sglang 0.5.2rc2__py3-none-any.whl → 0.5.3__py3-none-any.whl

sglang/srt/models/qwen2_5_vl.py

@@ -40,7 +40,6 @@ from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
     Qwen2_5_VisionRotaryEmbedding,
 )
 
-from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -61,6 +60,7 @@ 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 import Qwen2Model
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 
@@ -113,12 +113,13 @@ class Qwen2_5_VisionBlock(nn.Module):
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
         num_dummy_heads: int = 0,
+        rms_norm_eps: float = 1e-6,
     ) -> None:
         super().__init__()
         if norm_layer is None:
             norm_layer = partial(nn.LayerNorm, eps=1e-6)
-        self.norm1 = RMSNorm(dim, eps=1e-6)
-        self.norm2 = RMSNorm(dim, eps=1e-6)
+        self.norm1 = RMSNorm(dim, eps=rms_norm_eps)
+        self.norm2 = RMSNorm(dim, eps=rms_norm_eps)
 
         if attn_implementation is None:
             softmax_in_single_precision = False
@@ -264,7 +265,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
         self.fullatt_block_indexes = vision_config.fullatt_block_indexes
         self.window_size = vision_config.window_size
         self.patch_size = vision_config.patch_size
-        mlp_hidden_size: int = vision_config.intermediate_size
+        mlp_hidden_size: int = ((vision_config.intermediate_size + 7) // 8) * 8
         self.patch_embed = Qwen2_5_VisionPatchEmbed(
             patch_size=patch_size,
             temporal_patch_size=temporal_patch_size,
@@ -517,6 +518,9 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
+        # For EAGLE3 support
+        self.capture_aux_hidden_states = False
+
     def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
         pattern = MultiModalityDataPaddingPatternMultimodalTokens()
         return pattern.pad_input_tokens(input_ids, mm_inputs)
@@ -587,9 +591,13 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
             positions=positions,
         )
 
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
         if not get_embedding:
             return self.logits_processor(
-                input_ids, hidden_states, self.lm_head, forward_batch
+                input_ids, hidden_states, self.lm_head, forward_batch, aux_hidden_states
             )
         else:
             return self.pooler(hidden_states, forward_batch)
@@ -643,5 +651,21 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
+        self.capture_aux_hidden_states = True
+        self.model.capture_aux_hidden_states = True
+        if layer_ids is None:
+            num_layers = self.config.num_hidden_layers
+            self.model.layers_to_capture = [
+                2,
+                num_layers // 2,
+                num_layers - 3,
+            ]  # Specific layers for EAGLE3 support
+        else:
+            self.model.layers_to_capture = [val + 1 for val in layer_ids]
+
 
 EntryClass = [Qwen2_5_VLForConditionalGeneration]

sglang/srt/models/qwen2_audio.py

@@ -39,7 +39,6 @@ from transformers.models.qwen2_audio.modeling_qwen2_audio import (
     Qwen2AudioMultiModalProjector,
 )
 
-from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.activation import QuickGELU
 from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
@@ -61,6 +60,7 @@ 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 import Qwen2ForCausalLM
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 

sglang/srt/models/qwen2_moe.py

@@ -25,12 +25,14 @@ from torch import nn
 from transformers import PretrainedConfig
 
 from sglang.srt.distributed import (
+    get_moe_expert_parallel_world_size,
     get_pp_group,
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
 from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
+from sglang.srt.eplb.expert_location_dispatch import ExpertLocationDispatchInfo
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.communicator import (
     LayerCommunicator,
@@ -50,6 +52,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe import get_moe_a2a_backend
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
@@ -62,13 +65,16 @@ 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.cuda_graph_runner import get_is_capture_mode
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.two_batch_overlap import model_forward_maybe_tbo
-from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils import add_prefix, is_cuda, make_layers
 
 logger = logging.getLogger(__name__)
 
+_is_cuda = is_cuda()
+
 
 class Qwen2MoeMLP(nn.Module):
     def __init__(
@@ -79,6 +85,8 @@ class Qwen2MoeMLP(nn.Module):
         quant_config: Optional[QuantizationConfig] = None,
         reduce_results: bool = True,
         prefix: str = "",
+        tp_rank: Optional[int] = None,
+        tp_size: Optional[int] = None,
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
@@ -87,6 +95,8 @@ class Qwen2MoeMLP(nn.Module):
             bias=False,
             quant_config=quant_config,
             prefix=add_prefix("gate_up_proj", prefix),
+            tp_rank=tp_rank,
+            tp_size=tp_size,
         )
         self.down_proj = RowParallelLinear(
             intermediate_size,
@@ -95,6 +105,8 @@ class Qwen2MoeMLP(nn.Module):
             quant_config=quant_config,
             reduce_results=reduce_results,
             prefix=add_prefix("down_proj", prefix),
+            tp_rank=tp_rank,
+            tp_size=tp_size,
         )
         if hidden_act != "silu":
             raise ValueError(
@@ -122,11 +134,13 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         layer_id: int,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        alt_stream: Optional[torch.cuda.Stream] = None,
         prefix: str = "",
     ):
         super().__init__()
         self.tp_size = get_tensor_model_parallel_world_size()
         self.layer_id = layer_id
+        self.alt_stream = alt_stream
         if self.tp_size > config.num_experts:
             raise ValueError(
                 f"Tensor parallel size {self.tp_size} is greater than "
@@ -138,10 +152,11 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
             renormalize=config.norm_topk_prob,
         )
 
