sglang 0.3.4__py3-none-any.whl → 0.3.4.post2__py3-none-any.whl

sglang/bench_latency.py

@@ -227,8 +227,9 @@ 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)

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

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 
+import logging
+import os
 from enum import IntEnum, auto
 from typing import Optional
 
@@ -20,6 +22,8 @@ from transformers import PretrainedConfig
 
 from sglang.srt.hf_transformers_utils import get_config, get_context_length
 
+logger = logging.getLogger(__name__)
+
 
 class AttentionArch(IntEnum):
     MLA = auto()
@@ -46,10 +50,29 @@ class ModelConfig:
             model_override_args=model_override_args,
         )
         self.hf_text_config = get_hf_text_config(self.hf_config)
+        derived_context_len = get_context_length(self.hf_text_config)
+        allow_long_context = os.environ.get(
+            "SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", None
+        )
+
         if context_length is not None:
-            self.context_len = context_length
+            if context_length > derived_context_len:
+                if allow_long_context:
+                    logger.warning(
+                        f"Warning: User-specified context_length ({context_length}) is greater than the derived context_length ({derived_context_len}). "
+                        f"This may lead to incorrect model outputs or CUDA errors."
+                    )
+                    self.context_len = context_length
+                else:
+                    raise ValueError(
+                        f"User-specified context_length ({context_length}) is greater than the derived context_length ({derived_context_len}). "
+                        f"This may lead to incorrect model outputs or CUDA errors. Note that the derived context_length may differ from max_position_embeddings in the model's config. "
+                        f"To allow overriding this maximum, set the env var SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1"
+                    )
+            else:
+                self.context_len = context_length
         else:
-            self.context_len = get_context_length(self.hf_text_config)
+            self.context_len = derived_context_len
 
         # Unify the config keys for hf_text_config
         self.head_dim = getattr(
@@ -89,6 +112,8 @@ 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/constrained/fsm_cache.py

@@ -73,9 +73,16 @@ class FSMCache(BaseToolCache):
     def init_value(self, key):
         key_type, key_string = key
         if key_type == "json":
-            regex = build_regex_from_schema(
-                key_string, whitespace_pattern=self.constrained_json_whitespace_pattern
-            )
+            try:
+                regex = build_regex_from_schema(
+                    key_string,
+                    whitespace_pattern=self.constrained_json_whitespace_pattern,
+                )
+            except NotImplementedError as e:
+                logger.warning(
+                    f"skip invalid json schema: json_schema={key_string}, {e=}"
+                )
+                return None, key_string
         elif key_type == "regex":
             regex = key_string
         else:

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
@@ -162,6 +163,8 @@ def get_tokenizer(
             "Using a slow tokenizer. This might cause a significant "
             "slowdown. Consider using a fast tokenizer instead."
         )
+
+    attach_additional_stop_token_ids(tokenizer)
     return tokenizer
 
 
@@ -180,4 +183,16 @@ def get_processor(
         tokenizer_revision=tokenizer_revision,
         **kwargs,
     )
+
+    attach_additional_stop_token_ids(processor.tokenizer)
     return processor
+
+
+def attach_additional_stop_token_ids(tokenizer):
+    # Special handling for stop token <|eom_id|> generated by llama 3 tool use.
+    if "<|eom_id|>" in tokenizer.get_added_vocab():
+        tokenizer.additional_stop_token_ids = set(
+            [tokenizer.get_added_vocab()["<|eom_id|>"]]
+        )
+    else:
+        tokenizer.additional_stop_token_ids = None

sglang/srt/layers/attention/__init__.py

@@ -1,8 +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
 
 
@@ -19,13 +21,22 @@ class AttentionBackend(ABC):
         raise NotImplementedError()
 
     def init_forward_metadata_capture_cuda_graph(
-        self, bs: int, req_pool_indices: torch.Tensor, seq_lens: torch.Tensor
+        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: torch.Tensor, seq_lens: torch.Tensor
+        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()
@@ -39,7 +50,7 @@ class AttentionBackend(ABC):
         q: torch.Tensor,
         k: torch.Tensor,
         v: torch.Tensor,
-        layer: nn.Module,
+        layer: RadixAttention,
         forward_batch: ForwardBatch,
     ):
         """Run forward on an attention layer."""
@@ -53,7 +64,7 @@ class AttentionBackend(ABC):
         q: torch.Tensor,
         k: torch.Tensor,
         v: torch.Tensor,
-        layer: nn.Module,
+        layer: RadixAttention,
         forward_batch: ForwardBatch,
     ):
         """Run a forward for decode."""
@@ -64,7 +75,7 @@ class AttentionBackend(ABC):
         q: torch.Tensor,
         k: torch.Tensor,
         v: torch.Tensor,
-        layer: nn.Module,
+        layer: RadixAttention,
         forward_batch: ForwardBatch,
     ):
         """Run a forward for extend."""

sglang/srt/layers/attention/double_sparsity_backend.py

@@ -10,6 +10,7 @@ from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 if TYPE_CHECKING:
+    from sglang.srt.layers.radix_attention import RadixAttention
     from sglang.srt.model_executor.model_runner import ModelRunner
 
 
@@ -134,8 +135,13 @@ class DoubleSparseAttnBackend(AttentionBackend):
         )
 
     def init_forward_metadata_capture_cuda_graph(
-        self, bs: int, req_pool_indices: torch.Tensor, seq_lens: torch.Tensor
+        self,
+        bs: int,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        encoder_lens=None,
     ):
+        # NOTE: encoder_lens expected to be zeros or None
         self.forward_metadata = (
             self.cuda_graph_start_loc,
             self.cuda_graph_attn_logits,
@@ -144,15 +150,23 @@ class DoubleSparseAttnBackend(AttentionBackend):
         )
 
     def init_forward_metadata_replay_cuda_graph(
-        self, bs: int, req_pool_indices: torch.Tensor, seq_lens: torch.Tensor
+        self,
+        bs: int,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        seq_lens_sum: int,
+        encoder_lens=None,
     ):
+        # NOTE: encoder_lens expected to be zeros or None
         self.cuda_graph_start_loc.zero_()
         self.cuda_graph_start_loc[1:bs] = torch.cumsum(seq_lens[: bs - 1], dim=0)
 
     def get_cuda_graph_seq_len_fill_value(self):
         return 1
 
-    def forward_extend(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
+    def forward_extend(
+        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+    ):
         # TODO: reuse the buffer across layers
         if layer.qk_head_dim != layer.v_head_dim:
             o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
@@ -168,7 +182,7 @@ class DoubleSparseAttnBackend(AttentionBackend):
         )
 
         forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer.layer_id, forward_batch.out_cache_loc, k, v, k_label
+            layer, forward_batch.out_cache_loc, k, v, k_label
         )
 
         (
@@ -197,7 +211,9 @@ class DoubleSparseAttnBackend(AttentionBackend):
         )
         return o
 
-    def forward_decode(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
+    def forward_decode(
+        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+    ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
         q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)
@@ -227,7 +243,7 @@ class DoubleSparseAttnBackend(AttentionBackend):
         )
 
         forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer.layer_id, forward_batch.out_cache_loc, k, v, k_label
+            layer, forward_batch.out_cache_loc, k, v, k_label
         )
 
         # NOTE(Andy) shouldn't be used when max_len_in_batch < heavy_token_num