Improve multi-node stability (#1171)

python/sglang/launch_server.py

 """Launch the inference server."""
 
 import argparse
+import os
 
 from sglang.srt.server import launch_server
 from sglang.srt.server_args import ServerArgs
+from sglang.srt.utils import kill_child_process
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
@@ -11,4 +13,9 @@ if __name__ == "__main__":
     args = parser.parse_args()
     server_args = ServerArgs.from_cli_args(args)
 
-    launch_server(server_args)
+    try:
+        launch_server(server_args)
+    except Exception as e:
+        raise e
+    finally:
+        kill_child_process(os.getpid(), including_parent=False)

python/sglang/srt/hf_transformers_utils.py

@@ -233,6 +233,8 @@ class TiktokenTokenizer:
         }
         assert tok_dict["word_split"] == "V1"
 
+        default_allowed_special = None
+
         kwargs = {
             "name": name,
             "pat_str": tok_dict.get("pat_str", PAT_STR_B),
@@ -246,14 +248,18 @@ class TiktokenTokenizer:
                     for bytes_list in tok_dict["default_allowed_special"]
                 ]
             )
-        else:
-            default_allowed_special = None
         if "vocab_size" in tok_dict:
             kwargs["explicit_n_vocab"] = tok_dict["vocab_size"]
 
+        PAD = "<|pad|>"
+        EOS = "<|eos|>"
+        SEP = "<|separator|>"
+
+        DEFAULT_CONTROL_TOKENS = {"pad": PAD, "sep": EOS, "eos": SEP}
+
         tokenizer = tiktoken.Encoding(**kwargs)
         tokenizer._default_allowed_special = default_allowed_special or set()
-        tokenizer._default_allowed_special |= {"<|separator|>"}
+        tokenizer._control_tokens = DEFAULT_CONTROL_TOKENS
 
         def encode_patched(
             self,
@@ -270,14 +276,14 @@ class TiktokenTokenizer:
                 self,
                 text,
                 allowed_special=allowed_special,
-                disallowed_special=disallowed_special,
+                disallowed_special=(),
             )
 
         tokenizer.encode = functools.partial(encode_patched, tokenizer)
 
         # Convert to HF interface
         self.tokenizer = tokenizer
-        self.eos_token_id = tokenizer._special_tokens["<|eos|>"]
+        self.eos_token_id = tokenizer._special_tokens[EOS]
         self.vocab_size = tokenizer.n_vocab
         self.chat_template = Template(
             "{% for message in messages %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'].strip() + '<|separator|>\n\n' }}{% elif message['role'] == 'system' %}{{ 'System: ' + message['content'].strip() + '<|separator|>\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: '  + message['content'] + '<|separator|>\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}"

python/sglang/srt/managers/controller_multi.py

@@ -212,6 +212,4 @@ def start_controller_process(
     except Exception:
         logger.error("Exception in ControllerMulti:\n" + get_exception_traceback())
     finally:
-        for w in controller.workers:
-            os.kill(w.proc.pid, 9)
         kill_parent_process()

python/sglang/srt/managers/controller_single.py

@@ -167,6 +167,4 @@ def start_controller_process(
     except Exception:
         logger.error("Exception in ControllerSingle:\n" + get_exception_traceback())
     finally:
-        for t in controller.tp_procs:
-            os.kill(t.pid, 9)
         kill_parent_process()

python/sglang/srt/managers/schedule_batch.py

@@ -16,7 +16,6 @@ limitations under the License.
 """Meta data for requests and batches"""
 
 import logging
-import warnings
 from dataclasses import dataclass
 from typing import List, Optional, Union
 
@@ -270,7 +269,7 @@ class Req:
 
         if all_ids[prompt_tokens - 1] != self.origin_input_ids_unpadded[-1]:
             # TODO(lsyin): fix token fusion
-            logging.warning(
+            logger.warning(
                 "Token fusion between input and output, try to avoid this by removing the space at the end of the input."
             )
             return False
@@ -753,7 +752,7 @@ class ScheduleBatch:
                 )
             self.logit_bias = torch.concat([self.logit_bias, other.logit_bias])
 
-    def sample(self, logits: torch.Tensor, is_multi_node_tp=False):
+    def sample(self, logits: torch.Tensor):
         # TODO(lsyin): move this into a part of layer and run with CUDA Graph
         # Post process logits
         logits = logits.contiguous()
@@ -791,7 +790,7 @@ class ScheduleBatch:
             )
 
         if not torch.all(success):
-            logging.warning("Sampling failed, fallback to top_k=1 strategy")
+            logger.warning(f"Sampling failed. Fallback to top_k=1 strategy. {logits=}")
             probs = probs.masked_fill(torch.isnan(probs), 0.0)
             argmax_ids = torch.argmax(probs, dim=-1)
             batch_next_token_ids = torch.where(
@@ -808,16 +807,6 @@ class ScheduleBatch:
 
         self.penalizer_orchestrator.cumulate_output_tokens(batch_next_token_ids)
 
