sglang 0.1.18__py3-none-any.whl → 0.1.19__py3-none-any.whl

sglang/srt/models/qwen2_moe.py

@@ -0,0 +1,454 @@
+# coding=utf-8
+# Adapted from
+# https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/qwen2_moe.py
+"""Inference-only Qwen2MoE model compatible with HuggingFace weights."""
+from typing import Any, Dict, Iterable, List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.config import CacheConfig
+from vllm.distributed import (get_tensor_model_parallel_world_size,
+                              tensor_model_parallel_all_reduce)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors, SamplerOutput
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.controller.model_runner import InputMetadata
+
+class Qwen2MoeMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        reduce_results: bool = True,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config)
+        self.down_proj = RowParallelLinear(intermediate_size,
+                                           hidden_size,
+                                           bias=False,
+                                           quant_config=quant_config,
+                                           reduce_results=reduce_results)
+        if hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {hidden_act}. "
+                             "Only silu is supported for now.")
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class Qwen2MoeSparseMoeBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+
+        if self.tp_size > config.num_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {config.num_experts}.")
+
+        self.experts = FusedMoE(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)
+
+        self.gate = ReplicatedLinear(config.hidden_size,
+                                     config.num_experts,
+                                     bias=False,
+                                     quant_config=None)
+        if config.shared_expert_intermediate_size > 0:
+            self.shared_expert = Qwen2MoeMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.shared_expert_intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+            )
+        else:
+            self.shared_expert = None
+        self.shared_expert_gate = torch.nn.Linear(config.hidden_size,
+                                                  1,
+                                                  bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        shared_output = None
+        if self.shared_expert is not None:
+            shared_output = self.shared_expert(hidden_states)
+            if self.shared_expert_gate is not None:
+                shared_output = F.sigmoid(
+                    self.shared_expert_gate(hidden_states)) * shared_output
+
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+        final_hidden_states = self.experts(hidden_states=hidden_states,
+                                           router_logits=router_logits)
+        if shared_output is not None:
+            final_hidden_states = final_hidden_states + shared_output
+        if self.tp_size > 1:
+            final_hidden_states = tensor_model_parallel_all_reduce(
+                final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_dim)
+
+
+class Qwen2MoeAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = hidden_size // self.total_num_heads
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=True,
+            quant_config=quant_config,
+        )
+
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = RadixAttention(self.num_heads,
+                                   self.head_dim,
+                                   self.scaling,
+                                   num_kv_heads=self.num_kv_heads,
+                                   layer_id=layer_id)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Qwen2MoeDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.self_attn = Qwen2MoeAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+        )
+
+        # Note: Qwen/Qwen2-57B-A14B-Instruct does not have
+        # `mlp_only_layers` in the config.
+        mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
+                           config.mlp_only_layers)
+        if (layer_id not in mlp_only_layers) and (
+                config.num_experts > 0 and
+            (layer_id + 1) % config.decoder_sparse_step == 0):
+            self.mlp = Qwen2MoeSparseMoeBlock(config=config,
+                                              quant_config=quant_config)
+        else:
+            self.mlp = Qwen2MoeMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+            )
+        self.input_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size,
+                                                eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class Qwen2MoeModel(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList([
+            Qwen2MoeDecoderLayer(config,
+                                 layer_id,
+                                 cache_config,
+                                 quant_config=quant_config)
+            for layer_id in range(config.num_hidden_layers)
+        ])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(positions,
+                                            hidden_states,
+                                            input_metadata,
+                                            residual)
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class Qwen2MoeForCausalLM(nn.Module):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Qwen2MoeModel(config, cache_config, quant_config)
+        self.lm_head = ParallelLMHead(config.vocab_size,
+                                      config.hidden_size,
+                                      quant_config=quant_config)
+        self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata,
+                                   input_embeds)
+        return self.logits_processor(input_ids, hidden_states, self.lm_head.weight,
+                                     input_metadata)
+
+    def compute_logits(self, input_ids: torch.Tensor, hidden_states: torch.Tensor,
+                       input_metadata: InputMetadata) -> torch.Tensor:
+        logits = self.logits_processor(input_ids, hidden_states, self.lm_head.weight,
+                                       input_metadata)
+        return logits
+
+    def sample(
+        self,
+        logits: Optional[torch.Tensor],
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        expert_params_mapping = [
+            # These are the weights for the experts
+            # (param_name, weight_name, expert_id, shard_id)
+            ("experts.w13_weight" if weight_name in ["gate_proj", "up_proj"]
+             else "experts.w2_weight",
+             f"experts.{expert_id}.{weight_name}.weight", expert_id, shard_id)
+            for expert_id in range(self.config.num_experts) for shard_id,
+            weight_name in enumerate(["gate_proj", "down_proj", "up_proj"])
+        ]
+
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+                # Skip non-stacked layers and experts (experts handled below).
+                if weight_name not in name:
+                    continue
+                # We have mlp.experts[0].gate_proj in the checkpoint.
+                # Since we handle the experts below in expert_params_mapping,
+                # we need to skip here BEFORE we update the name, otherwise
+                # name will be updated to mlp.experts[0].gate_up_proj, which
+                # will then be updated below in expert_params_mapping
+                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
+                if "mlp.experts" in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if name not in params_dict:
+                    continue
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, expert_id, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(param,
+                                  loaded_weight,
+                                  weight_name,
+                                  shard_id=shard_id,
+                                  expert_id=expert_id)
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    if name not in params_dict:
+                        continue
+
+                    param = params_dict[name]
+                    weight_loader = getattr(param, "weight_loader",
+                                            default_weight_loader)
+                    weight_loader(param, loaded_weight)
+
+EntryClass = Qwen2MoeForCausalLM

