sglang 0.2.6__py3-none-any.whl → 0.2.7__py3-none-any.whl

sglang/__init__.py

@@ -1,4 +1,5 @@
 # SGL API Components
+
 from sglang.api import (
     Runtime,
     assistant,
@@ -14,48 +15,54 @@ from sglang.api import (
     select,
     set_default_backend,
     system,
+    system_begin,
+    system_end,
     user,
     user_begin,
     user_end,
     video,
 )
 
-# Global Configurations
-from sglang.global_config import global_config
-
-# SGL Backends
-from sglang.lang.backend.anthropic import Anthropic
-from sglang.lang.backend.litellm import LiteLLM
-from sglang.lang.backend.openai import OpenAI
-from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
-from sglang.lang.backend.vertexai import VertexAI
-
-from .version import __version__
-
-# public APIs management
+# SGLang DSL APIs
 __all__ = [
-    "global_config",
-    "Anthropic",
-    "LiteLLM",
-    "OpenAI",
-    "RuntimeEndpoint",
-    "VertexAI",
-    "function",
     "Runtime",
-    "set_default_backend",
+    "assistant",
+    "assistant_begin",
+    "assistant_end",
     "flush_cache",
-    "get_server_args",
+    "function",
     "gen",
     "gen_int",
     "gen_string",
+    "get_server_args",
     "image",
-    "video",
     "select",
+    "set_default_backend",
     "system",
+    "system_begin",
+    "system_end",
     "user",
-    "assistant",
     "user_begin",
     "user_end",
-    "assistant_begin",
-    "assistant_end",
+    "video",
 ]
+
+# Global Configurations
+from sglang.global_config import global_config
+
+__all__ += ["global_config"]
+
+from sglang.version import __version__
+
+__all__ += ["__version__"]
+
+# SGL Backends
+from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
+from sglang.utils import LazyImport
+
+Anthropic = LazyImport("sglang.lang.backend.anthropic", "Anthropic")
+LiteLLM = LazyImport("sglang.lang.backend.litellm", "LiteLLM")
+OpenAI = LazyImport("sglang.lang.backend.openai", "OpenAI")
+VertexAI = LazyImport("sglang.lang.backend.vertexai", "VertexAI")
+
+__all__ += ["Anthropic", "LiteLLM", "OpenAI", "VertexAI", "RuntimeEndpoint"]

sglang/api.py

@@ -75,7 +75,7 @@ def gen(
     choices: Optional[List[str]] = None,
     regex: Optional[str] = None,
 ):
-    """Call the model to generate. See the meaning of the arguments in docs/sampling_params.md"""
+    """Call the model to generate. See the meaning of the arguments in docs/en/sampling_params.md"""
 
     if choices:
         return SglSelect(name, choices, 0.0 if temperature is None else temperature)
@@ -210,6 +210,14 @@ def assistant(expr: Optional[SglExpr] = None):
     return _role_common("assistant", expr)
 
 
+def system_begin():
+    return SglRoleBegin("system")
+
+
+def system_end():
+    return SglRoleEnd("system")
+
+
 def user_begin():
     return SglRoleBegin("user")
 

sglang/bench_latency.py

@@ -37,9 +37,9 @@ import torch
 import torch.distributed as dist
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
-from sglang.srt.managers.controller.infer_batch import Batch, ForwardMode, Req
-from sglang.srt.managers.controller.model_runner import ModelRunner
+from sglang.srt.managers.schedule_batch import Batch, ForwardMode, Req
 from sglang.srt.model_config import ModelConfig
+from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.sampling_params import SamplingParams
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import suppress_other_loggers

sglang/bench_serving.py

@@ -1,5 +1,6 @@
 # Adapted from https://github.com/vllm-project/vllm/blob/6366efc67b0aedd2c1721c14385370e50b297fb3/benchmarks/backend_request_func.py
 # Adapted from https://github.com/vllm-project/vllm/blob/6366efc67b0aedd2c1721c14385370e50b297fb3/benchmarks/benchmark_serving.py
+
 """
 Benchmark online serving.
 
@@ -84,6 +85,9 @@ async def async_request_trt_llm(
             "min_length": request_func_input.output_len,
             "end_id": 1048576,
         }
+        if args.disable_ignore_eos:
+            del payload["min_length"]
+            del payload["end_id"]
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
 
@@ -149,7 +153,7 @@ async def async_request_openai_completions(
             "best_of": 1,
             "max_tokens": request_func_input.output_len,
             "stream": not args.disable_stream,
-            "ignore_eos": True,
+            "ignore_eos": not args.disable_ignore_eos,
         }
         headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
 
@@ -969,6 +973,11 @@ if __name__ == "__main__":
         action="store_true",
         help="Disable streaming mode.",
     )
+    parser.add_argument(
+        "--disable-ignore-eos",
+        action="store_true",
+        help="Disable ignoring EOS.",
+    )
 
     set_ulimit()
 

sglang/check_env.py

@@ -22,7 +22,7 @@ PACKAGE_LIST = [
     "huggingface_hub",
     "interegular",
     "packaging",
-    "pillow",
+    "PIL",
     "psutil",
     "pydantic",
     "uvicorn",

sglang/lang/backend/litellm.py

@@ -61,7 +61,7 @@ class LiteLLM(BaseBackend):
             model=self.model_name,
             messages=messages,
             **self.client_params,
-            **sampling_params.to_anthropic_kwargs(),
+            **sampling_params.to_litellm_kwargs(),
         )
         comp = ret.choices[0].message.content
 

sglang/lang/backend/openai.py

@@ -18,7 +18,7 @@ except ImportError as e:
     openai = tiktoken = e
 
 
-logger = logging.getLogger("openai")
+logger = logging.getLogger(__name__)
 
 
 def create_logit_bias_int(tokenizer):

sglang/lang/interpreter.py

@@ -553,6 +553,7 @@ class StreamExecutor:
                 "output_token_logprobs": output_token_logprobs,
             }
             self.variable_event[name].set()
+            self.stream_var_event[name].set()
         self.text_ += decision
 
     def _execute_variable(self, expr: SglVariable):
@@ -705,9 +706,9 @@ class ProgramState:
 
     def _role_common(self, name: str, expr: Optional[SglExpr] = None):
         if expr is not None:
-            self.stream_executor.submit(
-                SglExprList([SglRoleBegin(name), expr, SglRoleEnd(name)])
-            )
+            role_expr = SglExprList([SglRoleBegin(name), expr, SglRoleEnd(name)])
+            self.stream_executor.submit(role_expr)
+            return role_expr
         else:
 
             @contextmanager
@@ -778,7 +779,14 @@ class ProgramState:
                     if self.stream_executor.is_finished:
                         break
             else:
-                event = self.stream_executor.stream_var_event[var_name]
+                event = None
+                while not event:
+                    if var_name in self.stream_executor.stream_var_event:
+                        event = self.stream_executor.stream_var_event[var_name]
+                    if self.stream_executor.is_finished:
+                        yield ""
+                        return
+
                 while True:
                     event.wait()
                     event.clear()
@@ -813,7 +821,14 @@ class ProgramState:
                     if self.stream_executor.is_finished:
                         break
             else:
-                event = self.stream_executor.stream_var_event[var_name]
+                event = None
+                while not event:
+                    if var_name in self.stream_executor.stream_var_event:
+                        event = self.stream_executor.stream_var_event[var_name]
+                    if self.stream_executor.is_finished:
+                        yield ""
+                        return
+
                 while True:
                     await loop.run_in_executor(None, event.wait)
                     event.clear()

sglang/lang/ir.py

@@ -410,7 +410,7 @@ class SglGen(SglExpr):
         dtype: Optional[type] = None,
         regex: Optional[str] = None,
     ):
-        """Call the model to generate. See the meaning of the arguments in docs/sampling_params.md"""
+        """Call the model to generate. See the meaning of the arguments in docs/en/sampling_params.md"""
         super().__init__()
         self.name = name
         self.sampling_params = SglSamplingParams(

sglang/srt/constrained/__init__.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 import json
 from typing import Dict, Optional, Union
 

sglang/srt/constrained/base_cache.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """Base cache class."""
 
 import time

sglang/srt/constrained/fsm_cache.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """Cache for the compressed finite state machine."""
 
 from sglang.srt.constrained import RegexGuide, TransformerTokenizer

sglang/srt/constrained/jump_forward.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """
 Faster constrained decoding.
 Reference: https://lmsys.org/blog/2024-02-05-compressed-fsm/

sglang/srt/conversation.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """Conversation templates."""
 
 # Adapted from
@@ -421,3 +436,14 @@ register_conv_template(
         sep2="</s>",
     )
 )
+
+# Reference: https://github.com/InternLM/lmdeploy/blob/387bf54b4f124e72aab30ae9755f562e435d3d01/lmdeploy/model.py#L425-L442
+register_conv_template(
+    Conversation(
+        name="internlm2-chat",
+        system_template="<|im_start|>system\n{system_message}",
+        roles=("<|im_start|>user", "<|im_start|>assistant"),
+        sep="\n",
+        stop_str=["<|im_end|>", "<|action_end|>"],
+    )
+)

sglang/srt/hf_transformers_utils.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """Utilities for Huggingface Transformers."""
 
 import functools

sglang/srt/layers/context_flashattention_nopad.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 # Adapted from
 # https://github.com/ModelTC/lightllm/blob/f2a54f0912293f683bf1d1695fd12c4098a5bf82/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py#L1
 import torch

sglang/srt/layers/extend_attention.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 import torch
 import triton
 import triton.language as tl

sglang/srt/layers/fused_moe.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 # Adapted from
 # https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/layers/fused_moe/fused_moe.py#L1
 """Fused MoE kernel."""

sglang/srt/layers/linear.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 # temporarily adapted from https://github.com/vllm-project/vllm/blob/e76466dde2bc9525d55165ceaa600d298c7bf773/vllm/model_executor/layers/linear.py
 # FIXME: refactor the linear abstraction
 from abc import abstractmethod

sglang/srt/layers/logits_processor.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """Logits processing."""
 
