sglang 0.4.5__py3-none-any.whl → 0.4.5.post2__py3-none-any.whl

sglang/srt/models/qwen3_moe.py

@@ -0,0 +1,423 @@
+# Adapted from qwen2_moe.py
+
+# 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.
+# ==============================================================================
+
+
+"""Inference-only Qwen3MoE model compatible with HuggingFace weights."""
+
+from functools import partial
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    split_tensor_along_last_dim,
+    tensor_model_parallel_all_gather,
+    tensor_model_parallel_all_reduce,
+)
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.logits_processor import LogitsProcessor
+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
+from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+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
+from sglang.srt.models.qwen2_moe import Qwen2MoeModel
+from sglang.srt.utils import add_prefix
+
+Qwen3MoeConfig = None
+
+
+class Qwen3MoeSparseMoeBlock(nn.Module):
+    def __init__(
+        self,
+        config: Qwen3MoeConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        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,
+            prefix=add_prefix("experts", prefix),
+        )
+
+        self.gate = ReplicatedLinear(
+            config.hidden_size,
+            config.num_experts,
+            bias=False,
+            quant_config=None,
+            prefix=add_prefix("gate", prefix),
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        # 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 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 Qwen3MoeAttention(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,
+        head_dim: Optional[int] = None,
+        rms_norm_eps: float = 1e-06,
+        attention_bias: bool = False,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % self.tp_size == 0
+        self.num_heads = self.total_num_heads // self.tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= self.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 % self.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 self.tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
+        self.head_dim = head_dim or 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.tp_rank = get_tensor_model_parallel_rank()
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
+        )
+
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        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,
+            prefix=add_prefix("attn", prefix),
+        )
+
+        self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+        self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+
+    def _apply_qk_norm(
+        self, q: torch.Tensor, k: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        q_by_head = q.reshape(-1, self.head_dim)
+        q_by_head = self.q_norm(q_by_head)
+        q = q_by_head.view(q.shape)
+        k_by_head = k.reshape(-1, self.head_dim)
+        k_by_head = self.k_norm(k_by_head)
+        k = k_by_head.view(k.shape)
+        return q, k
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> 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._apply_qk_norm(q, k)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Qwen3MoeDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: Qwen3MoeConfig,
+        layer_id: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> 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)
+        head_dim = getattr(
+            config, "head_dim", config.hidden_size // config.num_attention_heads
+        )
+        rms_norm_eps = config.rms_norm_eps
+        attention_bias = config.attention_bias
+        self.self_attn = Qwen3MoeAttention(
+            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,
+            head_dim=head_dim,
+            rms_norm_eps=rms_norm_eps,
+            attention_bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("self_attn", prefix),
+        )
+
+        # 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 = Qwen3MoeSparseMoeBlock(
+                config=config,
+                quant_config=quant_config,
+                prefix=add_prefix("mlp", prefix),
+            )
+        else:
+            self.mlp = Qwen3MoeMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                prefix=add_prefix("mlp", prefix),
+            )
+        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,
+        forward_batch: ForwardBatch,
+        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,
+            forward_batch=forward_batch,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class Qwen3MoeModel(Qwen2MoeModel):
+    def __init__(
+        self,
+        config: Qwen3MoeConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(
+            config=config,
+            quant_config=quant_config,
+            prefix=prefix,
+            decoder_layer_type=Qwen3MoeDecoderLayer,
+        )
+
+
+class Qwen3MoeForCausalLM(nn.Module):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(
+        self,
+        config: Qwen3MoeConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Qwen3MoeModel(
+            config, quant_config, prefix=add_prefix("model", prefix)
+        )
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=add_prefix("lm_head", prefix),
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
+
+    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 = FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=self.config.num_experts,
+        )
+
+        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,
+                        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 = Qwen3MoeForCausalLM

sglang/srt/models/stablelm.py

@@ -149,6 +149,7 @@ class StablelmAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_key_value_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/xverse.py

@@ -153,6 +153,7 @@ class XverseAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/xverse_moe.py

@@ -252,6 +252,7 @@ class XverseAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/openai_api/adapter.py

@@ -983,6 +983,8 @@ def v1_chat_generate_request(
                     ):
                         encoded = encoded[1:]
                     prompt_ids += encoded
+                if tokenizer_manager.model_config.is_multimodal:
+                    prompt = tokenizer_manager.tokenizer.decode(prompt_ids)
                 stop = request.stop
                 image_data = None
                 audio_data = None
@@ -993,7 +995,8 @@ def v1_chat_generate_request(
                 image_data = conv.image_data
                 audio_data = conv.audio_data
                 modalities = conv.modalities
-                stop = conv.stop_str or []
+                stop = conv.stop_str or [] if not request.ignore_eos else []
+
                 if request.stop:
                     if isinstance(request.stop, str):
                         stop.append(request.stop)

sglang/srt/patch_torch.py

@@ -14,6 +14,7 @@
 from typing import Callable, Union
 
 import torch
+from packaging import version
 from torch.multiprocessing import reductions
 
 
@@ -69,3 +70,13 @@ def _device_from_maybe_uuid(device_maybe_uuid: Union[int, str]) -> int:
 
 def _modify_tuple(t, index: int, modifier: Callable):
     return *t[:index], modifier(t[index]), *t[index + 1 :]
+
+
+def monkey_patch_torch_compile():
+    if version.parse(torch.__version__) < version.parse("2.8.0"):
+        # These things are cacheable by torch.compile. torch.compile just doesn't know it.
+        # This was fixed in PyTorch 2.8, but until then, we monkey patch.
+        import torch._higher_order_ops.auto_functionalize as af
+
+        af.auto_functionalized_v2._cacheable = True
+        af.auto_functionalized._cacheable = True

sglang/srt/reasoning_parser.py

@@ -1,4 +1,3 @@
-import re
 from typing import Dict, Tuple
 
 

sglang/srt/sampling/sampling_batch_info.py

@@ -10,12 +10,11 @@ import torch
 import sglang.srt.sampling.penaltylib as penaltylib
 from sglang.srt.sampling.custom_logit_processor import CustomLogitProcessor
 
-logger = logging.getLogger(__name__)
-
-
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import ScheduleBatch
 
+logger = logging.getLogger(__name__)
+
 
 @dataclasses.dataclass
 class SamplingBatchInfo: