sglang 0.4.1__py3-none-any.whl → 0.4.1.post2__py3-none-any.whl

sglang/srt/models/llama_eagle.py

@@ -0,0 +1,132 @@
+"""
+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/SafeAILab/EAGLE/blob/main/eagle/model/cnets.py
+"""Inference-only LLaMA-EAGLE model compatible with HuggingFace weights."""
+
+from typing import Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.models.llama import LlamaDecoderLayer, LlamaForCausalLM
+
+
+class LlamaDecoderLayer(LlamaDecoderLayer):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config, layer_id, quant_config, prefix)
+
+        # Skip the input_layernorm
+        # https://github.com/SafeAILab/EAGLE/blob/35c78f6cdc19a73e05cf5c330b4c358dad970c6a/eagle/model/cnets.py#L427
+        if layer_id == 0:
+            del self.input_layernorm
+            setattr(self, "input_layernorm", lambda x: x)
+
+
+class LlamaModel(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                LlamaDecoderLayer(
+                    config, i, quant_config=quant_config, prefix=f"model.layers.{i}"
+                )
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.fc = torch.nn.Linear(config.hidden_size * 2, config.hidden_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+
+        hidden_states = self.fc(
+            torch.cat((hidden_states, forward_batch.spec_info.hidden_states), dim=-1)
+        )
+
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+                residual,
+            )
+        return hidden_states + residual
+
+
+class LlamaForCausalLMEagle(LlamaForCausalLM):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config=None,
+    ) -> None:
+        nn.Module.__init__(self)
+        self.config = config
+        self.quant_config = quant_config
+        self.model = LlamaModel(config, quant_config=quant_config)
+        # Llama 3.2 1B Instruct set tie_word_embeddings to True
+        # Llama 3.1 8B Instruct set tie_word_embeddings to False
+        if self.config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size, config.hidden_size, quant_config=quant_config
+            )
+        self.logits_processor = LogitsProcessor(config)
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        for name, loaded_weight in weights:
+            if "lm_head" not in name:
+                name = "model." + name
+                super().load_weights([(name, loaded_weight)])
+
+
+EntryClass = [LlamaForCausalLMEagle]

sglang/srt/openai_api/adapter.py

@@ -65,10 +65,13 @@ from sglang.srt.openai_api.protocol import (
     FileDeleteResponse,
     FileRequest,
     FileResponse,
+    FunctionResponse,
     LogProbs,
+    ToolCall,
     TopLogprob,
     UsageInfo,
 )
+from sglang.srt.utils import TOOLS_TAG_LIST, parse_tool_response
 from sglang.utils import get_exception_traceback
 
 logger = logging.getLogger(__name__)
@@ -517,6 +520,7 @@ def v1_generate_request(
                 "repetition_penalty": request.repetition_penalty,
                 "regex": request.regex,
                 "json_schema": request.json_schema,
+                "ebnf": request.ebnf,
                 "n": request.n,
                 "no_stop_trim": request.no_stop_trim,
                 "ignore_eos": request.ignore_eos,
@@ -692,6 +696,14 @@ def v1_generate_response(request, ret, tokenizer_manager, to_file=False):
 
 async def v1_completions(tokenizer_manager, raw_request: Request):
     request_json = await raw_request.json()
+    if "extra_body" in request_json:
+        extra = request_json["extra_body"]
+        if "ebnf" in extra:
+            request_json["ebnf"] = extra["ebnf"]
+        if "regex" in extra:
+            request_json["regex"] = extra["regex"]
+        # remove extra_body to avoid pydantic conflict
+        del request_json["extra_body"]
     all_requests = [CompletionRequest(**request_json)]
     adapted_request, request = v1_generate_request(all_requests)
 
@@ -870,6 +882,21 @@ def v1_chat_generate_request(
         #    None skips any image processing in GenerateReqInput.
         if not isinstance(request.messages, str):
             # Apply chat template and its stop strings.
+            tools = None
+            if request.tools and request.tool_choice != "none":
+                request.skip_special_tokens = False
+                if request.stream:
+                    logger.warning("Streaming is not supported with tools.")
+                    request.stream = False
+                if not isinstance(request.tool_choice, str):
+                    tools = [
+                        item.function.model_dump()
+                        for item in request.tools
+                        if item.function.name == request.tool_choice.function.name
+                    ]
+                else:
+                    tools = [item.function.model_dump() for item in request.tools]
+
             if chat_template_name is None:
                 openai_compatible_messages = []
                 for message in request.messages:
@@ -893,6 +920,7 @@ def v1_chat_generate_request(
                     openai_compatible_messages,
                     tokenize=True,
                     add_generation_prompt=True,
+                    tools=tools,
                 )
                 if assistant_prefix:
                     prompt_ids += tokenizer_manager.tokenizer.encode(assistant_prefix)
@@ -936,6 +964,7 @@ def v1_chat_generate_request(
             "frequency_penalty": request.frequency_penalty,
             "repetition_penalty": request.repetition_penalty,
             "regex": request.regex,
+            "ebnf": request.ebnf,
             "n": request.n,
             "no_stop_trim": request.no_stop_trim,
             "ignore_eos": request.ignore_eos,
@@ -1031,11 +1060,46 @@ def v1_chat_generate_response(request, ret, to_file=False, cache_report=False):
 
         finish_reason = ret_item["meta_info"]["finish_reason"]
 
+        tool_calls = None
+        text = ret_item["text"]
+
+        if isinstance(request, list):
+            tool_choice = request[idx].tool_choice
+            tools = request[idx].tools
+        else:
+            tool_choice = request.tool_choice
+            tools = request.tools
+
+        if tool_choice != "none" and any([i in text for i in TOOLS_TAG_LIST]):
+            if finish_reason == "stop":
+                finish_reason = "tool_calls"
+            try:
+                text, call_info_list = parse_tool_response(text, tools)  # noqa
+                tool_calls = [
+                    ToolCall(
+                        id=str(call_info[0]),
+                        function=FunctionResponse(
+                            name=call_info[1], arguments=call_info[2]
+                        ),
+                    )
+                    for call_info in call_info_list
+                ]
+            except Exception as e:
+                logger.error(f"Exception: {e}")
+                return create_error_response(
+                    HTTPStatus.BAD_REQUEST,
+                    "Failed to parse fc related info to json format!",
+                )
+
         if to_file:
             # to make the choice data json serializable
             choice_data = {
                 "index": 0,
-                "message": {"role": "assistant", "content": ret_item["text"]},
+                "message": {
+                    "role": "assistant",
+                    "content": ret_item["text"] if tool_calls is None else None,
+                    "tool_calls": tool_calls,
+                },
                 "logprobs": choice_logprobs,
                 "finish_reason": (finish_reason["type"] if finish_reason else ""),
                 "matched_stop": (
@@ -1047,7 +1111,11 @@ def v1_chat_generate_response(request, ret, to_file=False, cache_report=False):
         else:
             choice_data = ChatCompletionResponseChoice(
                 index=idx,
-                message=ChatMessage(role="assistant", content=ret_item["text"]),
+                message=ChatMessage(
+                    role="assistant",
+                    content=ret_item["text"] if tool_calls is None else None,
+                    tool_calls=tool_calls,
+                ),
                 logprobs=choice_logprobs,
                 finish_reason=(finish_reason["type"] if finish_reason else ""),
                 matched_stop=(
@@ -1108,6 +1176,15 @@ def v1_chat_generate_response(request, ret, to_file=False, cache_report=False):
 
 async def v1_chat_completions(tokenizer_manager, raw_request: Request):
     request_json = await raw_request.json()
+    if "extra_body" in request_json:
+        extra = request_json["extra_body"]
+        # For example, if 'ebnf' is given:
+        if "ebnf" in extra:
+            request_json["ebnf"] = extra["ebnf"]
+        if "regex" in extra:
+            request_json["regex"] = extra["regex"]
+        # remove extra_body to avoid pydantic conflict
+        del request_json["extra_body"]
     all_requests = [ChatCompletionRequest(**request_json)]
     adapted_request, request = v1_chat_generate_request(all_requests, tokenizer_manager)
 

sglang/srt/openai_api/protocol.py

@@ -179,6 +179,7 @@ class CompletionRequest(BaseModel):
     ignore_eos: bool = False
     skip_special_tokens: bool = True
     lora_path: Optional[Union[List[Optional[str]], Optional[str]]] = None
+    ebnf: Optional[str] = None
 
 
 class CompletionResponseChoice(BaseModel):
@@ -256,6 +257,34 @@ class ResponseFormat(BaseModel):
     json_schema: Optional[JsonSchemaResponseFormat] = None
 
 
+class Function(BaseModel):
+    """Function descriptions."""
+
+    description: Optional[str] = Field(default=None, examples=[None])
+    name: str
+    parameters: Optional[object] = None
+
+
+class Tool(BaseModel):
+    """Function wrapper."""
+
+    type: str = Field(default="function", examples=["function"])
+    function: Function
+
+
+class ToolChoiceFuncName(BaseModel):
+    """The name of tool choice function."""
+
+    name: str
+
+
+class ToolChoice(BaseModel):
+    """The tool choice definition."""
+
+    function: ToolChoiceFuncName
+    type: Literal["function"] = Field(default="function", examples=["function"])
+
+
 class ChatCompletionRequest(BaseModel):
     # Ordered by official OpenAI API documentation
     # https://platform.openai.com/docs/api-reference/chat/create
@@ -276,6 +305,10 @@ class ChatCompletionRequest(BaseModel):
     temperature: float = 0.7
     top_p: float = 1.0
     user: Optional[str] = None
+    tools: Optional[List[Tool]] = Field(default=None, examples=[None])
+    tool_choice: Union[ToolChoice, Literal["auto", "required", "none"]] = Field(
+        default="auto", examples=["none"]
+    )  # noqa
 
     # Extra parameters for SRT backend only and will be ignored by OpenAI models.
     top_k: int = -1
@@ -288,11 +321,28 @@ class ChatCompletionRequest(BaseModel):
     ignore_eos: bool = False
     skip_special_tokens: bool = True
     lora_path: Optional[Union[List[Optional[str]], Optional[str]]] = None
+    ebnf: Optional[str] = None
+
+
+class FunctionResponse(BaseModel):
+    """Function response."""
+
+    name: str
+    arguments: str
+
+
+class ToolCall(BaseModel):
+    """Tool call response."""
+
+    id: str
+    type: Literal["function"] = "function"
+    function: FunctionResponse
 
 
 class ChatMessage(BaseModel):
     role: Optional[str] = None
     content: Optional[str] = None
+    tool_calls: Optional[List[ToolCall]] = Field(default=None, examples=[None])
 
 
 class ChatCompletionResponseChoice(BaseModel):

sglang/srt/sampling/sampling_params.py

@@ -36,6 +36,7 @@ class SamplingParams:
         regex: Optional[str] = None,
         n: int = 1,
         json_schema: Optional[str] = None,
+        ebnf: Optional[str] = None,
         no_stop_trim: bool = False,
         ignore_eos: bool = False,
         skip_special_tokens: bool = True,
@@ -60,6 +61,7 @@ class SamplingParams:
         self.regex = regex
         self.n = n
         self.json_schema = json_schema
+        self.ebnf = ebnf
         self.no_stop_trim = no_stop_trim
 
         # Process some special cases
@@ -111,8 +113,13 @@ class SamplingParams:
                     f"min_new_tokens must be in (0, max_new_tokens({self.max_new_tokens})], got "
                     f"{self.min_new_tokens}."
                 )
-        if self.regex is not None and self.json_schema is not None:
-            raise ValueError("regex and json_schema cannot be both set.")
+        grammars = [
+            self.json_schema,
+            self.regex,
+            self.ebnf,
+        ]  # since mutually exclusive, only one can be set
+        if sum(x is not None for x in grammars) > 1:
+            raise ValueError("Only one of regex, json_schema, or ebnf can be set.")
 
     def normalize(self, tokenizer):
         # Process stop strings

sglang/srt/server.py

@@ -57,6 +57,7 @@ from sglang.srt.managers.io_struct import (
     OpenSessionReqInput,
     UpdateWeightFromDiskReqInput,
     UpdateWeightsFromDistributedReqInput,
+    UpdateWeightsFromTensorReqInput,
 )
 from sglang.srt.managers.scheduler import run_scheduler_process
 from sglang.srt.managers.tokenizer_manager import TokenizerManager
@@ -109,6 +110,7 @@ app.add_middleware(
 tokenizer_manager: TokenizerManager = None
 scheduler_info: Dict = None
 
+
 ##### Native API endpoints #####
 
 
@@ -245,16 +247,11 @@ async def get_weights_by_name(obj: GetWeightsByNameReqInput, request: Request):
     try:
         ret = await tokenizer_manager.get_weights_by_name(obj, request)
         if ret is None:
-            return ORJSONResponse(
-                {"error": {"message": "Get parameter by name failed"}},
-                status_code=HTTPStatus.BAD_REQUEST,
-            )
+            return _create_error_response("Get parameter by name failed")
         else:
             return ORJSONResponse(ret, status_code=200)
     except Exception as e:
-        return ORJSONResponse(
-            {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
-        )
+        return _create_error_response(e)
 
