sglang 0.3.3.post1__py3-none-any.whl → 0.3.4.post1__py3-none-any.whl

sglang/bench_latency.py

@@ -227,22 +227,24 @@ def extend(reqs, model_runner):
         req_to_token_pool=model_runner.req_to_token_pool,
         token_to_kv_pool=model_runner.token_to_kv_pool,
         tree_cache=None,
+        model_config=model_runner.model_config,
     )
-    batch.prepare_for_extend(model_runner.model_config.vocab_size)
+    batch.prepare_for_extend()
     model_worker_batch = batch.get_model_worker_batch()
     forward_batch = ForwardBatch.init_new(model_worker_batch, model_runner)
     logits_output = model_runner.forward(forward_batch)
-    next_token_ids = model_runner.sample(logits_output, forward_batch).tolist()
+    next_token_ids = model_runner.sample(logits_output, forward_batch)
     return next_token_ids, logits_output.next_token_logits, batch
 
 
 @torch.inference_mode()
 def decode(input_token_ids, batch, model_runner):
-    batch.prepare_for_decode(input_token_ids)
+    batch.output_ids = input_token_ids
+    batch.prepare_for_decode()
     model_worker_batch = batch.get_model_worker_batch()
     forward_batch = ForwardBatch.init_new(model_worker_batch, model_runner)
     logits_output = model_runner.forward(forward_batch)
-    next_token_ids = model_runner.sample(logits_output, forward_batch).tolist()
+    next_token_ids = model_runner.sample(logits_output, forward_batch)
     return next_token_ids, logits_output.next_token_logits
 
 
@@ -252,6 +254,7 @@ def correctness_test(
     bench_args,
     tp_rank,
 ):
+    configure_logger(server_args, prefix=f" TP{tp_rank}")
     rank_print = print if tp_rank == 0 else lambda *args, **kwargs: None
 
     # Load the model
@@ -279,8 +282,9 @@ def correctness_test(
     output_ids = [input_ids[i] + [next_token_ids[i]] for i in range(len(input_ids))]
     for _ in range(bench_args.output_len[0] - 1):
         next_token_ids, _ = decode(next_token_ids, batch, model_runner)
+        next_token_ids_list = next_token_ids.tolist()
         for i in range(len(reqs)):
-            output_ids[i].append(next_token_ids[i])
+            output_ids[i].append(next_token_ids_list[i])
 
     # Print
     for i in range(len(reqs)):
@@ -288,8 +292,15 @@ def correctness_test(
         rank_print(tokenizer.decode(output_ids[i]), "\n")
 
 
+def synchronize(device):
+    if device == "cuda":
+        torch.cuda.synchronize()
+    elif device == "xpu":
+        torch.xpu.synchronize()
+
+
 def latency_test_run_once(
-    run_name, model_runner, rank_print, reqs, batch_size, input_len, output_len
+    run_name, model_runner, rank_print, reqs, batch_size, input_len, output_len, device
 ):
     max_batch_size = model_runner.max_total_num_tokens // (input_len + output_len)
     if batch_size > max_batch_size:
@@ -312,10 +323,10 @@ def latency_test_run_once(
     tot_latency = 0
 
     # Prefill
-    torch.cuda.synchronize()
+    synchronize(device)
     tic = time.time()
     next_token_ids, _, batch = extend(reqs, model_runner)
-    torch.cuda.synchronize()
+    synchronize(device)
     prefill_latency = time.time() - tic
     tot_latency += prefill_latency
     throughput = input_len * batch_size / prefill_latency
@@ -328,10 +339,10 @@ def latency_test_run_once(
     # Decode
     decode_latencies = []
     for i in range(output_len - 1):
-        torch.cuda.synchronize()
+        synchronize(device)
         tic = time.time()
         next_token_ids, _ = decode(next_token_ids, batch, model_runner)
-        torch.cuda.synchronize()
+        synchronize(device)
         latency = time.time() - tic
         tot_latency += latency
         throughput = batch_size / latency
@@ -387,6 +398,7 @@ def latency_test(
         bench_args.batch_size[0],
         bench_args.input_len[0],
         8,  # shorter decoding to speed up the warmup
+        server_args.device,
     )
     rank_print("Benchmark ...")
 
