sglang 0.4.4.post3__tar.gz → 0.4.5__tar.gz

{sglang-0.4.4.post3/sglang.egg-info → sglang-0.4.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sglang
-Version: 0.4.4.post3
+Version: 0.4.5
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -234,18 +234,19 @@ Requires-Dist: pillow; extra == "runtime-common"
 Requires-Dist: prometheus-client>=0.20.0; extra == "runtime-common"
 Requires-Dist: psutil; extra == "runtime-common"
 Requires-Dist: pydantic; extra == "runtime-common"
+Requires-Dist: pynvml; extra == "runtime-common"
 Requires-Dist: python-multipart; extra == "runtime-common"
 Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
 Requires-Dist: soundfile==0.13.1; extra == "runtime-common"
 Requires-Dist: torchao>=0.7.0; extra == "runtime-common"
-Requires-Dist: transformers==4.50.0; extra == "runtime-common"
+Requires-Dist: transformers==4.51.0; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
 Requires-Dist: compressed-tensors; extra == "runtime-common"
 Requires-Dist: xgrammar==0.1.17; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
-Requires-Dist: sgl-kernel==0.0.5.post4; extra == "srt"
+Requires-Dist: sgl-kernel==0.0.8; extra == "srt"
 Requires-Dist: flashinfer_python==0.2.3; extra == "srt"
 Requires-Dist: torch==2.5.1; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"

{sglang-0.4.4.post3 → sglang-0.4.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.4.4.post3"
+version = "0.4.5"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -33,11 +33,12 @@ runtime_common = [
     "prometheus-client>=0.20.0",
     "psutil",
     "pydantic",
+    "pynvml",
     "python-multipart",
     "pyzmq>=25.1.2",
     "soundfile==0.13.1",
     "torchao>=0.7.0",
-    "transformers==4.50.0",
+    "transformers==4.51.0",
     "uvicorn",
     "uvloop",
     "compressed-tensors",
@@ -46,7 +47,7 @@ runtime_common = [
 
 srt = [
     "sglang[runtime_common]",
-    "sgl-kernel==0.0.5.post4",
+    "sgl-kernel==0.0.8",
     "flashinfer_python==0.2.3",
     "torch==2.5.1",
     "cuda-python",

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/bench_serving.py

@@ -44,6 +44,12 @@ ASSISTANT_SUFFIX = "Assistant:"
 global args
 
 
+# don't want to import sglang package here
+def _get_bool_env_var(name: str, default: str = "false") -> bool:
+    value = os.getenv(name, default)
+    return value.lower() in ("true", "1")
+
+
 @dataclass
 class RequestFuncInput:
     prompt: str
@@ -969,6 +975,7 @@ async def benchmark(
     extra_request_body: Dict[str, Any],
     profile: bool,
     pd_seperated: bool = False,
+    flush_cache: bool = False,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -986,13 +993,16 @@ async def benchmark(
             return await request_func(request_func_input=request_func_input, pbar=pbar)
 
     # Warmup
-    print("Starting initial single prompt test run...")
+    print(f"Starting warmup with {args.warmup_requests} sequences...")
+
+    # Use the first request for all warmup iterations
     test_prompt, test_prompt_len, test_output_len = input_requests[0]
     if lora_names != None and len(lora_names) != 0:
         lora_name = lora_names[0]
     else:
         lora_name = None
 
+    # Create the test input once
     test_input = RequestFuncInput(
         model=model_id,
         prompt=test_prompt,
@@ -1002,17 +1012,29 @@ async def benchmark(
         lora_name=lora_name,
         extra_request_body=extra_request_body,
     )
-    test_output = await request_func(request_func_input=test_input)
-    if not test_output.success:
+
+    # Run warmup requests
+    warmup_tasks = []
+    for _ in range(args.warmup_requests):
+        warmup_tasks.append(
+            asyncio.create_task(request_func(request_func_input=test_input))
+        )
+
+    warmup_outputs = await asyncio.gather(*warmup_tasks)
+
+    # Check if at least one warmup request succeeded
+    if not any(output.success for output in warmup_outputs):
         raise ValueError(
-            "Initial test run failed - Please make sure benchmark arguments "
-            f"are correctly specified. Error: {test_output.error}"
+            "Warmup failed - Please make sure benchmark arguments "
+            f"are correctly specified. Error: {warmup_outputs[0].error}"
         )
     else:
-        print("Initial test run completed. Starting main benchmark run...")
+        print(
+            f"Warmup completed with {args.warmup_requests} sequences. Starting main benchmark run..."
+        )
 