 
 @app.api_route("/open_session", methods=["GET", "POST"])
@@ -262,11 +259,13 @@ async def open_session(obj: OpenSessionReqInput, request: Request):
     """Open a session, and return its unique session id."""
     try:
         session_id = await tokenizer_manager.open_session(obj, request)
+        if session_id is None:
+            raise Exception(
+                "Failed to open the session. Check if a session with the same id is still open."
+            )
         return session_id
     except Exception as e:
-        return ORJSONResponse(
-            {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
-        )
+        return _create_error_response(e)
 
 
 @app.api_route("/close_session", methods=["GET", "POST"])
@@ -276,9 +275,7 @@ async def close_session(obj: CloseSessionReqInput, request: Request):
         await tokenizer_manager.close_session(obj, request)
         return Response(status_code=200)
     except Exception as e:
-        return ORJSONResponse(
-            {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
-        )
+        return _create_error_response(e)
 
 
 # fastapi implicitly converts json in the request to obj (dataclass)
@@ -312,9 +309,7 @@ async def generate_request(obj: GenerateReqInput, request: Request):
             return ret
         except ValueError as e:
             logger.error(f"Error: {e}")
-            return ORJSONResponse(
-                {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
-            )
+            return _create_error_response(e)
 
 
 @app.api_route("/encode", methods=["POST", "PUT"])
@@ -325,9 +320,7 @@ async def encode_request(obj: EmbeddingReqInput, request: Request):
         ret = await tokenizer_manager.generate_request(obj, request).__anext__()
         return ret
     except ValueError as e:
-        return ORJSONResponse(
-            {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
-        )
+        return _create_error_response(e)
 
 
 @app.api_route("/classify", methods=["POST", "PUT"])
@@ -338,9 +331,7 @@ async def classify_request(obj: EmbeddingReqInput, request: Request):
         ret = await tokenizer_manager.generate_request(obj, request).__anext__()
         return ret
     except ValueError as e:
-        return ORJSONResponse(
-            {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
-        )
+        return _create_error_response(e)
 
 
 ##### OpenAI-compatible API endpoints #####
@@ -416,6 +407,12 @@ async def retrieve_file_content(file_id: str):
     return await v1_retrieve_file_content(file_id)
 
 
+def _create_error_response(e):
+    return ORJSONResponse(
+        {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
+    )
+
+
 def launch_engine(
     server_args: ServerArgs,
 ):
@@ -493,7 +490,16 @@ def launch_engine(
     # Wait for model to finish loading
     scheduler_infos = []
     for i in range(len(scheduler_pipe_readers)):
-        data = scheduler_pipe_readers[i].recv()
+        try:
+            data = scheduler_pipe_readers[i].recv()
+        except EOFError as e:
+            logger.exception(e)
+            logger.error(
+                f"Rank {i} scheduler is dead. Please check if there are relevant logs."
+            )
+            scheduler_procs[i].join()
+            logger.error(f"Exit code: {scheduler_procs[i].exitcode}")
+            raise
 
         if data["status"] != "ready":
             raise RuntimeError(
@@ -501,7 +507,7 @@ def launch_engine(
             )
         scheduler_infos.append(data)
 
-    # Assume all schedulers have same max_total_num_tokens
+    # Assume all schedulers have same scheduler_info
     scheduler_info = scheduler_infos[0]
 
 
@@ -849,12 +855,10 @@ class Engine:
             group_name=group_name,
             backend=backend,
         )
-
-        async def _init_group():
-            return await tokenizer_manager.init_weights_update_group(obj, None)
-
         loop = asyncio.get_event_loop()
-        return loop.run_until_complete(_init_group())
+        return loop.run_until_complete(
+            tokenizer_manager.init_weights_update_group(obj, None)
+        )
 
     def update_weights_from_distributed(self, name, dtype, shape):
         """Update weights from distributed source."""
@@ -863,22 +867,24 @@ class Engine:
             dtype=dtype,
             shape=shape,
         )
+        loop = asyncio.get_event_loop()
+        return loop.run_until_complete(
+            tokenizer_manager.update_weights_from_distributed(obj, None)
+        )
 