-        if is_multi_node_tp:
-            # If the tensor parallelism spans across multiple nodes, there is some indeterminism
-            # that can cause the TP workers to generate different tokens, so we need to
-            # sync here
-            torch.distributed.all_reduce(
-                batch_next_token_ids,
-                op=dist.ReduceOp.MIN,
-                group=get_tensor_model_parallel_group().device_group,
-            )
-
         return batch_next_token_ids
 
 
@@ -835,7 +824,8 @@ def top_k_top_p_sampling_from_probs_torch(
     probs_sort.div_(probs_sort.max(dim=-1, keepdim=True)[0])
     try:
         sampled_index = torch.multinomial(probs_sort, num_samples=1)
-    except RuntimeError:
+    except RuntimeError as e:
+        logger.warning(f"Sampling error: {e}")
         batch_next_token_ids = torch.zeros(
             (probs_sort.shape[0],), dtype=torch.int32, device=probs.device
         )

python/sglang/srt/managers/tp_worker.py

@@ -133,6 +133,13 @@ class ModelTpServer:
             self.model_config.context_len - 1,
             self.max_total_num_tokens - 1,
         )
+
+        # Sync random seed
+        server_args.random_seed = broadcast_recv_input(
+            [server_args.random_seed],
+            self.tp_rank,
+            self.model_runner.tp_group.cpu_group,
+        )[0]
         set_random_seed(server_args.random_seed)
 
         # Print info
@@ -474,9 +481,7 @@ class ModelTpServer:
             # Forward and sample the next tokens
             if batch.extend_num_tokens != 0:
                 output = self.model_runner.forward(batch, ForwardMode.EXTEND)
-                next_token_ids = batch.sample(
-                    output.next_token_logits, self.model_runner.is_multi_node_tp
-                )
+                next_token_ids = batch.sample(output.next_token_logits)
 
                 # Move logprobs to cpu
                 if output.next_token_logprobs is not None:
@@ -636,9 +641,7 @@ class ModelTpServer:
 
         # Forward and sample the next tokens
         output = self.model_runner.forward(batch, ForwardMode.DECODE)
-        next_token_ids = batch.sample(
-            output.next_token_logits, self.model_runner.is_multi_node_tp
-        )
+        next_token_ids = batch.sample(output.next_token_logits)
 
         # Move logprobs to cpu
         if output.next_token_logprobs is not None:
@@ -879,6 +882,7 @@ def broadcast_recv_input(
 
             dist.broadcast(tensor_size, src=0, group=dist_group)
             dist.broadcast(tensor_data, src=0, group=dist_group)
+        return data
     else:
         tensor_size = torch.tensor([0], dtype=torch.long)
         dist.broadcast(tensor_size, src=0, group=dist_group)

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -84,13 +84,20 @@ def set_torch_compile_config():
 
 
 class CudaGraphRunner:
-    def __init__(self, model_runner, max_batch_size_to_capture, use_torch_compile):
+    def __init__(
+        self,
+        model_runner,
+        max_batch_size_to_capture: int,
+        use_torch_compile: bool,
+        disable_padding: bool,
+    ):
         self.model_runner = model_runner
         self.graphs = {}
         self.input_buffers = {}
         self.output_buffers = {}
         self.flashinfer_handlers = {}
         self.graph_memory_pool = None
+        self.disable_padding = disable_padding
 
         # Common inputs
         self.max_bs = max_batch_size_to_capture
@@ -142,7 +149,10 @@ class CudaGraphRunner:
             set_torch_compile_config()
 
     def can_run(self, batch_size):
-        return batch_size <= self.max_bs
+        if self.disable_padding:
+            return batch_size in self.graphs
+        else:
+            return batch_size <= self.max_bs
 
     def capture(self, batch_size_list):
         self.batch_size_list = batch_size_list

python/sglang/srt/model_executor/model_runner.py

@@ -465,6 +465,7 @@ class ModelRunner:
             self,
             max_batch_size_to_capture=max(batch_size_list),
             use_torch_compile=self.server_args.enable_torch_compile,
+            disable_padding=self.server_args.disable_cuda_graph_padding,
         )
         try:
             self.cuda_graph_runner.capture(batch_size_list)

python/sglang/srt/server.py

@@ -24,7 +24,6 @@ import json
 import logging
 import multiprocessing as mp
 import os
-import sys
 import threading
 import time
 from http import HTTPStatus
@@ -301,27 +300,29 @@ def launch_server(
     server_args.tokenizer_path = prepare_tokenizer(server_args.tokenizer_path)
 
