Allow disabling flashinfer sampling kernel (#778)

python/sglang/srt/layers/radix_attention.py

@@ -7,8 +7,11 @@ from torch import nn
 from sglang.global_config import global_config
 from sglang.srt.layers.extend_attention import extend_attention_fwd
 from sglang.srt.layers.token_attention import token_attention_fwd
-from sglang.srt.managers.controller.model_runner import ForwardMode, InputMetadata
-from sglang.srt.server import global_server_args_dict
+from sglang.srt.managers.controller.model_runner import (
+    ForwardMode,
+    InputMetadata,
+    global_server_args_dict,
+)
 
 
 class RadixAttention(nn.Module):

python/sglang/srt/layers/token_attention.py

@@ -5,7 +5,7 @@ import torch
 import triton
 import triton.language as tl
 
-from sglang.srt.server import global_server_args_dict
+from sglang.srt.managers.controller.infer_batch import global_server_args_dict
 
 if global_server_args_dict.get("attention_reduce_in_fp32", False):
     REDUCE_TRITON_TYPE = tl.float32

python/sglang/srt/managers/controller/infer_batch.py

@@ -17,6 +17,13 @@ from sglang.srt.memory_pool import ReqToTokenPool, TokenToKVPool
 
 INIT_INCREMENTAL_DETOKENIZATION_OFFSET = 5
 
+# Put some global args for easy access
+global_server_args_dict = {
+    "disable_flashinfer": False,
+    "disable_flashinfer_sampling": False,
+    "attention_reduce_in_fp32": False,
+}
+
 
 class ForwardMode(IntEnum):
     # Prefill a new sequence. This is deprecated now. "EXTEND" covers this case.
@@ -687,7 +694,7 @@ class Batch:
         # TODO(lmzheng): apply penalty
         probs = torch.softmax(logits, dim=-1)
 
-        if True:
+        if not global_server_args_dict["disable_flashinfer_sampling"]:
             max_top_k_round, batch_size = 32, probs.shape[0]
             uniform_samples = torch.rand(
                 (max_top_k_round, batch_size), device=probs.device

python/sglang/srt/managers/controller/model_runner.py

@@ -25,7 +25,12 @@ from vllm.distributed import (
 from vllm.model_executor.models import ModelRegistry
 
 from sglang.global_config import global_config
-from sglang.srt.managers.controller.infer_batch import Batch, ForwardMode, InputMetadata
+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 (
@@ -60,7 +65,13 @@ class ModelRunner:
         self.nccl_port = nccl_port
         self.server_args = server_args
         self.is_multimodal_model = is_multimodal_model(self.model_config)
-        monkey_patch_vllm_dummy_weight_loader()
+        global_server_args_dict.update(
+            {
+                "disable_flashinfer": server_args.disable_flashinfer,
+                "disable_flashinfer_sampling": server_args.disable_flashinfer_sampling,
+                "attention_reduce_in_fp32": server_args.attention_reduce_in_fp32,
+            }
+        )
 
         # Init torch distributed
         torch.cuda.set_device(self.gpu_id)
@@ -108,6 +119,7 @@ class ModelRunner:
             f"avail mem={get_available_gpu_memory(self.gpu_id):.2f} GB"
         )
 
+        monkey_patch_vllm_dummy_weight_loader()
         device_config = DeviceConfig()
         load_config = LoadConfig(load_format=self.server_args.load_format)
         vllm_model_config = VllmModelConfig(

python/sglang/srt/server.py

@@ -65,9 +65,6 @@ asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 app = FastAPI()
 tokenizer_manager = None
 
-# Put some args for easily access
-global_server_args_dict = {}
-
 
 @app.get("/health")
 async def health() -> Response:
@@ -150,14 +147,6 @@ def available_models():
     return ModelList(data=model_cards)
 
 
-def _set_global_server_args(server_args: ServerArgs):
-    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,
-    }
-
-
 def _set_torch_compile_config():
     # The following configurations are for torch compile optimizations
     import torch._dynamo.config
@@ -213,8 +202,6 @@ def launch_server(
     if server_args.enable_torch_compile:
         _set_torch_compile_config()
 
-    _set_global_server_args(server_args)
-
     # Allocate ports
     server_args.port, server_args.additional_ports = allocate_init_ports(
         server_args.port,

python/sglang/srt/server_args.py

@@ -52,13 +52,14 @@ class ServerArgs:
 
     # Optimization/debug options
     disable_flashinfer: bool = False
+    disable_flashinfer_sampling: bool = False
     disable_radix_cache: bool = False
     disable_regex_jump_forward: bool = False
     disable_cuda_graph: bool = False
     disable_disk_cache: bool = False
     enable_torch_compile: bool = False
-    attention_reduce_in_fp32: bool = False
     enable_p2p_check: bool = False
+    attention_reduce_in_fp32: bool = False
     efficient_weight_load: bool = False
 
     # Distributed args
@@ -303,7 +304,12 @@ class ServerArgs:
         parser.add_argument(
             "--disable-flashinfer",
             action="store_true",
-            help="Disable flashinfer inference kernels.",
+            help="Disable flashinfer attention kernels.",
+        )
+        parser.add_argument(
+            "--disable-flashinfer-sampling",
+            action="store_true",
+            help="Disable flashinfer sampling kernels.",
         )
         parser.add_argument(
             "--disable-radix-cache",
@@ -331,15 +337,15 @@ class ServerArgs:
             help="Optimize the model with torch.compile, experimental feature.",
         )
         parser.add_argument(
-            "--attention-reduce-in-fp32",
+            "--enable-p2p-check",
             action="store_true",
-            help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16."
-            "This only affects Triton attention kernels",
+            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
         )
         parser.add_argument(
-            "--enable-p2p-check",
+            "--attention-reduce-in-fp32",
             action="store_true",
-            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
+            help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16."
+            "This only affects Triton attention kernels",
         )
         parser.add_argument(
             "--efficient-weight-load",