sglang 0.4.6.post1__py3-none-any.whl → 0.4.6.post2__py3-none-any.whl

sglang/srt/models/llama.py

@@ -17,13 +17,14 @@
 """Inference-only LLaMA model compatible with HuggingFace weights."""
 
 import logging
-from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
 
 import torch
 from torch import nn
 from transformers import LlamaConfig
 
 from sglang.srt.distributed import (
+    get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
 )
@@ -39,11 +40,12 @@ from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import (
     default_weight_loader,
     kv_cache_scales_loader,
@@ -275,21 +277,31 @@ class LlamaModel(nn.Module):
         self.config = config
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
-            quant_config=quant_config,
-            prefix=add_prefix("embed_tokens", prefix),
-        )
-        self.layers = make_layers(
+        self.pp_group = get_pp_group()
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.layers, self.start_layer, self.end_layer = make_layers(
             config.num_hidden_layers,
             lambda idx, prefix: LlamaDecoderLayer(
-                config=config, layer_id=idx, quant_config=quant_config, prefix=prefix
+                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
             ),
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
             prefix="model.layers",
         )
 
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
         self.layers_to_capture = []
 
     def forward(
@@ -298,14 +310,23 @@ class LlamaModel(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]]]:
-        if input_embeds is None:
-            hidden_states = self.embed_tokens(input_ids)
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]], PPProxyTensors]:
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
         else:
-            hidden_states = input_embeds
-        residual = None
+            assert pp_proxy_tensors is not None
+            # FIXME(@ying): reduce the number of proxy tensors by not fusing layer norms
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+            deferred_norm = None
+
         aux_hidden_states = []
-        for i in range(len(self.layers)):
+        for i in range(self.start_layer, self.end_layer):
             if i in self.layers_to_capture:
                 aux_hidden_states.append(hidden_states + residual)
             layer = self.layers[i]
@@ -315,7 +336,16 @@ class LlamaModel(nn.Module):
                 forward_batch,
                 residual,
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+        else:
+            hidden_states, _ = self.norm(hidden_states, residual)
 
         if len(aux_hidden_states) == 0:
             return hidden_states
@@ -376,6 +406,7 @@ class LlamaForCausalLM(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
         self.model = self._init_model(config, quant_config, add_prefix("model", prefix))
@@ -419,23 +450,41 @@ class LlamaForCausalLM(nn.Module):
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
         get_embedding: bool = False,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> LogitsProcessorOutput:
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
+        )
+
         aux_hidden_states = None
         if self.capture_aux_hidden_states:
-            hidden_states, aux_hidden_states = self.model(
-                input_ids, positions, forward_batch, input_embeds
-            )
+            hidden_states, aux_hidden_states = hidden_states
+
+        if self.pp_group.is_last_rank:
+            if not get_embedding:
+                return self.logits_processor(
+                    input_ids,
+                    hidden_states,
+                    self.lm_head,
+                    forward_batch,
+                    aux_hidden_states,
+                )
+            else:
+                return self.pooler(hidden_states, forward_batch)
         else:
-            hidden_states = self.model(
-                input_ids, positions, forward_batch, input_embeds
-            )
+            return hidden_states
 
-        if not get_embedding:
-            return self.logits_processor(
-                input_ids, hidden_states, self.lm_head, forward_batch, aux_hidden_states
-            )
-        else:
-            return self.pooler(hidden_states, forward_batch)
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
 
     def get_input_embeddings(self) -> nn.Embedding:
         return self.model.embed_tokens
@@ -491,6 +540,16 @@ class LlamaForCausalLM(nn.Module):
         params_dict = dict(self.named_parameters())
 
         for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
             if "rotary_emb.inv_freq" in name or "projector" in name:
                 continue
             if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
@@ -637,6 +696,9 @@ class LlamaForCausalLM(nn.Module):
         self.model.load_kv_cache_scales(quantization_param_path)
 
     def set_eagle3_layers_to_capture(self):
+        if not self.pp_group.is_last_rank:
+            return
+
         self.capture_aux_hidden_states = True
         num_layers = self.config.num_hidden_layers
         self.model.layers_to_capture = [2, num_layers // 2, num_layers - 3]

sglang/srt/models/llama4.py

@@ -46,7 +46,7 @@ from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.models.llama import LlamaForCausalLM, LlamaMLP
 from sglang.srt.utils import add_prefix, fast_topk, get_compiler_backend, make_layers
 
@@ -431,6 +431,7 @@ class Llama4Model(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]]]:
         if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)

sglang/srt/models/llama_eagle.py

@@ -25,13 +25,14 @@ import torch
 from torch import nn
 from transformers import LlamaConfig
 
+from sglang.srt.distributed import get_pp_group
 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.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.models.llama import LlamaDecoderLayer, LlamaForCausalLM
 
 
@@ -86,6 +87,7 @@ class LlamaModel(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
         if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
@@ -118,6 +120,7 @@ class LlamaForCausalLMEagle(LlamaForCausalLM):
         nn.Module.__init__(self)
         self.config = config
         self.quant_config = quant_config
+        self.pp_group = get_pp_group()
         self.model = LlamaModel(
             config, quant_config=quant_config, prefix=add_prefix("model", prefix)
         )

sglang/srt/models/llama_eagle3.py

@@ -25,6 +25,7 @@ import torch
 from torch import nn
 from transformers import LlamaConfig
 
+from sglang.srt.distributed import get_pp_group
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
@@ -33,7 +34,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.models.llama import LlamaAttention, LlamaDecoderLayer, LlamaForCausalLM
 
 
@@ -118,6 +119,7 @@ class LlamaModel(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
         if input_embeds is None:
             embeds = self.embed_tokens(input_ids)
@@ -155,6 +157,7 @@ class LlamaForCausalLMEagle3(LlamaForCausalLM):
         nn.Module.__init__(self)
         self.config = config
         self.quant_config = quant_config
+        self.pp_group = get_pp_group()
 
         if self.config.num_hidden_layers != 1:
             raise ValueError("EAGLE3 currently only supports 1 layer")

sglang/srt/models/qwen2_moe.py

@@ -36,6 +36,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.ep_moe.layer import EPMoE
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
@@ -45,6 +46,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding,
 )
 from sglang.srt.managers.expert_distribution import ExpertDistributionRecorder
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, make_layers
@@ -108,12 +110,13 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
                 f"the number of experts {config.num_experts}."
             )
 
-        self.experts = FusedMoE(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        self.experts = MoEImpl(
             num_experts=config.num_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
@@ -427,7 +430,9 @@ class Qwen2MoeForCausalLM(nn.Module):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        expert_params_mapping = MoEImpl.make_expert_params_mapping(
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

sglang/srt/models/qwen2_vl.py

@@ -442,18 +442,6 @@ class Qwen2VLForConditionalGeneration(nn.Module):
         "up_proj": ("gate_up_proj", 1),
     }
 
-    def calculate_num_image_tokens(self, image_grid_thw: Tuple[int, int, int]):
-        processor = cached_get_processor(self.config._name_or_path)
-        grid_t, grid_h, grid_w = image_grid_thw
-        num_image_tokens = (
-            grid_t
-            * grid_h
-            * grid_w
-            // processor.image_processor.merge_size
-            // processor.image_processor.merge_size
-        )
-        return num_image_tokens
-
     def __init__(
         self,
         config: Qwen2VLConfig,

sglang/srt/models/qwen3_moe.py

@@ -40,6 +40,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.ep_moe.layer import EPMoE
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
@@ -48,6 +49,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2_moe import Qwen2MoeMLP as Qwen3MoeMLP
@@ -73,12 +75,13 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                 f"the number of experts {config.num_experts}."
             )
 
-        self.experts = FusedMoE(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        self.experts = MoEImpl(
             num_experts=config.num_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
@@ -356,7 +359,9 @@ class Qwen3MoeForCausalLM(nn.Module):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        expert_params_mapping = MoEImpl.make_expert_params_mapping(
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

sglang/srt/openai_api/adapter.py

@@ -36,6 +36,7 @@ from sglang.srt.conversation import (
     chat_template_exists,
     generate_chat_conv,
     generate_embedding_convs,
+    get_conv_template_by_model_path,
     register_conv_template,
 )
 from sglang.srt.function_call_parser import FunctionCallParser
@@ -163,10 +164,14 @@ def load_chat_template_for_openai_api(tokenizer_manager, chat_template_arg, mode
     else:
         chat_template_name = chat_template_arg
 
-    # Check chat-template
-    # TODO:
-    # 1. Do not import any code from sglang.lang
-    # 2. For VLM, when chat_template_arg is None, set it automatically by guessing from model_path.
+
+def guess_chat_template_name_from_model_path(model_path):
+    global chat_template_name
+    chat_template_name = get_conv_template_by_model_path(model_path)
+    if chat_template_name is not None:
+        logger.info(
+            f"Infer the chat template name from the model path and obtain the result: {chat_template_name}."
+        )
 
 
 async def v1_files_create(
@@ -894,6 +899,24 @@ async def v1_completions(tokenizer_manager, raw_request: Request):
     return response
 
 
+def _get_enable_thinking_from_request(request_obj):
+    """Extracts the 'enable_thinking' flag from request chat_template_kwargs.
+
+    Args:
+        request_obj: The request object (or an item from a list of requests).
+
+    Returns:
+        The boolean value of 'enable_thinking' if found and not True, otherwise True.
+    """
+    if (
+        hasattr(request_obj, "chat_template_kwargs")
+        and request_obj.chat_template_kwargs
+        and request_obj.chat_template_kwargs.get("enable_thinking") is not None
+    ):
+        return request_obj.chat_template_kwargs.get("enable_thinking")
+    return True
+
+
 def v1_chat_generate_request(
     all_requests: List[ChatCompletionRequest],
     tokenizer_manager,
@@ -1099,7 +1122,7 @@ def v1_chat_generate_request(
 
         sampling_params = {
             "temperature": request.temperature,
-            "max_new_tokens": request.max_tokens,
+            "max_new_tokens": request.max_tokens or request.max_completion_tokens,
             "min_new_tokens": request.min_tokens,
             "stop": stop,
             "stop_token_ids": request.stop_token_ids,
@@ -1258,31 +1281,16 @@ def v1_chat_generate_response(
         tool_calls = None
         text = ret_item["text"]
 
-        enable_thinking = True
         if isinstance(request, list):
             tool_choice = request[idx].tool_choice
             tools = request[idx].tools
             separate_reasoning = request[idx].separate_reasoning
-
-            if (
-                request[idx].chat_template_kwargs
-                and request[idx].chat_template_kwargs.get("enable_thinking") is not None
-            ):
-                enable_thinking = request[idx].chat_template_kwargs.get(
-                    "enable_thinking", True
-                )
+            enable_thinking = _get_enable_thinking_from_request(request[idx])
         else:
             tool_choice = request.tool_choice
             tools = request.tools
             separate_reasoning = request.separate_reasoning
-
-            if (
-                request.chat_template_kwargs
-                and request.chat_template_kwargs.get("enable_thinking") is not None
-            ):
-                enable_thinking = request.chat_template_kwargs.get(
-                    "enable_thinking", True
-                )
+            enable_thinking = _get_enable_thinking_from_request(request)
 
         reasoning_text = None
         if reasoning_parser and separate_reasoning and enable_thinking:
@@ -1521,9 +1529,12 @@ async def v1_chat_completions(
                     delta = text[len(stream_buffer) :]
                     new_stream_buffer = stream_buffer + delta
 
+                    enable_thinking = _get_enable_thinking_from_request(request)
+
                     if (
                         tokenizer_manager.server_args.reasoning_parser
                         and request.separate_reasoning
+                        and enable_thinking
                     ):
                         if index not in reasoning_parser_dict:
                             reasoning_parser_dict[index] = ReasoningParser(
@@ -1613,6 +1624,7 @@ async def v1_chat_completions(
 
                             tool_call = ToolCall(
                                 id=str(call_item.tool_index),
+                                index=call_item.tool_index,
                                 function=FunctionResponse(
                                     name=call_item.name,
                                     arguments=call_item.parameters,

sglang/srt/openai_api/protocol.py

@@ -320,7 +320,16 @@ class ChatCompletionRequest(BaseModel):
     logit_bias: Optional[Dict[str, float]] = None
     logprobs: bool = False
     top_logprobs: Optional[int] = None
-    max_tokens: Optional[int] = None
+    max_tokens: Optional[int] = Field(
+        default=None,
+        deprecated="max_tokens is deprecated in favor of the max_completion_tokens field",
+        description="The maximum number of tokens that can be generated in the chat completion. ",
+    )
+    max_completion_tokens: Optional[int] = Field(
+        default=None,
+        description="The maximum number of completion tokens for a chat completion request, "
+        "including visible output tokens and reasoning tokens. Input tokens are not included. ",
+    )
     n: int = 1
     presence_penalty: float = 0.0
     response_format: Optional[Union[ResponseFormat, StructuralTagResponseFormat]] = None
@@ -380,6 +389,7 @@ class ToolCall(BaseModel):
     """Tool call response."""
 
     id: str
+    index: Optional[int] = None
     type: Literal["function"] = "function"
     function: FunctionResponse