sglang 0.4.8.post1__py3-none-any.whl → 0.4.9.post1__py3-none-any.whl

sglang/srt/models/qwen3.py

@@ -2,7 +2,7 @@
 
 import logging
 from functools import partial
-from typing import Any, Dict, Iterable, Optional, Tuple
+from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -11,9 +11,9 @@ from sglang.srt.distributed import (
     get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
-    split_tensor_along_last_dim,
-    tensor_model_parallel_all_gather,
 )
+from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
+from sglang.srt.layers.dp_attention import get_attention_tp_rank, get_attention_tp_size
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
@@ -23,15 +23,17 @@ 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
+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.models.qwen2 import Qwen2MLP as Qwen3MLP
 from sglang.srt.models.qwen2 import Qwen2Model
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, is_cuda
 
 Qwen3Config = None
 
 logger = logging.getLogger(__name__)
+_is_cuda = is_cuda()
 
 
 class Qwen3Attention(nn.Module):
@@ -49,23 +51,27 @@ class Qwen3Attention(nn.Module):
         rms_norm_eps: float = None,
         attention_bias: bool = False,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> 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
+        attn_tp_rank = get_attention_tp_rank()
+        attn_tp_size = get_attention_tp_size()
+
+        assert self.total_num_heads % attn_tp_size == 0
+        self.num_heads = self.total_num_heads // attn_tp_size
         self.total_num_kv_heads = num_kv_heads
-        if self.total_num_kv_heads >= self.tp_size:
+        if self.total_num_kv_heads >= attn_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
+            assert self.total_num_kv_heads % attn_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)
+            assert attn_tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // attn_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
@@ -84,6 +90,8 @@ class Qwen3Attention(nn.Module):
             self.total_num_kv_heads,
             bias=attention_bias,
             quant_config=quant_config,
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
             prefix=add_prefix("qkv_proj", prefix),
         )
         self.o_proj = RowParallelLinear(
@@ -91,6 +99,9 @@ class Qwen3Attention(nn.Module):
             hidden_size,
             bias=attention_bias,
             quant_config=quant_config,
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
+            reduce_results=False,
             prefix=add_prefix("o_proj", prefix),
         )
 
@@ -109,15 +120,27 @@ class Qwen3Attention(nn.Module):
             layer_id=layer_id,
             prefix=add_prefix("attn", prefix),
         )
+        self.alt_stream = alt_stream
 
     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)
+        # overlap qk norm
+        if self.alt_stream is not None and get_is_capture_mode():
+            current_stream = torch.cuda.current_stream()
+            self.alt_stream.wait_stream(current_stream)
+            q_by_head = q.reshape(-1, self.head_dim)
+            q_by_head = self.q_norm(q_by_head)
+            with torch.cuda.stream(self.alt_stream):
+                k_by_head = k.reshape(-1, self.head_dim)
+                k_by_head = self.k_norm(k_by_head)
+            current_stream.wait_stream(self.alt_stream)
+        else:
+            q_by_head = q.reshape(-1, self.head_dim)
+            q_by_head = self.q_norm(q_by_head)
+            k_by_head = k.reshape(-1, self.head_dim)
+            k_by_head = self.k_norm(k_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
 
@@ -143,6 +166,7 @@ class Qwen3DecoderLayer(nn.Module):
         layer_id: int = 0,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -163,6 +187,7 @@ class Qwen3DecoderLayer(nn.Module):
             rms_norm_eps=config.rms_norm_eps,
             attention_bias=config.attention_bias,
             prefix=add_prefix("self_attn", prefix),
+            alt_stream=alt_stream,
         )
         self.mlp = Qwen3MLP(
             hidden_size=self.hidden_size,
@@ -176,6 +201,18 @@ class Qwen3DecoderLayer(nn.Module):
             config.hidden_size, eps=config.rms_norm_eps
         )
 
+        self.layer_scatter_modes = LayerScatterModes.init_new(
+            layer_id=layer_id,
+            num_layers=config.num_hidden_layers,
+            is_layer_sparse=False,
+            is_previous_layer_sparse=False,
+        )
+        self.layer_communicator = LayerCommunicator(
+            layer_scatter_modes=self.layer_scatter_modes,
+            input_layernorm=self.input_layernorm,
+            post_attention_layernorm=self.post_attention_layernorm,
+        )
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -184,20 +221,24 @@ class Qwen3DecoderLayer(nn.Module):
         residual: Optional[torch.Tensor],
     ) -> Tuple[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,
+        hidden_states, residual = self.layer_communicator.prepare_attn(
+            hidden_states, residual, forward_batch
         )
+        if hidden_states.shape[0] != 0:
+            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, residual = self.layer_communicator.prepare_mlp(
+            hidden_states, residual, forward_batch
+        )
         hidden_states = self.mlp(hidden_states)
+        hidden_states, residual = self.layer_communicator.postprocess_layer(
+            hidden_states, residual, forward_batch
+        )
         return hidden_states, residual
 
 
@@ -208,11 +249,13 @@ class Qwen3Model(Qwen2Model):
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
+        alt_stream = torch.cuda.Stream() if _is_cuda else None
         super().__init__(
             config=config,
             quant_config=quant_config,
             prefix=prefix,
             decoder_layer_type=Qwen3DecoderLayer,
+            alt_stream=alt_stream,
         )
 
 
@@ -282,6 +325,9 @@ class Qwen3ForCausalLM(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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.model.get_input_embeddings(input_ids)
 
@@ -303,10 +349,18 @@ class Qwen3ForCausalLM(nn.Module):
             pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            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
+                    input_ids,
+                    hidden_states,
+                    self.lm_head,
+                    forward_batch,
+                    aux_hidden_states,
                 )
             else:
                 return self.pooler(hidden_states, forward_batch)
@@ -404,5 +458,20 @@ class Qwen3ForCausalLM(nn.Module):
     def load_kv_cache_scales(self, quantization_param_path: str) -> None:
         self.model.load_kv_cache_scales(quantization_param_path)
 
+    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
+        if not self.pp_group.is_last_rank:
+            return
+
+        self.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 = Qwen3ForCausalLM

sglang/srt/models/qwen3_moe.py

@@ -18,7 +18,7 @@
 """Inference-only Qwen3MoE model compatible with HuggingFace weights."""
 
 import logging
-from typing import Any, Dict, Iterable, Optional, Tuple
+from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -32,6 +32,9 @@ from sglang.srt.distributed import (
     tensor_model_parallel_all_gather,
     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, LayerScatterModes
 from sglang.srt.layers.dp_attention import (
@@ -63,12 +66,8 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.expert_distribution import (
-    get_global_expert_distribution_recorder,
-)
-from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
-from sglang.srt.managers.expert_location_dispatch import ExpertLocationDispatchInfo
 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,
     ForwardMode,
@@ -78,11 +77,12 @@ 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.two_batch_overlap import MaybeTboDeepEPDispatcher
-from sglang.srt.utils import DeepEPMode, add_prefix, is_non_idle_and_non_empty
+from sglang.srt.utils import DeepEPMode, add_prefix, is_cuda, is_non_idle_and_non_empty
 
 Qwen3MoeConfig = None
 
 logger = logging.getLogger(__name__)
+_is_cuda = is_cuda()
 
 
 class Qwen3MoeSparseMoeBlock(nn.Module):
@@ -117,6 +117,15 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                 if global_server_args_dict["enable_deepep_moe"]
                 else {}
             ),
+            # Additional args for FusedMoE
+            **(
+                dict(
+                    enable_flashinfer_moe=True,
+                    enable_ep_moe=global_server_args_dict["enable_ep_moe"],
+                )
+                if global_server_args_dict["enable_flashinfer_moe"]
+                else {}
+            ),
         )
 
         self.gate = ReplicatedLinear(
@@ -220,7 +229,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                 hidden_states=hidden_states,
                 topk_idx=topk_idx,
                 topk_weights=topk_weights,
-                forward_mode=forward_mode,
+                forward_batch=forward_batch,
             )
         final_hidden_states = self.experts(
             hidden_states=hidden_states,
@@ -231,14 +240,14 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             masked_m=masked_m,
             expected_m=expected_m,
             num_recv_tokens_per_expert=num_recv_tokens_per_expert,
-            forward_mode=forward_mode,
+            forward_batch=forward_batch,
         )
         if self.ep_size > 1:
             final_hidden_states = self.deepep_dispatcher.combine(
                 hidden_states=final_hidden_states,
                 topk_idx=topk_idx,
                 topk_weights=topk_weights,
-                forward_mode=forward_mode,
+                forward_batch=forward_batch,
             )
         return final_hidden_states
 
@@ -284,7 +293,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                 hidden_states=state.pop("hidden_states_mlp_input"),
                 topk_idx=state.pop("topk_idx_local"),
                 topk_weights=state.pop("topk_weights_local"),
-                forward_mode=state.forward_batch.forward_mode,
+                forward_batch=state.forward_batch,
                 tbo_subbatch_index=state.get("tbo_subbatch_index"),
             )
 
@@ -316,7 +325,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             masked_m=state.pop("masked_m"),
             expected_m=state.pop("expected_m"),
             num_recv_tokens_per_expert=state.pop("num_recv_tokens_per_expert"),
-            forward_mode=state.forward_batch.forward_mode,
+            forward_batch=state.forward_batch,
         )
 
     def op_combine_a(self, state):
@@ -325,7 +334,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                 hidden_states=state.pop("hidden_states_experts_output"),
                 topk_idx=state.pop("topk_idx_dispatched"),
                 topk_weights=state.pop("topk_weights_dispatched"),
-                forward_mode=state.forward_batch.forward_mode,
+                forward_batch=state.forward_batch,
                 tbo_subbatch_index=state.get("tbo_subbatch_index"),
             )
 
@@ -354,6 +363,7 @@ class Qwen3MoeAttention(nn.Module):
         attention_bias: bool = False,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -423,15 +433,27 @@ class Qwen3MoeAttention(nn.Module):
 
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+        self.alt_stream = alt_stream
 
     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)
+        # overlap qk norm
+        if self.alt_stream is not None and get_is_capture_mode():
+            current_stream = torch.cuda.current_stream()
+            self.alt_stream.wait_stream(current_stream)
+            q_by_head = q.reshape(-1, self.head_dim)
+            q_by_head = self.q_norm(q_by_head)
+            with torch.cuda.stream(self.alt_stream):
+                k_by_head = k.reshape(-1, self.head_dim)
+                k_by_head = self.k_norm(k_by_head)
+            current_stream.wait_stream(self.alt_stream)
+        else:
+            q_by_head = q.reshape(-1, self.head_dim)
+            q_by_head = self.q_norm(q_by_head)
+            k_by_head = k.reshape(-1, self.head_dim)
+            k_by_head = self.k_norm(k_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
 
@@ -491,6 +513,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
         layer_id: int,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -516,6 +539,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
             attention_bias=attention_bias,
             quant_config=quant_config,
             prefix=add_prefix("self_attn", prefix),
+            alt_stream=alt_stream,
         )
 
         self.layer_id = layer_id
@@ -623,9 +647,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
 
     def op_mlp(self, state):
         hidden_states = state.pop("hidden_states_mlp_input")
-        state.hidden_states_mlp_output = self.mlp(
-            hidden_states, state.forward_batch.forward_mode
-        )
+        state.hidden_states_mlp_output = self.mlp(hidden_states, state.forward_batch)
 
     def op_comm_postprocess_layer(self, state):
         hidden_states, residual = self.layer_communicator.postprocess_layer(
@@ -659,11 +681,13 @@ class Qwen3MoeModel(Qwen2MoeModel):
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
+        alt_stream = torch.cuda.Stream() if _is_cuda else None
         super().__init__(
             config=config,
             quant_config=quant_config,
             prefix=prefix,
             decoder_layer_type=Qwen3MoeDecoderLayer,
+            alt_stream=alt_stream,
         )
 
 
@@ -691,6 +715,7 @@ class Qwen3MoeForCausalLM(nn.Module):
             use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
         )
         self.logits_processor = LogitsProcessor(config)
+        self.capture_aux_hidden_states = False
 
     @torch.no_grad()
     def forward(
@@ -709,9 +734,13 @@ class Qwen3MoeForCausalLM(nn.Module):
             pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
         if self.pp_group.is_last_rank:
             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 hidden_states
@@ -724,6 +753,24 @@ class Qwen3MoeForCausalLM(nn.Module):
     def end_layer(self):
         return self.model.end_layer
 
+    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):
+        if not self.pp_group.is_last_rank:
+            return
+
+        self.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]
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)

sglang/srt/models/vila.py

@@ -270,15 +270,10 @@ class VILAForConditionalGeneration(nn.Module):
                 weight_loader(param, loaded_weight)
 
     def pad_input_ids(
-        self,
-        input_ids: List[int],
-        image_inputs: MultimodalInputs,
+        self, input_ids: List[int], mm_inputs: MultimodalInputs
     ) -> List[int]:
-        pattern = MultiModalityDataPaddingPatternMultimodalTokens(
-            token_ids=[self.config.image_token_id],
-        )
-
-        return pattern.pad_input_tokens(input_ids, image_inputs)
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
 
     ##### BEGIN COPY modeling_vila.py #####
 

sglang/srt/{mm_utils.py → multimodal/mm_utils.py}

@@ -28,12 +28,12 @@ LLaVA-Onevision : https://arxiv.org/pdf/2408.03326
 
 """
 import ast
-import base64
 import math
 import re
 from io import BytesIO
 
 import numpy as np
+import pybase64
 from PIL import Image
 
 from sglang.srt.utils import flatten_nested_list
@@ -252,7 +252,7 @@ def process_anyres_image(image, processor, grid_pinpoints):
 
 
 def load_image_from_base64(image):
-    return Image.open(BytesIO(base64.b64decode(image)))
+    return Image.open(BytesIO(pybase64.b64decode(image, validate=True)))
 
 
 def expand2square(pil_img, background_color):