-        async def _update_weights():
-            return await tokenizer_manager.update_weights_from_distributed(obj, None)
-
+    def update_weights_from_tensor(self, name, tensor):
+        """Update weights from distributed source."""
+        obj = UpdateWeightsFromTensorReqInput(name=name, tensor=tensor)
         loop = asyncio.get_event_loop()
-        return loop.run_until_complete(_update_weights())
+        return loop.run_until_complete(
+            tokenizer_manager.update_weights_from_tensor(obj, None)
+        )
 
     def get_weights_by_name(self, name, truncate_size=100):
         """Get weights by parameter name."""
         obj = GetWeightsByNameReqInput(name=name, truncate_size=truncate_size)
-
-        async def _get_weights():
-            return await tokenizer_manager.get_weights_by_name(obj, None)
-
         loop = asyncio.get_event_loop()
-        return loop.run_until_complete(_get_weights())
+        return loop.run_until_complete(tokenizer_manager.get_weights_by_name(obj, None))
 
 
 class Runtime:
@@ -888,7 +894,7 @@ class Runtime:
     using the commond line interface.
 
     It is mainly used for the frontend language.
-    You should use the Engine class if you want to do normal offline processing.
+    You should use the Engine class above if you want to do normal offline processing.
     """
 
     def __init__(

sglang/srt/server_args.py

@@ -55,7 +55,7 @@ class ServerArgs:
     is_embedding: bool = False
     revision: Optional[str] = None
 
-    # Port
+    # Port for the HTTP server
     host: str = "127.0.0.1"
     port: int = 30000
 
@@ -68,6 +68,7 @@ class ServerArgs:
     schedule_policy: str = "lpm"
     schedule_conservativeness: float = 1.0
     cpu_offload_gb: int = 0
+    prefill_only_one_req: bool = False
 
     # Other runtime options
     tp_size: int = 1
@@ -94,6 +95,7 @@ class ServerArgs:
     # Data parallelism
     dp_size: int = 1
     load_balance_method: str = "round_robin"
+
     # Expert parallelism
     ep_size: int = 1
 
@@ -217,6 +219,13 @@ class ServerArgs:
             )
             self.disable_cuda_graph = True
 
+        # Expert parallelism
+        if self.enable_ep_moe:
+            self.ep_size = self.tp_size
+            logger.info(
+                f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
+            )
+
         # Others
         if self.enable_dp_attention:
             self.dp_size = self.tp_size
@@ -229,12 +238,6 @@ class ServerArgs:
                 "Data parallel size is adjusted to be the same as tensor parallel size. "
                 "Overlap scheduler is disabled."
             )
-        # Expert parallelism
-        if self.enable_ep_moe:
-            self.ep_size = self.tp_size
-            logger.info(
-                f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
-            )
 
         # GGUF
         if (
@@ -430,13 +433,18 @@ class ServerArgs:
             default=ServerArgs.schedule_conservativeness,
             help="How conservative the schedule policy is. A larger value means more conservative scheduling. Use a larger value if you see requests being retracted frequently.",
         )
-
         parser.add_argument(
             "--cpu-offload-gb",
             type=int,
             default=ServerArgs.cpu_offload_gb,
             help="How many GBs of RAM to reserve for CPU offloading",
         )
+        parser.add_argument(
+            "--prefill-only-one-req",
+            type=bool,
+            help="If true, we only prefill one request at one prefill batch",
+            default=ServerArgs.prefill_only_one_req,
+        )
 
         # Other runtime options
         parser.add_argument(
@@ -555,6 +563,7 @@ class ServerArgs:
                 "shortest_queue",
             ],
         )
+
         # Expert parallelism
         parser.add_argument(
             "--expert-parallel-size",
@@ -777,28 +786,6 @@ class ServerArgs:
             help="Delete the model checkpoint after loading the model.",
         )
 
-        # Deprecated arguments
-        parser.add_argument(
-            "--enable-overlap-schedule",
-            action=DeprecatedAction,
-            help="'--enable-overlap-schedule' is deprecated. It is enabled by default now. Please drop this argument.",
-        )
-        parser.add_argument(
-            "--disable-flashinfer",
-            action=DeprecatedAction,
-            help="'--disable-flashinfer' is deprecated. Please use '--attention-backend triton' instead.",
-        )
-        parser.add_argument(
-            "--disable-flashinfer-sampling",
-            action=DeprecatedAction,
-            help="'--disable-flashinfer-sampling' is deprecated. Please use '--sampling-backend pytroch' instead.",
-        )
-        parser.add_argument(
-            "--disable-disk-cache",
-            action=DeprecatedAction,
-            help="'--disable-disk-cache' is deprecated. Please use '--disable-outlines-disk-cache' instead.",
-        )
-
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
         args.tp_size = args.tensor_parallel_size