sglang/srt/openai_api_adapter.py

@@ -164,7 +164,7 @@ async def v1_completions(tokenizer_manager, raw_request: Request):
                         logprobs = None
 
                     delta = text[len(stream_buffer) :]
-                    stream_buffer = content["text"]
+                    stream_buffer = stream_buffer + delta
                     choice_data = CompletionResponseStreamChoice(
                         index=0,
                         text=delta,
@@ -323,7 +323,7 @@ async def v1_chat_completions(tokenizer_manager, raw_request: Request):
 
                     text = content["text"]
                     delta = text[len(stream_buffer) :]
-                    stream_buffer = text
+                    stream_buffer = stream_buffer + delta
                     choice_data = ChatCompletionResponseStreamChoice(
                         index=0,
                         delta=DeltaMessage(content=delta),

sglang/srt/openai_protocol.py

@@ -134,7 +134,7 @@ class ChatCompletionRequest(BaseModel):
     logit_bias: Optional[Dict[str, float]] = None
     logprobs: Optional[bool] = False
     top_logprobs: Optional[int] = None
-    max_tokens: Optional[int] = None
+    max_tokens: Optional[int] = 16
     n: Optional[int] = 1
     presence_penalty: Optional[float] = 0.0
     response_format: Optional[ResponseFormat] = None

sglang/srt/server.py

@@ -51,13 +51,12 @@ from sglang.srt.utils import (
     allocate_init_ports,
     assert_pkg_version,
     enable_show_time_cost,
-    send_addrs_to_rank_0,
     receive_addrs,
+    send_addrs_to_rank_0,
     start_rpyc_service_process,
 )
 from sglang.utils import get_exception_traceback
 
-
 logger = logging.getLogger(__name__)
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
@@ -152,9 +151,13 @@ def launch_server(server_args: ServerArgs, pipe_finish_writer, model_overide_arg
     if server_args.disable_disk_cache:
         disable_cache()
     if not server_args.disable_flashinfer:
-        assert_pkg_version("flashinfer", "0.0.8", "Please uninstall the old version and "
-                           "reinstall the latest version by following the instructions "
-                           "at https://docs.flashinfer.ai/installation.html.")
+        assert_pkg_version(
+            "flashinfer",
+            "0.0.8",
+            "Please uninstall the old version and "
+            "reinstall the latest version by following the instructions "
+            "at https://docs.flashinfer.ai/installation.html.",
+        )
     if server_args.chat_template:
         # TODO: replace this with huggingface transformers template
         load_chat_template_for_openai_api(server_args.chat_template)
@@ -176,7 +179,9 @@ def launch_server(server_args: ServerArgs, pipe_finish_writer, model_overide_arg
             ModelPortArgs(
                 nccl_port=ports[3 + i * (tp_size_local + 1)],
                 model_tp_ips=[None] * tp_size_local,
-                model_tp_ports=ports[3 + i * (tp_size_local + 1) + 1 : 3 + (i + 1) * (tp_size_local + 1)],
+                model_tp_ports=ports[
+                    3 + i * (tp_size_local + 1) + 1 : 3 + (i + 1) * (tp_size_local + 1)
+                ],
             )
         )
     port_args = PortArgs(
@@ -194,9 +199,13 @@ def launch_server(server_args: ServerArgs, pipe_finish_writer, model_overide_arg
         else:
             receive_addrs(model_port_args[0], server_args)
         for i in range(tp_size_local):
-            start_rpyc_service_process(ModelTpService, model_port_args[0].model_tp_ports[i])
+            start_rpyc_service_process(
+                ModelTpService, model_port_args[0].model_tp_ports[i]
+            )
         if server_args.node_rank != 0:
-            logger.info(f"[node_rank={server_args.node_rank}]: Listen for connections...")
+            logger.info(
+                f"[node_rank={server_args.node_rank}]: Listen for connections..."
+            )
             while True:
                 pass
 

sglang/srt/server_args.py

@@ -55,6 +55,7 @@ class ServerArgs:
     disable_regex_jump_forward: bool = False
     disable_disk_cache: bool = False
     attention_reduce_in_fp32: bool = False
+    enable_p2p_check: bool = False
 
     # Distributed args
     nccl_init_addr: Optional[str] = None
@@ -137,17 +138,16 @@ class ServerArgs:
             "--dtype",
             type=str,
             default=ServerArgs.dtype,
-            choices=[
-                "auto", "half", "float16", "bfloat16", "float", "float32"
-            ],
-            help='Data type for model weights and activations.\n\n'
+            choices=["auto", "half", "float16", "bfloat16", "float", "float32"],
+            help="Data type for model weights and activations.\n\n"
             '* "auto" will use FP16 precision for FP32 and FP16 models, and '
-            'BF16 precision for BF16 models.\n'
+            "BF16 precision for BF16 models.\n"
             '* "half" for FP16. Recommended for AWQ quantization.\n'
             '* "float16" is the same as "half".\n'
             '* "bfloat16" for a balance between precision and range.\n'
             '* "float" is shorthand for FP32 precision.\n'
-            '* "float32" for FP32 precision.')
+            '* "float32" for FP32 precision.',
+        )
         parser.add_argument(
             "--trust-remote-code",
             action="store_true",
@@ -271,19 +271,12 @@ class ServerArgs:
         parser.add_argument(
             "--nccl-init-addr",
             type=str,
-            help="The nccl init address of multi-node server."
-        )
-        parser.add_argument(
-            "--nnodes",
-            type=int,
-            default=1,
-            help="The number of nodes."
+            help="The nccl init address of multi-node server.",
         )
         parser.add_argument(
-            "--node-rank",
-            type=int,
-            help="The node rank."
+            "--nnodes", type=int, default=1, help="The number of nodes."
         )
+        parser.add_argument("--node-rank", type=int, help="The node rank.")
 
         # Optimization/debug options
         parser.add_argument(
@@ -312,6 +305,11 @@ class ServerArgs:
             help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16."
             "This only affects Triton attention kernels",
         )
+        parser.add_argument(
+            "--enable-p2p-check",
+            action="store_true",
+            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
+        )
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):