-        self.experts = get_moe_impl_class()(
+        self.experts = get_moe_impl_class(quant_config)(
             layer_id=self.layer_id,
             top_k=config.num_experts_per_tok,
-            num_experts=config.num_experts,
+            num_experts=config.num_experts
+            + global_server_args_dict["ep_num_redundant_experts"],
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
             quant_config=quant_config,
@@ -163,19 +178,32 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
                 quant_config=quant_config,
                 reduce_results=False,
                 prefix=add_prefix("shared_expert", prefix),
+                **(
+                    dict(tp_rank=0, tp_size=1)
+                    if get_moe_a2a_backend().is_deepep()
+                    else {}
+                ),
             )
         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,
-        forward_batch: Optional[ForwardBatch] = None,
-        use_reduce_scatter: bool = False,
-    ) -> torch.Tensor:
-        num_tokens, hidden_dim = hidden_states.shape
-        hidden_states = hidden_states.view(-1, hidden_dim)
+        if get_moe_a2a_backend().is_deepep():
+            # TODO: we will support tp < ep in the future
+            self.ep_size = get_moe_expert_parallel_world_size()
+            self.num_experts = (
+                config.num_experts + global_server_args_dict["ep_num_redundant_experts"]
+            )
+            self.top_k = config.num_experts_per_tok
+
+    def get_moe_weights(self):
+        return [
+            x.data
+            for name, x in self.experts.named_parameters()
+            if name not in ["correction_bias"]
+        ]
+
+    def _forward_shared_experts(self, hidden_states: torch.Tensor):
         shared_output = None
         if self.shared_expert is not None:
             shared_output = self.shared_expert(hidden_states)
@@ -183,11 +211,85 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
                 shared_output = (
                     F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_output
                 )
+        return shared_output
+
+    def _forward_deepep(self, hidden_states: torch.Tensor, forward_batch: ForwardBatch):
+        shared_output = None
+        if hidden_states.shape[0] > 0:
+            # router_logits: (num_tokens, n_experts)
+            router_logits, _ = self.gate(hidden_states)
+            shared_output = self._forward_shared_experts(hidden_states)
+            topk_weights, topk_idx, _ = self.topk(
+                hidden_states,
+                router_logits,
+                num_token_non_padded=forward_batch.num_token_non_padded,
+                expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
+                    layer_id=self.layer_id,
+                ),
+            )
+        else:
+            topk_weights, topk_idx, _ = self.topk.empty_topk_output(
+                hidden_states.device
+            )
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states,
+            topk_idx=topk_idx,
+            topk_weights=topk_weights,
+            forward_batch=forward_batch,
+        )
+
+        if shared_output is not None:
+            final_hidden_states.add_(shared_output)
 
+        return final_hidden_states
+
+    def _forward_router_experts(self, hidden_states: torch.Tensor):
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
         topk_output = self.topk(hidden_states, router_logits)
-        final_hidden_states = self.experts(hidden_states, topk_output)
+        return self.experts(hidden_states, topk_output)
+
+    def forward_normal_dual_stream(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        current_stream = torch.cuda.current_stream()
+        self.alt_stream.wait_stream(current_stream)
+        shared_output = self._forward_shared_experts(hidden_states.clone())
+
+        with torch.cuda.stream(self.alt_stream):
+            router_output = self._forward_router_experts(hidden_states)
+
+        current_stream.wait_stream(self.alt_stream)
+
+        return router_output, shared_output
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: Optional[ForwardBatch] = None,
+        use_reduce_scatter: bool = False,
+    ) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        if get_moe_a2a_backend().is_deepep():
+            return self._forward_deepep(hidden_states, forward_batch)
+
+        DUAL_STREAM_TOKEN_THRESHOLD = 1024
+        if (
+            self.alt_stream is not None
+            and hidden_states.shape[0] > 0
+            and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
+            and get_is_capture_mode()
+        ):
+            final_hidden_states, shared_output = self.forward_normal_dual_stream(
+                hidden_states
+            )
+        else:
+            shared_output = self._forward_shared_experts(hidden_states)
+            final_hidden_states = self._forward_router_experts(hidden_states)
+
         if shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output
         if self.tp_size > 1 and not use_reduce_scatter:
@@ -346,6 +448,7 @@ class Qwen2MoeDecoderLayer(nn.Module):
                 layer_id=layer_id,
                 config=config,
                 quant_config=quant_config,
+                alt_stream=alt_stream,
                 prefix=add_prefix("mlp", prefix),
             )
         else:
@@ -528,8 +631,12 @@ class Qwen2MoeForCausalLM(nn.Module):
         self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
+        alt_stream = torch.cuda.Stream() if _is_cuda else None
         self.model = Qwen2MoeModel(
-            config, quant_config, prefix=add_prefix("model", prefix)
+            config,
+            quant_config,
+            prefix=add_prefix("model", prefix),
+            alt_stream=alt_stream,
         )
         self.lm_head = ParallelLMHead(
             config.vocab_size,

sglang/srt/models/qwen2_vl.py

@@ -33,7 +33,6 @@ from einops import rearrange
 from transformers import Qwen2VLConfig
 from transformers.models.qwen2_vl.configuration_qwen2_vl import Qwen2VLVisionConfig
 
-from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.activation import QuickGELU
 from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
@@ -50,6 +49,7 @@ 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 import Qwen2Model
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 

sglang/srt/models/qwen3.py

@@ -1,6 +1,5 @@
 # Adapted from qwen2.py
 import logging
-from functools import partial
 from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
@@ -30,12 +29,19 @@ from sglang.srt.model_loader.weight_utils import (
 )
 from sglang.srt.models.qwen2 import Qwen2MLP as Qwen3MLP
 from sglang.srt.models.qwen2 import Qwen2Model
-from sglang.srt.utils import add_prefix, is_cuda
+from sglang.srt.utils import (
+    add_prefix,
+    get_cmo_stream,
+    is_cuda,
+    is_npu,
+    wait_cmo_stream,
+)
 
 Qwen3Config = None
 
 logger = logging.getLogger(__name__)
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
 
 class Qwen3Attention(nn.Module):
@@ -235,9 +241,18 @@ class Qwen3DecoderLayer(nn.Module):
 
         # Fully Connected
         hidden_states, residual = self.layer_communicator.prepare_mlp(
-            hidden_states, residual, forward_batch
+            hidden_states,
+            residual,
+            forward_batch,
+            cache=(
+                [self.mlp.gate_up_proj.weight, self.mlp.down_proj.weight]
+                if _is_npu
+                else None
+            ),
         )
         hidden_states = self.mlp(hidden_states)
+        if _is_npu and get_cmo_stream():
+            wait_cmo_stream()
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             hidden_states, residual, forward_batch
         )

sglang/srt/models/qwen3_moe.py

@@ -51,7 +51,7 @@ from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
 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.rotary_embedding import MRotaryEmbedding, get_rope
 from sglang.srt.layers.utils import get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
 from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -60,6 +60,10 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTe
 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.models.utils import (
+    create_fused_set_kv_buffer_arg,
+    enable_fused_set_kv_buffer,
+)
 from sglang.srt.utils import (
     add_prefix,
     is_cuda,
@@ -98,7 +102,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             use_grouped_topk=False,
         )
 
-        self.experts = get_moe_impl_class()(
+        self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.num_experts
             + global_server_args_dict["ep_num_redundant_experts"],
             top_k=config.num_experts_per_tok,
@@ -354,6 +358,10 @@ class Qwen3MoeAttention(nn.Module):
             rope_scaling=rope_scaling,
             dual_chunk_attention_config=dual_chunk_attention_config,
         )
+        self.compatible_with_fused_kv_buffer = (
+            False if isinstance(self.rotary_emb, MRotaryEmbedding) else True
+        )
+
         self.attn = RadixAttention(
             self.num_heads,
             self.head_dim,
@@ -412,7 +420,21 @@ class Qwen3MoeAttention(nn.Module):
         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)
+        q, k = self.rotary_emb(
+            positions,
+            q,
+            k,
+            fused_set_kv_buffer_arg=(
+                create_fused_set_kv_buffer_arg(
+                    value=v,
+                    layer=self.attn,
+                    forward_batch=forward_batch,
+                )
+                if enable_fused_set_kv_buffer(forward_batch)
+                and self.compatible_with_fused_kv_buffer
+                else None
+            ),
+        )
         inner_state = q, k, v, forward_batch
         return None, forward_batch, inner_state
 
@@ -420,7 +442,13 @@ class Qwen3MoeAttention(nn.Module):
         hidden_states, forward_batch, inner_state = intermediate_state
         if inner_state is None:
             return hidden_states
-        attn_output = self.attn(*inner_state)
+        attn_output = self.attn(
+            *inner_state,
+            save_kv_cache=not (
+                enable_fused_set_kv_buffer(forward_batch)
+                and self.compatible_with_fused_kv_buffer
+            ),
+        )
         output, _ = self.o_proj(attn_output)
         return output