     # Flush cache
-    if "sglang" in backend:
+    if ("sglang" in backend and _get_bool_env_var("SGLANG_IS_IN_CI")) or flush_cache:
         requests.post(base_url + "/flush_cache", headers=get_auth_headers())
 
     time.sleep(1.0)
@@ -1246,6 +1268,10 @@ def run_benchmark(args_: argparse.Namespace):
     if not hasattr(args, "max_concurrency"):
         args.max_concurrency = None
 
+    # Set default value for warmup_requests if not present
+    if not hasattr(args, "warmup_requests"):
+        args.warmup_requests = 1
+
     print(f"benchmark_args={args}")
 
     # Set global environments
@@ -1347,6 +1373,10 @@ def run_benchmark(args_: argparse.Namespace):
     tokenizer = get_tokenizer(tokenizer_id)
     input_requests = get_dataset(args, tokenizer)
 
+    # compatible with SimpleNamespace
+    if not hasattr(args, "flush_cache"):
+        args.flush_cache = False
+
     return asyncio.run(
         benchmark(
             backend=backend,
@@ -1362,6 +1392,7 @@ def run_benchmark(args_: argparse.Namespace):
             extra_request_body=extra_request_body,
             profile=args.profile,
             pd_seperated=args.pd_seperated,
+            flush_cache=args.flush_cache,
         )
     )
 
@@ -1543,6 +1574,17 @@ if __name__ == "__main__":
         action="store_true",
         help="Benchmark PD disaggregation server",
     )
+    parser.add_argument(
+        "--flush-cache",
+        action="store_true",
+        help="Flush the cache before running the benchmark",
+    )
+    parser.add_argument(
+        "--warmup-requests",
+        type=int,
+        default=1,
+        help="Number of warmup requests to run before the benchmark",
+    )
 
     group = parser.add_argument_group("generated-shared-prefix dataset arguments")
     group.add_argument(

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/lang/chat_template.py

@@ -294,6 +294,30 @@ register_chat_template(
     )
 )
 
+# Reference: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/blob/main/chat_template.json
+register_chat_template(
+    ChatTemplate(
+        name="llama-4",
+        default_system_prompt=None,
+        role_prefix_and_suffix={
+            "system": (
+                "<|header_start|>system<|header_end|>\n\n",
+                "<|eot|>",
+            ),
+            "user": (
+                "<|header_start|>user<|header_end|>\n\n",
+                "<|eot|>",
+            ),
+            "assistant": (
+                "<|header_start|>assistant<|header_end|>\n\n",
+                "<|eot|>",
+            ),
+        },
+        stop_str=("<|eot|>",),
+        image_token="<|image|>",
+    )
+)
+
 # Reference: https://modelscope.cn/models/01ai/Yi-1.5-34B-Chat/file/view/master?fileName=tokenizer_config.json&status=1
 register_chat_template(
     ChatTemplate(

sglang-0.4.5/sglang/srt/_custom_ops.py

@@ -0,0 +1,117 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/_custom_ops.py
+import logging
+import os
+from typing import List, Tuple
+
+import torch
+import torch.library
+
+from sglang.srt.utils import get_bool_env_var, is_hip, is_hpu
+
+logger = logging.getLogger(__name__)
+use_vllm_custom_allreduce = get_bool_env_var(
+    "USE_VLLM_CUSTOM_ALLREDUCE", default="false"
+)
+
+if not is_hpu():
+    # ROCm does not use vllm custom allreduce
+    if use_vllm_custom_allreduce and not is_hip():
+        try:
+            import vllm._C
+        except ImportError as e:
+            logger.warning("Failed to import from vllm._C with %r", e)
+    else:
+        try:
+            import sgl_kernel
+        except ImportError as e:
+            logger.warning("Failed to import from custom_ar with %r", e)
+
+
+if not is_hip():
+    if use_vllm_custom_allreduce:
+        custom_op = torch.ops._C_custom_ar
+    else:
+        custom_op = sgl_kernel.allreduce
+
+    # custom allreduce
+    def init_custom_ar(
+        ipc_tensors: List[torch.Tensor],
+        rank_data: torch.Tensor,
+        rank: int,
+        full_nvlink: bool,
+    ) -> int:
+        return custom_op.init_custom_ar(ipc_tensors, rank_data, rank, full_nvlink)
+
+    def all_reduce(
+        fa: int,
+        inp: torch.Tensor,
+        out: torch.Tensor,
+        reg_buffer: int,
+        reg_buffer_sz_bytes: int,
+    ) -> None:
+        custom_op.all_reduce(fa, inp, out, reg_buffer, reg_buffer_sz_bytes)
+
+    def dispose(fa: int) -> None:
+        custom_op.dispose(fa)
+
+    def meta_size() -> int:
+        return custom_op.meta_size()
+
+    def register_buffer(fa: int, ipc_tensors: List[int]) -> None:
+        return custom_op.register_buffer(fa, ipc_tensors)
+
+    def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
+        return custom_op.get_graph_buffer_ipc_meta(fa)
+
+    def register_graph_buffers(
+        fa: int, handles: List[List[int]], offsets: List[List[int]]
+    ) -> None:
+        custom_op.register_graph_buffers(fa, handles, offsets)
+
+else:
+    # ROCM custom allreduce
+
+    def init_custom_ar(
+        meta: torch.Tensor,
+        rank_data: torch.Tensor,
+        handles: List[str],
+        offsets: List[int],
+        rank: int,
+        full_nvlink: bool,
+    ) -> int:
+        return sgl_kernel.allreduce.init_custom_ar(
+            meta, rank_data, handles, offsets, rank, full_nvlink
+        )
+
+    def all_reduce_reg(fa: int, inp: torch.Tensor, out: torch.Tensor) -> None:
+        sgl_kernel.allreduce.all_reduce_reg(fa, inp, out)
+
+    def all_reduce_unreg(
+        fa: int, inp: torch.Tensor, reg_buffer: torch.Tensor, out: torch.Tensor
+    ) -> None:
+        sgl_kernel.allreduce.all_reduce_unreg(fa, inp, reg_buffer, out)
+
+    def dispose(fa: int) -> None:
+        sgl_kernel.allreduce.dispose(fa)
+
+    def meta_size() -> int:
+        return sgl_kernel.allreduce.meta_size()
+
+    def register_buffer(
+        fa: int, t: torch.Tensor, handles: List[str], offsets: List[int]
+    ) -> None:
+        return sgl_kernel.allreduce.register_buffer(fa, t, handles, offsets)
+
+    def get_graph_buffer_ipc_meta(fa: int) -> Tuple[torch.Tensor, List[int]]:
+        return sgl_kernel.allreduce.get_graph_buffer_ipc_meta(fa)
+
+    def register_graph_buffers(
+        fa: int, handles: List[str], offsets: List[List[int]]
+    ) -> None:
+        sgl_kernel.allreduce.register_graph_buffers(fa, handles, offsets)
+
+    def allocate_meta_buffer(size: int) -> torch.Tensor:
+        return sgl_kernel.allreduce.allocate_meta_buffer(size)
+
+    def get_meta_buffer_ipc_handle(inp: torch.Tensor) -> torch.Tensor:
+        return sgl_kernel.allreduce.get_meta_buffer_ipc_handle(inp)

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/configs/model_config.py

@@ -65,6 +65,9 @@ class ModelConfig:
             **kwargs,
         )
         self.hf_text_config = get_hf_text_config(self.hf_config)
+        self.attention_chunk_size = getattr(
+            self.hf_text_config, "attention_chunk_size", None
+        )
 
         # Check model type
         self.is_generation = is_generation_model(
@@ -258,6 +261,7 @@ class ModelConfig:
             "experts_int8",
             "w8a8_int8",
             "w8a8_fp8",
+            "moe_wna16",
         ]
         compatible_quantization_methods = {
             "w8a8_int8": ["compressed-tensors", "compressed_tensors"],
@@ -466,6 +470,7 @@ multimodal_model_archs = [
     "Gemma3ForConditionalGeneration",
     "Grok1VForCausalLM",
     "Grok1AForCausalLM",
+    # TODO: add multimodal support for "Llama4ForConditionalGeneration",
     "LlavaLlamaForCausalLM",
     "LlavaMistralForCausalLM",
     "LlavaQwenForCausalLM",

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/constrained/base_grammar_backend.py

@@ -169,7 +169,9 @@ class BaseGrammarBackend(ABC):
             self.cache.clear()
 
 
-def create_grammar_backend(server_args: ServerArgs, tokenizer, vocab_size):
+def create_grammar_backend(
+    server_args: ServerArgs, tokenizer, vocab_size: int
+) -> Optional[BaseGrammarBackend]:
     if server_args.grammar_backend == "outlines":
         from sglang.srt.constrained.outlines_backend import OutlinesGrammarBackend
 
@@ -188,6 +190,8 @@ def create_grammar_backend(server_args: ServerArgs, tokenizer, vocab_size):
             tokenizer=tokenizer,
             whitespace_pattern=server_args.constrained_json_whitespace_pattern,
         )
+    elif server_args.grammar_backend == "none":
+        return None
     else:
         raise ValueError(f"Invalid grammar backend: {server_args.grammar_backend}")
 

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/conversation.py

@@ -33,6 +33,7 @@ class SeparatorStyle(IntEnum):
     ADD_NEW_LINE_SINGLE = auto()
     LLAMA2 = auto()
     LLAMA3 = auto()
+    LLAMA4 = auto()
     CHATGLM = auto()
     CHATML = auto()
     CHATINTERN = auto()
@@ -156,19 +157,30 @@ class Conversation:
                 else:
                     ret += role + ":"
             return ret
+        elif self.sep_style == SeparatorStyle.LLAMA4:
+            # begin_of_text is added by default
+            if self.system_message:
+                ret = system_prompt
+            else:
+                ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += f"<|header_start|>{role}<|header_end|>\n\n"
+                    ret += f"{message.strip()}<|eot|>"
+                else:
+                    ret += f"<|header_start|>{role}<|header_end|>\n\n"
+            return ret
         elif self.sep_style == SeparatorStyle.LLAMA3:
-            ret = "<|begin_of_text|>"
             if self.system_message:
-                ret += system_prompt
+                ret = system_prompt
             else:
-                ret += ""
+                ret = ""
             for i, (role, message) in enumerate(self.messages):
                 if message:
                     ret += f"<|start_header_id|>{role}<|end_header_id|>\n\n"
                     ret += f"{message.strip()}<|eot_id|>"
                 else:
                     ret += f"<|start_header_id|>{role}<|end_header_id|>\n\n"
-            # print(ret)
             return ret
         elif self.sep_style == SeparatorStyle.LLAMA2:
             seps = [self.sep, self.sep2]
@@ -561,6 +573,19 @@ register_conv_template(
     )
 )
 
+# reference: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/blob/main/chat_template.json
+register_conv_template(
+    Conversation(
+        name="llama-4",
+        system_template="<|header_start|>system<|header_end|>\n\n{system_message}<|eot|>",
+        roles=("user", "assistant"),
+        sep_style=SeparatorStyle.LLAMA4,
+        sep="",
+        stop_str=["<|end_of_text|>", "<|eot|>", "<|eom|>"],
+        image_token="<|image|>",
+    )
+)
+
 register_conv_template(
     Conversation(
         name="chatml",

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/custom_op.py

@@ -50,6 +50,7 @@ if _is_cuda:
     def scaled_fp8_quant(
         input: torch.Tensor,
         scale: Optional[torch.Tensor] = None,
+        num_token_padding: Optional[int] = None,
         use_per_token_if_dynamic: bool = False,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
@@ -59,6 +60,8 @@ if _is_cuda:
             input (torch.Tensor): Input tensor to be quantized
             scale (Optional[torch.Tensor]): Pre-computed scaling factor for static quantization.
                 If None, scales will be computed dynamically.
+            num_token_padding (Optional[int]): If specified, pad the first dimension
+                of the output to at least this value.
             use_per_token_if_dynamic (bool): When using dynamic scaling (scale=None),
                 determines the quantization granularity:
                 - True: compute scale per token
@@ -75,6 +78,8 @@ if _is_cuda:
         assert input.ndim == 2, f"Expected 2D input tensor, got {input.ndim}D"
         shape = input.shape
         out_dtype = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
+        if num_token_padding:
+            shape = (max(num_token_padding, input.shape[0]), shape[1])
         output = torch.empty(shape, device=input.device, dtype=out_dtype)
 
         if scale is None:

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/distributed/device_communicators/custom_all_reduce.py

@@ -18,7 +18,7 @@ from sglang.srt.distributed.device_communicators.custom_all_reduce_utils import
     gpu_p2p_access_check,
 )
 from sglang.srt.distributed.parallel_state import in_the_same_node_as
-from sglang.srt.utils import cuda_device_count_stateless, is_cuda, is_hip
+from sglang.srt.utils import is_cuda, is_hip
 
 logger = logging.getLogger(__name__)
 
@@ -217,7 +217,7 @@ class CustomAllreduce:
         if cuda_visible_devices:
             device_ids = list(map(int, cuda_visible_devices.split(",")))
         else:
-            device_ids = list(range(cuda_device_count_stateless()))
+            device_ids = list(range(torch.cuda.device_count()))
 
         physical_device_id = device_ids[device.index]
         tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
@@ -257,7 +257,7 @@ class CustomAllreduce:
         self.world_size = world_size
         self.full_nvlink = full_nvlink
 
-        if ops.use_vllm_custom_allreduce and not _is_hip:
+        if not _is_hip:
             # Buffers memory are owned by this Python class and passed to C++.
             # Meta data composes of two parts: meta data for synchronization and a
             # temporary buffer for storing intermediate allreduce results.
@@ -280,56 +280,24 @@ class CustomAllreduce:
             )
             ops.register_buffer(self._ptr, self.buffer_ptrs)
         else:
-            if _is_hip:
-                # meta data buffers need to be "uncached" for signal on MI200
-                self.meta = ops.allocate_meta_buffer(ops.meta_size() + max_size)
-                self.buffer = torch.empty(
-                    max_size, dtype=torch.uint8, device=self.device
-                )
-                handle = ops.get_meta_buffer_ipc_handle(self.meta)
-                shard_data = (
-                    bytes(handle),  # ipc handle to base ptr
-                    0,  # offset of base ptr
-                )
-                handles, offsets = self._gather_ipc_meta(shard_data)
-                self.rank_data = torch.empty(
-                    8 * 1024 * 1024, dtype=torch.uint8, device=self.device
-                )
-                self._ptr = ops.init_custom_ar(
-                    self.meta, self.rank_data, handles, offsets, rank, self.full_nvlink
-                )
-                self.register_buffer(self.buffer)
-                self.MSCCL = os.getenv("RCCL_MSCCL_ENABLE", "1") == "1"
-            else:
-                # From TensorRT-LLM getMaxRequiredWorkspaceSize
-                self.max_required_workspace_size = [16 * 1024 * 1024, 8 * 1024 * 1024]
-
-                # sizeof(uint32_t) * (MAX_ALL_REDUCE_BLOCKS + 2) * MAX_RANKS_PER_NODE;
-                self.barrier_max_size = 8 * (36 + 2) * 8
-
-                self.buffer_ptrs = self.create_shared_buffer(max_size, group=group)
-                self.tmp_result_buffer_ptrs = self.create_shared_buffer(
-                    max_size, group=group
-                )
-                self.rank_data_base = torch.empty(
-                    8 * 1024 * 1024, dtype=torch.uint8, device=self.device
-                )
-                self.barrier_in_ptrs = self.create_shared_buffer(
-                    self.barrier_max_size, group=group
-                )
-                self.barrier_out_ptrs = self.create_shared_buffer(
-                    self.barrier_max_size, group=group
-                )
+            # meta data buffers need to be "uncached" for signal on MI200
+            self.meta = ops.allocate_meta_buffer(ops.meta_size() + max_size)
+            self.buffer = torch.empty(max_size, dtype=torch.uint8, device=self.device)
+            handle = ops.get_meta_buffer_ipc_handle(self.meta)
+            shard_data = (
+                bytes(handle),  # ipc handle to base ptr
+                0,  # offset of base ptr
+            )
+            handles, offsets = self._gather_ipc_meta(shard_data)
+            self.rank_data = torch.empty(
+                8 * 1024 * 1024, dtype=torch.uint8, device=self.device
+            )
+            self._ptr = ops.init_custom_ar(
+                self.meta, self.rank_data, handles, offsets, rank, self.full_nvlink
+            )
+            self.register_buffer(self.buffer)
+            self.MSCCL = os.getenv("RCCL_MSCCL_ENABLE", "1") == "1"
 
-                self._ptr = ops.init_custom_ar(
-                    rank,
-                    world_size,
-                    self.rank_data_base,
-                    self.buffer_ptrs,
-                    self.tmp_result_buffer_ptrs,
-                    self.barrier_in_ptrs,
-                    self.barrier_out_ptrs,
-                )
         self.disabled = False
 
     @staticmethod
@@ -455,7 +423,7 @@ class CustomAllreduce:
             return False
         # for 4 or more non NVLink-capable GPUs, custom allreduce provides
         # little performance improvement over NCCL.
-        if ops.use_vllm_custom_allreduce and not _is_hip:
+        if not _is_hip:
             if self.world_size == 2 or self.full_nvlink:
                 return inp_size < self.max_size
             return False
@@ -471,18 +439,6 @@ class CustomAllreduce:
                     return inp_size < self.max_size
             return False
 
-        if self.world_size == 2:
-            return (
-                inp_size < self.max_size
-                and inp_size < self.max_required_workspace_size[0]
-            )
-
-        if self.full_nvlink:
-            return (
-                inp_size < self.max_size
-                and inp_size < self.max_required_workspace_size[1]
-            )
-
         return False
 
     # all reduce, assuming inp tensor is IPC registered with register_buffer,
@@ -515,15 +471,12 @@ class CustomAllreduce:
         """
         if out is None:
             out = torch.empty_like(inp)
-        if ops.use_vllm_custom_allreduce:
-            if registered:
-                ops.all_reduce(self._ptr, inp, out, 0, 0)
-            else:
-                ops.all_reduce(
-                    self._ptr, inp, out, self.buffer_ptrs[self.rank], self.max_size
-                )
+        if registered:
+            ops.all_reduce(self._ptr, inp, out, 0, 0)
         else:
-            ops.all_reduce(self._ptr, inp, out)
+            ops.all_reduce(
+                self._ptr, inp, out, self.buffer_ptrs[self.rank], self.max_size
+            )
         return out
 
     def custom_all_reduce(self, input: torch.Tensor) -> Optional[torch.Tensor]:
@@ -554,14 +507,9 @@ class CustomAllreduce:
     def close(self):
         if not self.disabled and self._ptr:
             ops.dispose(self._ptr)
-            if ops.use_vllm_custom_allreduce:
+            if _is_cuda:
                 self.free_shared_buffer(self.meta_ptrs)
                 self.free_shared_buffer(self.buffer_ptrs)
-            elif _is_cuda:
-                self.free_shared_buffer(self.buffer_ptrs)
-                self.free_shared_buffer(self.tmp_result_buffer_ptrs)
-                self.free_shared_buffer(self.barrier_in_ptrs)
-                self.free_shared_buffer(self.barrier_out_ptrs)
             self._ptr = 0
 
     def __del__(self):

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/distributed/device_communicators/custom_all_reduce_utils.py

@@ -11,11 +11,11 @@ import tempfile
 from itertools import product
 from typing import Dict, List, Optional, Sequence
 
+import torch
 import torch.distributed as dist
 import torch.multiprocessing as mp
 
 from sglang.srt.distributed.device_communicators.cuda_wrapper import CudaRTLibrary
-from sglang.srt.utils import cuda_device_count_stateless
 
 logger = logging.getLogger(__name__)
 
@@ -218,7 +218,7 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
 
     is_distributed = dist.is_initialized()
 
-    num_dev = cuda_device_count_stateless()
+    num_dev = torch.cuda.device_count()
     cuda_visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES", None)
     if cuda_visible_devices is None:
         cuda_visible_devices = ",".join(str(i) for i in range(num_dev))

{sglang-0.4.4.post3 → sglang-0.4.5}/sglang/srt/entrypoints/engine.py

@@ -151,10 +151,6 @@ class Engine:
         The arguments of this function is the same as `sglang/srt/managers/io_struct.py::GenerateReqInput`.
         Please refer to `GenerateReqInput` for the documentation.
         """
-        modalities_list = []
-        if image_data is not None:
-            modalities_list.append("image")
-
         obj = GenerateReqInput(
             text=prompt,
             input_ids=input_ids,
@@ -165,7 +161,6 @@ class Engine:
             top_logprobs_num=top_logprobs_num,
             token_ids_logprob=token_ids_logprob,
             lora_path=lora_path,
-            modalities=modalities_list,
             custom_logit_processor=custom_logit_processor,
             return_hidden_states=return_hidden_states,
             stream=stream,