[Build] Add srt_empty extra group for device-agnostic install (#31300)

Co-authored-by: yixiaodapeng <yixiaodapeng@example.com>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>

python/pyproject_other.toml

@@ -17,14 +17,16 @@ classifiers = [
 dependencies = ["aiohttp", "IPython", "numpy", "requests", "setproctitle", "tqdm"]
 
 [project.optional-dependencies]
-runtime_common = [
+# runtime_base: torch-free subset of runtime_common.
+# Used by srt_empty for device-agnostic install on non-NVIDIA platforms.
+# NOTE: Do NOT add packages that depend on torch/triton here — they belong in runtime_common.
+runtime_base = [
   "aiohttp",
   "anthropic>=0.20.0",
   "apache-tvm-ffi",
   "av",
   "blobfile==3.0.0",
   "build",
-  "compressed-tensors",
   "datasets",
   "easydict",
   "einops",
@@ -42,7 +44,6 @@ runtime_common = [
   "openai==2.6.1",
   "openai-harmony==0.0.4",
   "orjson",
-  "outlines==0.1.11",
   "packaging",
   "partial_json_parser",
   "pillow",
@@ -60,16 +61,29 @@ runtime_common = [
   "smg-grpc-servicer>=0.5.0",
   "soundfile==0.13.1",
   "tiktoken",
-  "timm==1.0.16",
-  "torchao==0.9.0",
   "tqdm",
   "transformers==5.12.1",
   "uvicorn",
   "uvloop",
   "xxhash",
+]
+
+# runtime_common: backward-compatible — same install result as before the split.
+runtime_common = [
+  "sglang[runtime_base]",
+  "compressed-tensors",
+  "outlines==0.1.11",
+  "timm==1.0.16",
+  "torchao==0.9.0",
   "xgrammar==0.2.1",
 ]
 
+# srt_empty: device-agnostic install — pure Python packages only, no torch dependency chain.
+# Enables OOT plugins (e.g. sglang-plugin-FL) to install sglang without conflicting with
+# vendor-specific PyTorch builds (torch_npu, torch_musa, etc.).
+# Usage: cp pyproject_other.toml pyproject.toml && pip install -e ".[srt_empty]"
+srt_empty = ["sglang[runtime_base]"]
+
 diffusion_common = [
   "addict",
   "cloudpickle",

python/sglang/srt/managers/scheduler.py

@@ -25,7 +25,7 @@ from collections import deque
 from contextlib import contextmanager, nullcontext
 from functools import partial
 from http import HTTPStatus
-from typing import Any, Deque, Dict, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Deque, Dict, List, Optional, Tuple, Union
 
 from sglang.srt.runtime_context import (
     get_device,
@@ -50,12 +50,17 @@ import psutil  # isort: skip
 import setproctitle
 import torch
 import torch.distributed
-from torch.cuda import Stream as CudaStream
 from torch.distributed import barrier
 
-from sglang.kernels.ops.mamba.triton_ops import (
-    initialize_mamba_selective_state_update_backend,
-)
+if TYPE_CHECKING:
+    from torch.cuda import Stream as CudaStream
+
+try:
+    from sglang.kernels.ops.mamba.triton_ops import (
+        initialize_mamba_selective_state_update_backend,
+    )
+except ImportError:
+    initialize_mamba_selective_state_update_backend = None
 from sglang.srt.configs.model_config import (
     ModelConfig,
     ModelImpl,
@@ -851,7 +856,8 @@ class Scheduler(
                 )
 
     def init_mamba_backend(self) -> None:
-        initialize_mamba_selective_state_update_backend(self.server_args)
+        if initialize_mamba_selective_state_update_backend is not None:
+            initialize_mamba_selective_state_update_backend(self.server_args)
 
     def init_moe_gemm_config(self):
         # For the MM models, check the text_config for MoE settings

python/sglang/srt/server_args.py

@@ -31,7 +31,6 @@ import uuid
 from functools import cached_property
 from typing import Any, Callable, Dict, List, Literal, Optional, Union
 
-from sglang.kernels.ops.attention.fla.chunk_delta_h import CHUNK_SIZE as FLA_CHUNK_SIZE
 from sglang.kernels.ops.kv_canary.consts import RealKvHashMode
 from sglang.srt.arg_groups.arg_utils import NS, A, Arg, add_cli_args_from_dataclass
 from sglang.srt.arg_groups.argparse_actions import (
@@ -8675,6 +8674,14 @@ class ServerArgs:
         # It is used to determine the caching point in a sequence during prefill.
         if not hasattr(self, "_mamba_cache_chunk_size"):
 
+            try:
+                from sglang.kernels.ops.attention.fla.chunk_delta_h import (
+                    CHUNK_SIZE as FLA_CHUNK_SIZE,
+                )
+            except ImportError:
+                # Must match sglang.kernels.ops.attention.fla.chunk_delta_h.CHUNK_SIZE
+                FLA_CHUNK_SIZE = 64
+
             hf_config = self.get_model_config().hf_config
             chunk_size = getattr(hf_config, "mamba_chunk_size", FLA_CHUNK_SIZE)
             page_size = resolved_view(self).page_size