]
-            tp_rank_range = list(range(server_args.node_rank * tp_size_local,
-                                  (server_args.node_rank + 1) * tp_size_local))
-            procs = launch_tp_servers(gpu_ids, tp_rank_range, server_args,
-                                      port_args.model_port_args[0], model_overide_args)
+            gpu_ids = [
+                i for _ in range(server_args.nnodes) for i in range(tp_size_local)
+            ]
+            tp_rank_range = list(
+                range(
+                    server_args.node_rank * tp_size_local,
+                    (server_args.node_rank + 1) * tp_size_local,
+                )
+            )
+            procs = launch_tp_servers(
+                gpu_ids,
+                tp_rank_range,
+                server_args,
+                port_args.model_port_args[0],
+                model_overide_args,
+            )
             while True:
                 pass
 

scripts/convert_yi_vl.py

@@ -10,16 +10,15 @@ import os
 
 from transformers import AutoConfig, AutoTokenizer
 
+
 def add_image_token(model_path: str):
     tokenizer = AutoTokenizer.from_pretrained(model_path)
-    tokenizer.add_tokens(
-        ["<image_placeholder>"],
-        special_tokens=True
-    )
+    tokenizer.add_tokens(["<image_placeholder>"], special_tokens=True)
 
     print(tokenizer)
     tokenizer.save_pretrained(model_path)
 
+
 def edit_model_config(model_path):
     config = AutoConfig.from_pretrained(model_path)
 
@@ -29,10 +28,11 @@ def edit_model_config(model_path):
     print(config)
     config.save_pretrained(model_path)
 
+
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--model-path", type=str)
     args = parser.parse_args()
 
     add_image_token(args.model_path)
-    edit_model_config(args.model_path)
\ No newline at end of file
+    edit_model_config(args.model_path)