     # Launch processes for multi-node tensor parallelism
-    if server_args.nnodes > 1:
-        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,
-                )
+    if server_args.nnodes > 1 and 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,
-                ports[3],
-                model_overide_args,
-            )
-            while True:
-                pass
+        )
+        procs = launch_tp_servers(
+            gpu_ids,
+            tp_rank_range,
+            server_args,
+            ports[3],
+            model_overide_args,
+        )
+
+        try:
+            for p in procs:
+                p.join()
+        finally:
+            kill_child_process(os.getpid(), including_parent=False)
+            return
 
     # Launch processes
     tokenizer_manager = TokenizerManager(server_args, port_args, model_overide_args)
@@ -356,15 +357,11 @@ def launch_server(
     if controller_init_state != "init ok" or detoken_init_state != "init ok":
         proc_controller.kill()
         proc_detoken.kill()
-        print(
-            f"Initialization failed. controller_init_state: {controller_init_state}",
-            flush=True,
+        raise RuntimeError(
+            "Initialization failed. "
+            f"controller_init_state: {controller_init_state}, "
+            f"detoken_init_state: {detoken_init_state}"
         )
-        print(
-            f"Initialization failed. detoken_init_state: {detoken_init_state}",
-            flush=True,
-        )
-        sys.exit(1)
     assert proc_controller.is_alive() and proc_detoken.is_alive()
 
     # Add api key authorization
@@ -373,12 +370,12 @@ def launch_server(
 
     # Send a warmup request
     t = threading.Thread(
-        target=_wait_and_warmup, args=(server_args, pipe_finish_writer)
+        target=_wait_and_warmup, args=(server_args, pipe_finish_writer, os.getpid())
     )
     t.start()
 
-    # Listen for requests
     try:
+        # Listen for requests
         uvicorn.run(
             app,
             host=server_args.host,
@@ -426,7 +423,7 @@ def _set_envs_and_config(server_args: ServerArgs):
         )
 
 
-def _wait_and_warmup(server_args, pipe_finish_writer):
+def _wait_and_warmup(server_args, pipe_finish_writer, pid):
     headers = {}
     url = server_args.url()
     if server_args.api_key:
@@ -449,8 +446,9 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
     if not success:
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(last_traceback)
-        print(f"Initialization failed. warmup error: {last_traceback}", flush=True)
-        sys.exit(1)
+        logger.error(f"Initialization failed. warmup error: {last_traceback}")
+        kill_child_process(pid, including_parent=False)
+        return
 
     # Send a warmup request
     request_name = "/generate" if model_info["is_generation"] else "/encode"
@@ -475,12 +473,13 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
                 timeout=600,
             )
             assert res.status_code == 200, f"{res}"
-    except Exception as e:
+    except Exception:
         last_traceback = get_exception_traceback()
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(last_traceback)
-        print(f"Initialization failed. warmup error: {last_traceback}", flush=True)
-        sys.exit(1)
+        logger.error(f"Initialization failed. warmup error: {last_traceback}")
+        kill_child_process(pid, including_parent=False)
+        return
 
     logger.info("The server is fired up and ready to roll!")
     if pipe_finish_writer is not None:

python/sglang/srt/server_args.py

@@ -79,6 +79,7 @@ class ServerArgs:
     disable_radix_cache: bool = False
     disable_regex_jump_forward: bool = False
     disable_cuda_graph: bool = False
+    disable_cuda_graph_padding: bool = False
     disable_disk_cache: bool = False
     enable_mixed_chunk: bool = False
     enable_torch_compile: bool = False
@@ -393,6 +394,11 @@ class ServerArgs:
             action="store_true",
             help="Disable cuda graph.",
         )
+        parser.add_argument(
+            "--disable-cuda-graph-padding",
+            action="store_true",
+            help="Disable cuda graph when padding is needed. Still uses cuda graph when padding is not needed.",
+        )
         parser.add_argument(
             "--disable-disk-cache",
             action="store_true",

python/sglang/srt/utils.py

@@ -369,14 +369,11 @@ def kill_parent_process():
     """Kill the parent process and all children of the parent process."""
     current_process = psutil.Process()
     parent_process = current_process.parent()
-    children = parent_process.children(recursive=True)
-    for child in children:
-        if child.pid != current_process.pid:
-            os.kill(child.pid, 9)
-    os.kill(parent_process.pid, 9)
+    kill_child_process(parent_process.pid, skip_pid=current_process.pid)
 
 
-def kill_child_process(pid, including_parent=True):
+def kill_child_process(pid, including_parent=True, skip_pid=None):
+    """Kill the process and all its children process."""
     try:
         parent = psutil.Process(pid)
     except psutil.NoSuchProcess:
@@ -384,6 +381,8 @@ def kill_child_process(pid, including_parent=True):
 
     children = parent.children(recursive=True)
     for child in children:
+        if child.pid == skip_pid:
+            continue
         try:
             child.kill()
         except psutil.NoSuchProcess: