sglang 0.1.19__py3-none-any.whl → 0.1.20__py3-none-any.whl

sglang/srt/managers/controller/model_runner.py

@@ -4,11 +4,9 @@ import importlib
 import importlib.resources
 import logging
 import pkgutil
-from dataclasses import dataclass
 from functools import lru_cache
-from typing import List, Optional, Type
+from typing import Optional, Type
 
-import numpy as np
 import torch
 import torch.nn as nn
 from vllm.config import DeviceConfig, LoadConfig
@@ -17,7 +15,8 @@ from vllm.distributed import init_distributed_environment, initialize_model_para
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.models import ModelRegistry
 
-from sglang.srt.managers.controller.infer_batch import Batch, ForwardMode
+from sglang.global_config import global_config
+from sglang.srt.managers.controller.infer_batch import Batch, ForwardMode, InputMetadata, global_server_args_dict
 from sglang.srt.memory_pool import ReqToTokenPool, TokenToKVPool
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import (
@@ -29,210 +28,6 @@ from sglang.srt.utils import (
 
 logger = logging.getLogger("srt.model_runner")
 
-# for server args in model endpoints
-global_server_args_dict = {}
-
-
-@dataclass
-class InputMetadata:
-    forward_mode: ForwardMode
-    batch_size: int
-    total_num_tokens: int
-    max_seq_len: int
-    req_pool_indices: torch.Tensor
-    start_loc: torch.Tensor
-    seq_lens: torch.Tensor
-    prefix_lens: torch.Tensor
-    positions: torch.Tensor
-    req_to_token_pool: ReqToTokenPool
-    token_to_kv_pool: TokenToKVPool
-
-    # for extend
-    extend_seq_lens: torch.Tensor = None
-    extend_start_loc: torch.Tensor = None
-    max_extend_len: int = 0
-
-    out_cache_loc: torch.Tensor = None
-    out_cache_cont_start: torch.Tensor = None
-    out_cache_cont_end: torch.Tensor = None
-
-    other_kv_index: torch.Tensor = None
-    return_logprob: bool = False
-    top_logprobs_nums: List[int] = None
-
-    # for flashinfer
-    qo_indptr: torch.Tensor = None
-    kv_indptr: torch.Tensor = None
-    kv_indices: torch.Tensor = None
-    kv_last_page_len: torch.Tensor = None
-    flashinfer_prefill_wrapper_ragged: "BatchPrefillWithRaggedKVCacheWrapper" = None
-    flashinfer_prefill_wrapper_paged: "BatchPrefillWithPagedKVCacheWrapper" = None
-    flashinfer_decode_wrapper: "BatchDecodeWithPagedKVCacheWrapper" = None
-
-    def init_flashinfer_args(self, num_qo_heads, num_kv_heads, head_dim):
-        if (
-            self.forward_mode == ForwardMode.PREFILL
-            or self.forward_mode == ForwardMode.EXTEND
-        ):
-            paged_kernel_lens = self.prefix_lens
-            self.no_prefix = torch.all(self.prefix_lens == 0)
-        else:
-            paged_kernel_lens = self.seq_lens
-
-        self.kv_indptr = torch.zeros(
-            (self.batch_size + 1,), dtype=torch.int32, device="cuda"
-        )
-        self.kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
-        self.kv_last_page_len = torch.ones(
-            (self.batch_size,), dtype=torch.int32, device="cuda"
-        )
-        req_pool_indices_cpu = self.req_pool_indices.cpu().numpy()
-        paged_kernel_lens_cpu = paged_kernel_lens.cpu().numpy()
-        self.kv_indices = torch.cat(
-            [
-                self.req_to_token_pool.req_to_token[
-                    req_pool_indices_cpu[i], : paged_kernel_lens_cpu[i]
-                ]
-                for i in range(self.batch_size)
-            ],
-            dim=0,
-        ).contiguous()
-
-        if (
-            self.forward_mode == ForwardMode.PREFILL
-            or self.forward_mode == ForwardMode.EXTEND
-        ):
-            # extend part
-            self.qo_indptr = torch.zeros(
-                (self.batch_size + 1,), dtype=torch.int32, device="cuda"
-            )
-            self.qo_indptr[1:] = torch.cumsum(self.extend_seq_lens, dim=0)
-
-            self.flashinfer_prefill_wrapper_ragged.end_forward()
-            self.flashinfer_prefill_wrapper_ragged.begin_forward(
-                self.qo_indptr,
-                self.qo_indptr.clone(),
-                num_qo_heads,
-                num_kv_heads,
-                head_dim,
-            )
-
-            # cached part
-            self.flashinfer_prefill_wrapper_paged.end_forward()
-            self.flashinfer_prefill_wrapper_paged.begin_forward(
-                self.qo_indptr,
-                self.kv_indptr,
-                self.kv_indices,
-                self.kv_last_page_len,
-                num_qo_heads,
-                num_kv_heads,
-                head_dim,
-                1,
-            )
-        else:
-            self.flashinfer_decode_wrapper.end_forward()
-            self.flashinfer_decode_wrapper.begin_forward(
-                self.kv_indptr,
-                self.kv_indices,
-                self.kv_last_page_len,
-                num_qo_heads,
-                num_kv_heads,
-                head_dim,
-                1,
-                pos_encoding_mode="NONE",
-                data_type=self.token_to_kv_pool.kv_data[0].dtype,
-            )
-
-    def init_extend_args(self):
-        self.extend_seq_lens = self.seq_lens - self.prefix_lens
-        self.extend_start_loc = torch.zeros_like(self.seq_lens)
-        self.extend_start_loc[1:] = torch.cumsum(self.extend_seq_lens[:-1], dim=0)
-        self.max_extend_len = int(torch.max(self.extend_seq_lens))
-
-    @classmethod
-    def create(
-        cls,
-        model_runner,
-        tp_size,
-        forward_mode,
-        req_pool_indices,
-        seq_lens,
-        prefix_lens,
-        position_ids_offsets,
-        out_cache_loc,
-        out_cache_cont_start=None,
-        out_cache_cont_end=None,
-        top_logprobs_nums=None,
-        return_logprob=False,
-        flashinfer_prefill_wrapper_ragged=None,
-        flashinfer_prefill_wrapper_paged=None,
-        flashinfer_decode_wrapper=None,
-    ):
-        batch_size = len(req_pool_indices)
-        start_loc = torch.zeros((batch_size,), dtype=torch.int32, device="cuda")
-        start_loc[1:] = torch.cumsum(seq_lens[:-1], dim=0)
-        total_num_tokens = int(torch.sum(seq_lens))
-        max_seq_len = int(torch.max(seq_lens))
-
-        if forward_mode == ForwardMode.DECODE:
-            positions = ((seq_lens - 1) + position_ids_offsets).to(torch.int64)
-            other_kv_index = model_runner.req_to_token_pool.req_to_token[
-                req_pool_indices[0], seq_lens[0] - 1
-            ].item()
-        else:
-            seq_lens_cpu = seq_lens.cpu().numpy()
-            prefix_lens_cpu = prefix_lens.cpu().numpy()
-            position_ids_offsets_cpu = position_ids_offsets.cpu().numpy()
-            positions = torch.tensor(
-                np.concatenate(
-                    [
-                        np.arange(
-                            prefix_lens_cpu[i] + position_ids_offsets_cpu[i],
-                            seq_lens_cpu[i] + position_ids_offsets_cpu[i],
-                        )
-                        for i in range(batch_size)
-                    ],
-                    axis=0,
-                ),
-                device="cuda",
-            )
-            other_kv_index = None
-
-        ret = cls(
-            forward_mode=forward_mode,
-            batch_size=batch_size,
-            total_num_tokens=total_num_tokens,
-            max_seq_len=max_seq_len,
-            req_pool_indices=req_pool_indices,
-            start_loc=start_loc,
-            seq_lens=seq_lens,
-            prefix_lens=prefix_lens,
-            positions=positions,
-            req_to_token_pool=model_runner.req_to_token_pool,
-            token_to_kv_pool=model_runner.token_to_kv_pool,
-            out_cache_loc=out_cache_loc,
-            out_cache_cont_start=out_cache_cont_start,
-            out_cache_cont_end=out_cache_cont_end,
-            other_kv_index=other_kv_index,
-            return_logprob=return_logprob,
-            top_logprobs_nums=top_logprobs_nums,
-            flashinfer_prefill_wrapper_ragged=flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=flashinfer_decode_wrapper,
-        )
-
-        if forward_mode == ForwardMode.EXTEND:
-            ret.init_extend_args()
-
-        if not global_server_args_dict.get("disable_flashinfer", False):
-            ret.init_flashinfer_args(
-                model_runner.model_config.num_attention_heads // tp_size,
-                model_runner.model_config.get_num_kv_heads(tp_size),
-                model_runner.model_config.head_dim,
-            )
-
-        return ret
-
 
 class ModelRunner:
     def __init__(
@@ -245,6 +40,7 @@ class ModelRunner:
         nccl_port: int,
         server_args: ServerArgs,
     ):
+        # Parse args
         self.model_config = model_config
         self.mem_fraction_static = mem_fraction_static
         self.gpu_id = gpu_id
@@ -256,7 +52,6 @@ class ModelRunner:
         monkey_patch_vllm_dummy_weight_loader()
 
         # Init torch distributed
-        logger.info(f"[gpu_id={self.gpu_id}] Set cuda device.")
         torch.cuda.set_device(self.gpu_id)
         logger.info(f"[gpu_id={self.gpu_id}] Init nccl begin.")
 
@@ -287,11 +82,8 @@ class ModelRunner:
                 )
 
         # Set some global args
-        global global_server_args_dict
-        global_server_args_dict = {
-            "disable_flashinfer": server_args.disable_flashinfer,
-            "attention_reduce_in_fp32": server_args.attention_reduce_in_fp32,
-        }
+        global_server_args_dict["disable_flashinfer"] = server_args.disable_flashinfer
+        global_server_args_dict["attention_reduce_in_fp32"] = server_args.attention_reduce_in_fp32
 
         # Load the model and create memory pool
         self.load_model()
@@ -299,6 +91,9 @@ class ModelRunner:
         self.init_cublas()
         self.init_flash_infer()
 
+        # Capture cuda graphs
+        self.init_cuda_graphs()
+
     def load_model(self):
         logger.info(
             f"[gpu_id={self.gpu_id}] Load weight begin. "
@@ -391,67 +186,60 @@ class ModelRunner:
         return c
 
     def init_flash_infer(self):
-        if not global_server_args_dict.get("disable_flashinfer", False):
-            from flashinfer import (
-                BatchDecodeWithPagedKVCacheWrapper,
-                BatchPrefillWithPagedKVCacheWrapper,
-                BatchPrefillWithRaggedKVCacheWrapper,
-            )
-            from flashinfer.decode import _grouped_size_compiled_for_decode_kernels
+        if self.server_args.disable_flashinfer:
+            self.flashinfer_prefill_wrapper_ragged = None
+            self.flashinfer_prefill_wrapper_paged = None
+            self.flashinfer_decode_wrapper = None
+            return
 
-            if not _grouped_size_compiled_for_decode_kernels(
-                self.model_config.num_attention_heads // self.tp_size,
-                self.model_config.get_num_kv_heads(self.tp_size),
-            ):
-                use_tensor_cores = True
-            else:
-                use_tensor_cores = False
+        from flashinfer import (
+            BatchDecodeWithPagedKVCacheWrapper,
+            BatchPrefillWithPagedKVCacheWrapper,
+            BatchPrefillWithRaggedKVCacheWrapper,
+        )
+        from flashinfer.decode import _grouped_size_compiled_for_decode_kernels
 
-            workspace_buffers = torch.empty(
-                2, 96 * 1024 * 1024, dtype=torch.uint8, device="cuda"
-            )
-            self.flashinfer_prefill_wrapper_ragged = (
-                BatchPrefillWithRaggedKVCacheWrapper(workspace_buffers[0], "NHD")
-            )
-            self.flashinfer_prefill_wrapper_paged = BatchPrefillWithPagedKVCacheWrapper(
-                workspace_buffers[1], "NHD"
-            )
-            self.flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
-                workspace_buffers[0], "NHD", use_tensor_cores=use_tensor_cores
-            )
+        if not _grouped_size_compiled_for_decode_kernels(
+            self.model_config.num_attention_heads // self.tp_size,
+            self.model_config.get_num_kv_heads(self.tp_size),
+        ):
+            use_tensor_cores = True
         else:
-            self.flashinfer_prefill_wrapper_ragged = (
-                self.flashinfer_prefill_wrapper_paged
-            ) = None
-            self.flashinfer_decode_wrapper = None
+            use_tensor_cores = False
 
-    @torch.inference_mode()
-    def forward_prefill(self, batch: Batch):
-        input_metadata = InputMetadata.create(
-            self,
-            forward_mode=ForwardMode.PREFILL,
-            tp_size=self.tp_size,
-            req_pool_indices=batch.req_pool_indices,
-            seq_lens=batch.seq_lens,
-            prefix_lens=batch.prefix_lens,
-            position_ids_offsets=batch.position_ids_offsets,
-            out_cache_loc=batch.out_cache_loc,
-            top_logprobs_nums=batch.top_logprobs_nums,
-            return_logprob=batch.return_logprob,
-            flashinfer_prefill_wrapper_ragged=self.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=self.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=self.flashinfer_decode_wrapper,
+        self.flashinfer_workspace_buffers = torch.empty(
+            2, global_config.flashinfer_workspace_size, dtype=torch.uint8, device="cuda"
         )
-        return self.model.forward(
-            batch.input_ids, input_metadata.positions, input_metadata
+        self.flashinfer_prefill_wrapper_ragged = BatchPrefillWithRaggedKVCacheWrapper(
+            self.flashinfer_workspace_buffers[0], "NHD"
+        )
+        self.flashinfer_prefill_wrapper_paged = BatchPrefillWithPagedKVCacheWrapper(
+            self.flashinfer_workspace_buffers[1], "NHD"
         )
+        self.flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
+            self.flashinfer_workspace_buffers[0], "NHD", use_tensor_cores=use_tensor_cores
+        )
+
+    def init_cuda_graphs(self):
+        from sglang.srt.managers.controller.cuda_graph_runner import CudaGraphRunner
+
+        if self.server_args.disable_cuda_graph or self.server_args.disable_flashinfer:
+            self.cuda_graph_runner = None
+            return
+
+        logger.info(f"[gpu_id={self.gpu_id}] Capture cuda graph begin.")
+        batch_size_list = [1, 2, 4] + [i * 8 for i in range(1, 16)]
+        self.cuda_graph_runner = CudaGraphRunner(self, max_batch_size_to_capture=max(batch_size_list))
+        self.cuda_graph_runner.capture(batch_size_list)
 
     @torch.inference_mode()
-    def forward_extend(self, batch: Batch):
+    def forward_decode(self, batch: Batch):
+        if self.cuda_graph_runner and self.cuda_graph_runner.can_run(len(batch.reqs)):
+            return self.cuda_graph_runner.replay(batch)
+
         input_metadata = InputMetadata.create(
             self,
-            forward_mode=ForwardMode.EXTEND,
-            tp_size=self.tp_size,
+            forward_mode=ForwardMode.DECODE,
             req_pool_indices=batch.req_pool_indices,
             seq_lens=batch.seq_lens,
             prefix_lens=batch.prefix_lens,
@@ -459,32 +247,23 @@ class ModelRunner:
             out_cache_loc=batch.out_cache_loc,
             top_logprobs_nums=batch.top_logprobs_nums,
             return_logprob=batch.return_logprob,
-            flashinfer_prefill_wrapper_ragged=self.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=self.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=self.flashinfer_decode_wrapper,
         )
         return self.model.forward(
             batch.input_ids, input_metadata.positions, input_metadata
         )
 
     @torch.inference_mode()
-    def forward_decode(self, batch: Batch):
+    def forward_extend(self, batch: Batch):
         input_metadata = InputMetadata.create(
             self,
-            forward_mode=ForwardMode.DECODE,
-            tp_size=self.tp_size,
+            forward_mode=ForwardMode.EXTEND,
             req_pool_indices=batch.req_pool_indices,
             seq_lens=batch.seq_lens,
             prefix_lens=batch.prefix_lens,
             position_ids_offsets=batch.position_ids_offsets,
             out_cache_loc=batch.out_cache_loc,
-            out_cache_cont_start=batch.out_cache_cont_start,
-            out_cache_cont_end=batch.out_cache_cont_end,
             top_logprobs_nums=batch.top_logprobs_nums,
             return_logprob=batch.return_logprob,
-            flashinfer_prefill_wrapper_ragged=self.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=self.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=self.flashinfer_decode_wrapper,
         )
         return self.model.forward(
             batch.input_ids, input_metadata.positions, input_metadata
@@ -495,17 +274,13 @@ class ModelRunner:
         input_metadata = InputMetadata.create(
             self,
             forward_mode=ForwardMode.EXTEND,
-            tp_size=self.tp_size,
             req_pool_indices=batch.req_pool_indices,
             seq_lens=batch.seq_lens,
             prefix_lens=batch.prefix_lens,
             position_ids_offsets=batch.position_ids_offsets,
             out_cache_loc=batch.out_cache_loc,
-            top_logprobs_nums=batch.top_logprobs_nums,
             return_logprob=batch.return_logprob,
-            flashinfer_prefill_wrapper_ragged=self.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=self.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=self.flashinfer_decode_wrapper,
+            top_logprobs_nums=batch.top_logprobs_nums,
         )
         return self.model.forward(
             batch.input_ids,
@@ -523,8 +298,6 @@ class ModelRunner:
             return self.forward_decode(batch)
         elif forward_mode == ForwardMode.EXTEND:
             return self.forward_extend(batch)
-        elif forward_mode == ForwardMode.PREFILL:
-            return self.forward_prefill(batch)
         else:
             raise ValueError(f"Invaid forward mode: {forward_mode}")
 

sglang/srt/managers/controller/tp_worker.py

@@ -98,7 +98,7 @@ class ModelTpServer:
             )
         self.max_total_num_tokens = self.model_runner.max_total_num_tokens
         self.max_prefill_tokens = (
-            4096
+            8192
             if server_args.max_prefill_tokens is None
             else server_args.max_prefill_tokens
         )
@@ -314,11 +314,9 @@ class ModelTpServer:
         self.forward_queue.append(req)
 
     def get_new_fill_batch(self) -> Optional[Batch]:
-        if (
-            self.running_batch is not None
-            and len(self.running_batch.reqs) > self.max_running_requests
-        ):
-            return None
+        running_bs = len(self.running_batch.reqs) if self.running_batch is not None else 0
+        if running_bs >= self.max_running_requests:
+            return
 
         # Compute matched prefix length
         for req in self.forward_queue:
@@ -394,6 +392,10 @@ class ModelTpServer:
                     new_batch_input_tokens += req.extend_input_len
             else:
                 break
+
+            if running_bs + len(can_run_list) >= self.max_running_requests:
+                break
+
         if len(can_run_list) == 0:
             return None
 

sglang/srt/memory_pool.py

@@ -38,15 +38,24 @@ class ReqToTokenPool:
 
 class TokenToKVPool:
     def __init__(self, size, dtype, head_num, head_dim, layer_num):
-        self.mem_state = torch.zeros((size,), dtype=torch.int16, device="cuda")
+        self.size = size
+        # mem_state is the reference counter.
+        # We also add one slot. This slot is used for writing dummy output from padded tokens.
+        self.mem_state = torch.zeros((self.size + 1,), dtype=torch.int16, device="cuda")
         self.total_ref_ct = 0
 
         # [size, key/value, head_num, head_dim] for each layer
         self.kv_data = [
-            torch.empty((size, 2, head_num, head_dim), dtype=dtype, device="cuda")
+            torch.empty((size + 1, 2, head_num, head_dim), dtype=dtype, device="cuda")
             for _ in range(layer_num)
         ]
 
+        # Prefetch buffer
+        self.prefetch_buffer = torch.empty(0, device="cuda", dtype=torch.int32)
+        self.prefetch_chunk_size = 512
+
+        self.clear()
+
     def get_key_buffer(self, layer_id):
         return self.kv_data[layer_id][:, 0]
 
@@ -54,14 +63,29 @@ class TokenToKVPool:
         return self.kv_data[layer_id][:, 1]
 
     def alloc(self, need_size):
-        select_index = torch.nonzero(self.mem_state == 0).squeeze(1)[:need_size]
-        if select_index.shape[0] < need_size:
+        buffer_len = len(self.prefetch_buffer)
+        if need_size <= buffer_len:
+            select_index = self.prefetch_buffer[:need_size]
+            self.prefetch_buffer = self.prefetch_buffer[need_size:]
+            return select_index
+
+        addition_size = need_size - buffer_len
+        alloc_size = max(addition_size, self.prefetch_chunk_size)
+        select_index = torch.nonzero(self.mem_state == 0).squeeze(1)[:alloc_size].to(torch.int32)
+
+        if select_index.shape[0] < addition_size:
             return None
 
         self.add_refs(select_index)
-        return select_index.to(torch.int32)
+
+        self.prefetch_buffer = torch.cat((self.prefetch_buffer, select_index))
+        ret_index = self.prefetch_buffer[:need_size]
+        self.prefetch_buffer = self.prefetch_buffer[need_size:]
+
+        return ret_index
 
     def alloc_contiguous(self, need_size):
+        # NOTE: This function is deprecated.
         empty_index = torch.nonzero(self.mem_state == 0).squeeze(1)[:need_size]
         if empty_index.shape[0] < need_size:
             return None
@@ -84,7 +108,7 @@ class TokenToKVPool:
         return len(torch.nonzero(self.mem_state).squeeze(1))
 
     def available_size(self):
-        return torch.sum(self.mem_state == 0).item()
+        return torch.sum(self.mem_state == 0).item() + len(self.prefetch_buffer)
 
     def add_refs(self, token_index: torch.Tensor):
         self.total_ref_ct += len(token_index)
@@ -101,3 +125,6 @@ class TokenToKVPool:
     def clear(self):
         self.mem_state.fill_(0)
         self.total_ref_ct = 0
+
+        # We also add one slot. This slot is used for writing dummy output from padded tokens.
+        self.add_refs(torch.tensor([0], dtype=torch.int32))

sglang/srt/server.py

@@ -146,6 +146,7 @@ def launch_server(server_args: ServerArgs, pipe_finish_writer, model_overide_arg
 
     # Set global environments
     os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+    os.environ["NCCL_CUMEM_ENABLE"] = "0"
     if server_args.show_time_cost:
         enable_show_time_cost()
     if server_args.disable_disk_cache:

sglang/srt/server_args.py

@@ -29,7 +29,7 @@ class ServerArgs:
     max_prefill_tokens: Optional[int] = None
     max_running_requests: Optional[int] = None
     schedule_heuristic: str = "lpm"
-    schedule_conservativeness: float = 1.0
+    schedule_conservativeness: float = 0.8
 
     # Other runtime options
     tp_size: int = 1
@@ -53,6 +53,7 @@ class ServerArgs:
     disable_flashinfer: bool = False
     disable_radix_cache: bool = False
     disable_regex_jump_forward: bool = False
+    disable_cuda_graph: bool = False
     disable_disk_cache: bool = False
     attention_reduce_in_fp32: bool = False
     enable_p2p_check: bool = False
@@ -67,13 +68,13 @@ class ServerArgs:
             self.tokenizer_path = self.model_path
         if self.mem_fraction_static is None:
             if self.tp_size >= 8:
-                self.mem_fraction_static = 0.80
+                self.mem_fraction_static = 0.78
             elif self.tp_size >= 4:
-                self.mem_fraction_static = 0.82
+                self.mem_fraction_static = 0.80
             elif self.tp_size >= 2:
                 self.mem_fraction_static = 0.85
             else:
-                self.mem_fraction_static = 0.90
+                self.mem_fraction_static = 0.88
         if isinstance(self.additional_ports, int):
             self.additional_ports = [self.additional_ports]
         elif self.additional_ports is None:
@@ -294,6 +295,11 @@ class ServerArgs:
             action="store_true",
             help="Disable regex jump-forward",
         )
+        parser.add_argument(
+            "--disable-cuda-graph",
+            action="store_true",
+            help="Disable cuda graph.",
+        )
         parser.add_argument(
             "--disable-disk-cache",
             action="store_true",

{sglang-0.1.19.dist-info → sglang-0.1.20.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.1.19
+Version: 0.1.20
 Summary: A structured generation langauge for LLMs.
 License: Apache License
                                    Version 2.0, January 2004