 import dataclasses
@@ -10,7 +25,7 @@ from vllm.distributed import (
     tensor_model_parallel_all_gather,
 )
 
-from sglang.srt.managers.controller.model_runner import ForwardMode, InputMetadata
+from sglang.srt.model_executor.model_runner import ForwardMode, InputMetadata
 
 
 @dataclasses.dataclass
@@ -77,33 +92,46 @@ class LogitsProcessor(nn.Module):
 
     @staticmethod
     def get_top_logprobs(all_logprobs, logits_metadata: LogitsMetadata):
-        # TODO: vectorize the code below
         if logits_metadata.forward_mode == ForwardMode.DECODE:
             output_top_logprobs = []
-            for i in range(all_logprobs.shape[0]):
-                k = logits_metadata.top_logprobs_nums[i]
-                t = all_logprobs[i].topk(k)
-                v_cpu = t.values.tolist()
-                p_cpu = t.indices.tolist()
-                output_top_logprobs.append(list(zip(v_cpu, p_cpu)))
+            max_k = max(logits_metadata.top_logprobs_nums)
+            ret = all_logprobs.topk(max_k, dim=1)
+            values = ret.values.tolist()
+            indices = ret.indices.tolist()
+            for i, k in enumerate(logits_metadata.top_logprobs_nums):
+                output_top_logprobs.append(list(zip(values[i][:k], indices[i][:k])))
             return None, output_top_logprobs
         else:
+            # TODO: vectorize the code below
             input_top_logprobs, output_top_logprobs = [], []
             pt = 0
             extend_seq_lens_cpu = logits_metadata.extend_seq_lens.tolist()
+
+            max_k = max(logits_metadata.top_logprobs_nums)
+            ret = all_logprobs.topk(max_k, dim=1)
+            values = ret.values.tolist()
+            indices = ret.indices.tolist()
+
             for i, extend_seq_len in enumerate(extend_seq_lens_cpu):
                 if extend_seq_len == 0:
                     input_top_logprobs.append([])
                     output_top_logprobs.append([])
                     continue
                 k = logits_metadata.top_logprobs_nums[i]
-                t = all_logprobs[pt : pt + extend_seq_len].topk(k)
-                vs_cpu = t.values.tolist()
-                ps_cpu = t.indices.tolist()
                 input_top_logprobs.append(
-                    [list(zip(vs_cpu[j], ps_cpu[j])) for j in range(len(vs_cpu) - 1)]
+                    [
+                        list(zip(values[pt + j][:k], indices[pt + j][:k]))
+                        for j in range(extend_seq_len - 1)
+                    ]
+                )
+                output_top_logprobs.append(
+                    list(
+                        zip(
+                            values[pt + extend_seq_len - 1][:k],
+                            indices[pt + extend_seq_len - 1][:k],
+                        )
+                    )
                 )
-                output_top_logprobs.append(list(zip(vs_cpu[-1], ps_cpu[-1])))
                 pt += extend_seq_len
 
             return input_top_logprobs, output_top_logprobs

sglang/srt/layers/quantization/__init__.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 # temporarily adapted from vLLM
 # FIXME: in progress of refactoring the model loader
 

sglang/srt/layers/quantization/fp8.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 # adapted from https://github.com/vllm-project/vllm/blob/e76466dde2bc9525d55165ceaa600d298c7bf773/vllm/model_executor/layers/quantization/fp8.py
 # FIXME refactor in progress
 from typing import Any, Dict, List, Optional, Union

sglang/srt/layers/radix_attention.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 """Radix attention."""
 
 import torch
@@ -7,7 +22,7 @@ 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 (
+from sglang.srt.model_executor.model_runner import (
     ForwardMode,
     InputMetadata,
     global_server_args_dict,
@@ -88,7 +103,7 @@ class RadixAttention(nn.Module):
         return o
 
     def extend_forward_flashinfer(self, q, k, v, input_metadata: InputMetadata):
-        if not input_metadata.use_ragged:
+        if not input_metadata.flashinfer_use_ragged:
             self.store_kv_cache(k, v, input_metadata)
 
             o = input_metadata.flashinfer_prefill_wrapper_paged.forward(

sglang/srt/layers/token_attention.py

@@ -1,3 +1,18 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
 # Adapted from
 # https://github.com/ModelTC/lightllm/blob/f2a54f0912293f683bf1d1695fd12c4098a5bf82/lightllm/models/llama/triton_kernel/token_attention_nopad_att1.py
 # https://github.com/ModelTC/lightllm/blob/f2a54f0912293f683bf1d1695fd12c4098a5bf82/lightllm/models/llama/triton_kernel/token_attention_softmax_and_reducev.py
@@ -5,7 +20,7 @@ import torch
 import triton
 import triton.language as tl
 
-from sglang.srt.managers.controller.infer_batch import global_server_args_dict
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 
 if global_server_args_dict.get("attention_reduce_in_fp32", False):
     REDUCE_TRITON_TYPE = tl.float32