@@ -397,7 +409,14 @@ def latency_test(
     ):
         reqs = prepare_synthetic_inputs_for_latency_test(bs, il)
         ret = latency_test_run_once(
-            bench_args.run_name, model_runner, rank_print, reqs, bs, il, ol
+            bench_args.run_name,
+            model_runner,
+            rank_print,
+            reqs,
+            bs,
+            il,
+            ol,
+            server_args.device,
         )
         if ret is not None:
             result_list.append(ret)

sglang/bench_server_latency.py

@@ -6,6 +6,8 @@ It accepts arguments similar to those of launch_server.py.
 Usage:
 
 python3 -m sglang.bench_server_latency --model meta-llama/Meta-Llama-3.1-8B --batch-size 1 16 64 --input-len 1024 --output-len 8
+
+python3 -m sglang.bench_server_latency --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8
 """
 
 import argparse
@@ -32,6 +34,8 @@ class BenchArgs:
     input_len: Tuple[int] = (1024,)
     output_len: Tuple[int] = (16,)
     result_filename: str = "result.jsonl"
+    base_url: str = ""
+    skip_warmup: bool = False
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -48,6 +52,8 @@ class BenchArgs:
         parser.add_argument(
             "--result-filename", type=str, default=BenchArgs.result_filename
         )
+        parser.add_argument("--base-url", type=str, default=BenchArgs.base_url)
+        parser.add_argument("--skip-warmup", action="store_true")
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
@@ -139,17 +145,21 @@ def run_one_case(
 
 
 def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
-    proc, base_url = launch_server_process(server_args)
+    if bench_args.base_url:
+        proc, base_url = None, bench_args.base_url
+    else:
+        proc, base_url = launch_server_process(server_args)
 
     # warmup
-    run_one_case(
-        base_url,
-        batch_size=16,
-        input_len=1024,
-        output_len=16,
-        run_name="",
-        result_filename="",
-    )
+    if not bench_args.skip_warmup:
+        run_one_case(
+            base_url,
+            batch_size=16,
+            input_len=1024,
+            output_len=16,
+            run_name="",
+            result_filename="",
+        )
 
     # benchmark
     try:
@@ -165,7 +175,8 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
                 bench_args.result_filename,
             )
     finally:
-        kill_child_process(proc.pid)
+        if proc:
+            kill_child_process(proc.pid)
 
     print(f"\nResults are saved to {bench_args.result_filename}")
 

sglang/bench_serving.py

@@ -222,6 +222,85 @@ async def async_request_openai_completions(
     return output
 
 
+async def async_request_sglang_generate(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    prompt = request_func_input.prompt
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        payload = {
+            "text": prompt,
+            "sampling_params": {
+                "temperature": 0.0,
+                "max_new_tokens": request_func_input.output_len,
+                "ignore_eos": not args.disable_ignore_eos,
+            },
+            "stream": not args.disable_stream,
+            **request_func_input.extra_request_body,
+        }
+        headers = {}
+
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        generated_text = ""
+        ttft = 0.0
+        st = time.perf_counter()
+        most_recent_timestamp = st
+        try:
+            async with session.post(
+                url=api_url, json=payload, headers=headers
+            ) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+                        # print(chunk_bytes)
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data: ")
+                        latency = time.perf_counter() - st
+                        if chunk == "[DONE]":
+                            pass
+                        else:
+                            data = json.loads(chunk)
+
+                            # NOTE: Some completion API might have a last
+                            # usage summary response without a token so we
+                            # want to check a token was generated
+                            if data["text"]:
+                                timestamp = time.perf_counter()
+                                # First token
+                                if ttft == 0.0:
+                                    ttft = time.perf_counter() - st
+                                    output.ttft = ttft
+
+                                # Decoding phase
+                                else:
+                                    output.itl.append(timestamp - most_recent_timestamp)
+
+                                most_recent_timestamp = timestamp
+                                generated_text = data["text"]
+
+                    output.generated_text = generated_text
+                    output.success = True
+                    output.latency = latency
+                    output.output_len = request_func_input.output_len
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
 async def async_request_gserver(
     request_func_input: RequestFuncInput,
     pbar: Optional[tqdm] = None,
@@ -264,7 +343,9 @@ def get_tokenizer(
 
 
 ASYNC_REQUEST_FUNCS = {
-    "sglang": async_request_openai_completions,
+    "sglang": async_request_sglang_generate,
+    "sglang-native": async_request_sglang_generate,
+    "sglang-oai": async_request_openai_completions,
     "vllm": async_request_openai_completions,
     "lmdeploy": async_request_openai_completions,
     "trt": async_request_trt_llm,
@@ -387,6 +468,8 @@ def sample_sharegpt_requests(
             continue
         filtered_dataset.append((prompt, prompt_len, output_len))
 
+    print(f"#Input tokens: {np.sum([x[1] for x in filtered_dataset])}")
+    print(f"#Output tokens: {np.sum([x[2] for x in filtered_dataset])}")
     return filtered_dataset
 
 
@@ -587,6 +670,8 @@ async def benchmark(
     else:
         print("Initial test run completed. Starting main benchmark run...")
 
+    time.sleep(1.5)
+
     pbar = None if disable_tqdm else tqdm(total=len(input_requests))
 
     benchmark_start_time = time.perf_counter()
@@ -782,24 +867,33 @@ def run_benchmark(args_: argparse.Namespace):
     if args.port is None:
         args.port = {
             "sglang": 30000,
+            "sglang-native": 30000,
+            "sglang-oai": 30000,
             "lmdeploy": 23333,
             "vllm": 8000,
             "trt": 8000,
             "gserver": 9988,
         }.get(args.backend, 30000)
 
-    api_url = (
-        f"{args.base_url}/v1/completions"
-        if args.base_url
-        else f"http://{args.host}:{args.port}/v1/completions"
-    )
     model_url = (
         f"{args.base_url}/v1/models"
         if args.base_url
         else f"http://{args.host}:{args.port}/v1/models"
     )
 
-    if args.backend == "trt":
+    if args.backend in ["sglang", "sglang-native"]:
+        api_url = (
+            f"{args.base_url}/generate"
+            if args.base_url
+            else f"http://{args.host}:{args.port}/generate"
+        )
+    elif args.backend in ["sglang-oai", "vllm", "lmdeploy"]:
+        api_url = (
+            f"{args.base_url}/v1/completions"
+            if args.base_url
+            else f"http://{args.host}:{args.port}/v1/completions"
+        )
+    elif args.backend == "trt":
         api_url = (
             f"{args.base_url}/v2/models/ensemble/generate_stream"
             if args.base_url

sglang/global_config.py

@@ -19,7 +19,6 @@ class GlobalConfig:
         self.new_token_ratio_decay = 0.001
 
         # Runtime constants: others
-        self.num_continue_decode_steps = 10
         self.retract_decode_steps = 20
         self.flashinfer_workspace_size = os.environ.get(
             "FLASHINFER_WORKSPACE_SIZE", 384 * 1024 * 1024

sglang/lang/chat_template.py

@@ -133,6 +133,22 @@ register_chat_template(
     )
 )
 
+# Reference: https://huggingface.co/docs/transformers/main/model_doc/qwen2_vl#usage-example
+register_chat_template(
+    ChatTemplate(
+        name="qwen2-vl",
+        default_system_prompt="You are a helpful assistant.",
+        role_prefix_and_suffix={
+            "system": ("<|im_start|>system\n", "<|im_end|>\n"),
+            "user": ("<|im_start|>user\n", "<|im_end|>\n"),
+            "assistant": ("<|im_start|>assistant\n", "<|im_end|>\n"),
+        },
+        style=ChatTemplateStyle.PLAIN,
+        stop_str=("<|im_end|>"),
+        image_token="<|vision_start|><|image_pad|><|vision_end|>",
+    )
+)
+
 
 register_chat_template(
     ChatTemplate(
@@ -213,6 +229,7 @@ register_chat_template(
             ),
         },
         stop_str=("<|eot_id|>",),
+        image_token="<|image|>",
     )
 )
 

sglang/launch_server_llavavid.py

@@ -14,7 +14,7 @@ if __name__ == "__main__":
     model_override_args["num_frames"] = 16
     model_override_args["model_type"] = "llavavid"
     if model_override_args["num_frames"] == 32:
-        model_override_args["rope_scaling"] = {"factor": 2.0, "type": "linear"}
+        model_override_args["rope_scaling"] = {"factor": 2.0, "rope_type": "linear"}
         model_override_args["max_sequence_length"] = 4096 * 2
         model_override_args["tokenizer_model_max_length"] = 4096 * 2
         model_override_args["model_max_length"] = 4096 * 2

sglang/srt/configs/__init__.py

@@ -1,5 +1,8 @@
 from sglang.srt.configs.exaone import ExaoneConfig
+from sglang.srt.configs.qwen2vl import Qwen2VLConfig, Qwen2VLVisionConfig
 
 __all__ = [
     "ExaoneConfig",
+    "Qwen2VLConfig",
+    "Qwen2VLVisionConfig",
 ]

sglang/srt/configs/model_config.py

@@ -89,6 +89,8 @@ class ModelConfig:
         self.num_hidden_layers = self.hf_text_config.num_hidden_layers
         self.vocab_size = self.hf_text_config.vocab_size
 
+        self.is_encoder_decoder = self.hf_config.model_type in ["mllama"]
+
     # adapted from https://github.com/vllm-project/vllm/blob/main/vllm/config.py#L289
     def get_total_num_kv_heads(self) -> int:
         """Returns the total number of KV heads."""

sglang/srt/configs/qwen2vl.py

@@ -0,0 +1,133 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team.
+# All rights reserved.
+#
+# 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.
+"""Qwen2VL model configuration"""
+
+import os
+from typing import Union
+
+from transformers import PretrainedConfig
+
+
+class Qwen2VLVisionConfig(PretrainedConfig):
+    model_type = "qwen2_vl"
+
+    def __init__(
+        self,
+        depth=32,
+        embed_dim=1280,
+        hidden_size=3584,
+        hidden_act="quick_gelu",
+        mlp_ratio=4,
+        num_heads=16,
+        in_channels=3,
+        patch_size=14,
+        spatial_merge_size=2,
+        temporal_patch_size=2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.depth = depth
+        self.embed_dim = embed_dim
+        self.hidden_size = hidden_size
+        self.hidden_act = hidden_act
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.in_channels = in_channels
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.temporal_patch_size = temporal_patch_size
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+    ) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        if config_dict.get("model_type") == "qwen2_vl":
+            config_dict = config_dict["vision_config"]
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class Qwen2VLConfig(PretrainedConfig):
+    model_type = "qwen2_vl"
+
+    def __init__(
+        self,
+        vocab_size=152064,
+        hidden_size=8192,
+        intermediate_size=29568,
+        num_hidden_layers=80,
+        num_attention_heads=64,
+        num_key_value_heads=8,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-05,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=1000000.0,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=80,
+        attention_dropout=0.0,
+        vision_config=None,
+        rope_scaling=None,
+        **kwargs,
+    ):
+        if isinstance(vision_config, dict):
+            self.vision_config = Qwen2VLVisionConfig(**vision_config)
+        elif vision_config is None:
+            self.vision_config = Qwen2VLVisionConfig()
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.rope_scaling = rope_scaling
+
+        # NOTE: the following section from original transformers config
+        # for Qwen2-VL is commented out to address rope config loading issue
+        #
+        # if self.rope_scaling is not None and "type" in self.rope_scaling:
+        #     if self.rope_scaling["type"] == "mrope":
+        #         self.rope_scaling["type"] = "default"
+        #     self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        # rope_config_validation(self)
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

sglang/srt/conversation.py

@@ -509,6 +509,19 @@ register_conv_template(
     )
 )
 
+register_conv_template(
+    Conversation(
+        name="llama_3_vision",
+        system_message="You are a helpful language and vision assistant. You are able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language.",
+        system_template="<|start_header_id|>system<|end_header_id|>\n\n{system_message}<|eot_id|>",
+        roles=("user", "assistant"),
+        sep_style=SeparatorStyle.LLAMA3,
+        sep="",
+        stop_str=["<|end_of_text|>", "<|eot_id|>"],
+        image_token="<|image|>",
+    )
+)
+
 register_conv_template(
     Conversation(
         name="llava_llama_3",
@@ -530,3 +543,17 @@ register_conv_template(
         stop_str=["<|im_end|>", "<|action_end|>"],
     )
 )
+
+# Reference: https://huggingface.co/docs/transformers/main/model_doc/qwen2_vl#usage-example
+register_conv_template(
+    Conversation(
+        name="qwen2-vl",
+        system_message="You are a helpful assistant.",
+        system_template="<|im_start|>system\n{system_message}",
+        roles=("<|im_start|>user", "<|im_start|>assistant"),
+        sep="<|im_end|>\n",
+        sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
+        stop_str=["<|im_end|>"],
+        image_token="<|vision_start|><|image_pad|><|vision_end|>",
+    )
+)

sglang/srt/hf_transformers_utils.py

@@ -33,12 +33,13 @@ from transformers import (
 try:
     from vllm.transformers_utils.configs import ChatGLMConfig, DbrxConfig
 
-    from sglang.srt.configs import ExaoneConfig
+    from sglang.srt.configs import ExaoneConfig, Qwen2VLConfig
 
     _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
         ChatGLMConfig.model_type: ChatGLMConfig,
         DbrxConfig.model_type: DbrxConfig,
         ExaoneConfig.model_type: ExaoneConfig,
+        Qwen2VLConfig.model_type: Qwen2VLConfig,
     }
 except ImportError:
     # We want this file to run without vllm dependency

sglang/srt/layers/attention/__init__.py

@@ -1,7 +1,10 @@
 from abc import ABC, abstractmethod
+from typing import Optional
 
+import torch
 from torch import nn
 
+from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 
@@ -18,13 +21,22 @@ class AttentionBackend(ABC):
         raise NotImplementedError()
 
     def init_forward_metadata_capture_cuda_graph(
-        self, bs: int, req_pool_indices, seq_lens
+        self,
+        bs: int,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        encoder_lens: Optional[torch.Tensor] = None,
     ):
         """Init the metadata for a forward pass for capturing a cuda graph."""
         raise NotImplementedError()
 
     def init_forward_metadata_replay_cuda_graph(
-        self, bs: int, req_pool_indices, seq_lens
+        self,
+        bs: int,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        seq_lens_sum: int,
+        encoder_lens: Optional[torch.Tensor] = None,
     ):
         """Init the metadata for a forward pass for replying a cuda graph."""
         raise NotImplementedError()
@@ -33,17 +45,38 @@ class AttentionBackend(ABC):
         """Get the fill value for padded seq lens. Typically, it is 0 or 1."""
         raise NotImplementedError()
 
-    def forward(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
+    def forward(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+    ):
         """Run forward on an attention layer."""
         if forward_batch.forward_mode.is_decode():
             return self.forward_decode(q, k, v, layer, forward_batch)
         else:
             return self.forward_extend(q, k, v, layer, forward_batch)
 
-    def forward_decode(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
+    def forward_decode(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+    ):
         """Run a forward for decode."""
         raise NotImplementedError()
 
-    def forward_extend(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
+    def forward_extend(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+    ):
         """Run a forward for extend."""
         raise NotImplementedError()