sglang 0.5.3rc2__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/models/qwen3_vl.py

@@ -15,12 +15,11 @@
 """Inference-only Qwen3-VL model compatible with HuggingFace weights."""
 import logging
 from functools import lru_cache, partial
-from typing import Callable, Iterable, List, Literal, Optional, Tuple, TypedDict, Union
+from typing import Callable, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
 from einops import rearrange
 from transformers.activations import ACT2FN
 from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
@@ -38,16 +37,20 @@ from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternMultimodalTokens,
     general_mm_embed_routine,
 )
-from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
+from sglang.srt.managers.schedule_batch import (
+    Modality,
+    MultimodalDataItem,
+    MultimodalInputs,
+)
 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_vl import Qwen2VLVideoInputs
 from sglang.srt.models.qwen3 import Qwen3Model
 from sglang.srt.utils import add_prefix
 from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 
+
 # === Vision Encoder === #
 
 
@@ -189,14 +192,14 @@ class Qwen3_VisionBlock(nn.Module):
             position_embeddings=position_embeddings,
         )
         attn = rearrange(attn, "b s ... -> s b ...")
-        x = x + attn
+        x += attn
         norm2 = self.norm2(x)
         mlp = self.mlp(norm2)
-        x = x + mlp
+        x += mlp
         return x
 
 
-class Qwen3_VisionPatchMerger(nn.Module):
+class Qwen3VLMoeVisionPatchMerger(nn.Module):
 
     def __init__(
         self,
@@ -246,7 +249,7 @@ class Qwen3_VisionPatchMerger(nn.Module):
         return out
 
 
-class Qwen3_VisionTransformer(nn.Module):
+class Qwen3VLMoeVisionModel(nn.Module):
 
     def __init__(
         self,
@@ -263,10 +266,10 @@ class Qwen3_VisionTransformer(nn.Module):
         self.spatial_merge_size = vision_config.spatial_merge_size
         self.spatial_merge_unit = self.spatial_merge_size**2
         self.temporal_patch_size = vision_config.temporal_patch_size
+        # layer indexes of which layer's output should be deep-stacked
         self.deepstack_visual_indexes = vision_config.deepstack_visual_indexes
         self.patch_embed = Qwen3VLVisionPatchEmbed(config=vision_config)
         self.pos_embed = nn.Embedding(self.num_position_embeddings, self.hidden_size)
-
         norm_layer = partial(nn.LayerNorm, eps=norm_eps)
         head_dim = self.hidden_size // self.num_heads
         self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
@@ -286,7 +289,7 @@ class Qwen3_VisionTransformer(nn.Module):
                 for layer_idx in range(vision_config.depth)
             ]
         )
-        self.merger = Qwen3_VisionPatchMerger(
+        self.merger = Qwen3VLMoeVisionPatchMerger(
             dim=vision_config.out_hidden_size,
             context_dim=self.hidden_size,
             norm_layer=norm_layer,
@@ -297,7 +300,7 @@ class Qwen3_VisionTransformer(nn.Module):
 
         self.deepstack_merger_list = nn.ModuleList(
             [
-                Qwen3_VisionPatchMerger(
+                Qwen3VLMoeVisionPatchMerger(
                     dim=vision_config.out_hidden_size,
                     context_dim=self.hidden_size,
                     spatial_merge_size=self.spatial_merge_size,
@@ -441,7 +444,7 @@ class Qwen3_VisionTransformer(nn.Module):
         x = self.patch_embed(x)
 
         pos_embeds = self.fast_pos_embed_interpolate(grid_thw)
-        x = x + pos_embeds
+        x += pos_embeds
         rotary_pos_emb = self.rot_pos_emb(grid_thw)
 
         seq_len, _ = x.size()
@@ -452,15 +455,16 @@ class Qwen3_VisionTransformer(nn.Module):
         position_embeddings = (emb.cos(), emb.sin())
 
         # compute cu_seqlens
+        cu_seqlens = torch.repeat_interleave(
+            grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
+        ).cumsum(dim=0)
         cu_seqlens = torch.cat(
             [
-                torch.tensor([0], device=grid_thw.device),
-                (grid_thw[:, 0] * grid_thw[:, 1] * grid_thw[:, 2]).cumsum(dim=0),
+                torch.zeros(1, dtype=torch.int32, device=cu_seqlens.device),
+                cu_seqlens.to(torch.int32),
             ]
         )
-        cu_seqlens = F.pad(cu_seqlens, (1, 0), "constant", 0)
 
-        # max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
         x = x.unsqueeze(1)
 
         deepstack_feature_lists = []
@@ -574,10 +578,7 @@ class Qwen3LLMModel(Qwen3Model):
                 and layer_idx in self.deepstack_embed_to_decoder_layer
             ):
                 sep = self.hidden_size * layer_idx
-                hidden_states = (
-                    hidden_states
-                    + input_deepstack_embeds[:, sep : sep + self.hidden_size]
-                )
+                hidden_states += input_deepstack_embeds[:, sep : sep + self.hidden_size]
 
         if not self.pp_group.is_last_rank:
             return PPProxyTensors(
@@ -605,37 +606,43 @@ class Qwen3VLForConditionalGeneration(nn.Module):
         config: Qwen3VLConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        language_model_cls=Qwen3LLMModel,
     ) -> None:
         super().__init__()
 
-        self.config = config
-        self.visual = Qwen3_VisionTransformer(
+        self.visual = Qwen3VLMoeVisionModel(
             config.vision_config,
-            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
             # NOTE: Qwen3-VL vision encoder currently supports BitsAndBytes 4-bit quantization.
             # Other quantization methods (e.g., GPTQ, AWQ) are untested and may not be supported.
             quant_config=quant_config,
+            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
             prefix=add_prefix("visual", prefix),
         )
 
-        self.model = Qwen3LLMModel(
-            config=config,
+        # TODO: make it more elegant
+        if language_model_cls is Qwen3LLMModel:
+            self.config: Qwen3VLConfig = config  # for qwen3-vl
+        else:
+            self.config = config.text_config  # for qwen3-omni
+
+        self.model = language_model_cls(
+            config=self.config,
             quant_config=quant_config,
             prefix=add_prefix("model", prefix),
         )
 
-        if config.tie_word_embeddings:
+        if self.config.tie_word_embeddings:
             self.lm_head = self.model.embed_tokens
         else:
             self.lm_head = ParallelLMHead(
-                config.vocab_size,
-                config.hidden_size,
+                self.config.vocab_size,
+                self.config.hidden_size,
                 quant_config=quant_config,
                 prefix=add_prefix("lm_head", prefix),
             )
         self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
 
-        self.logits_processor = LogitsProcessor(config)
+        self.logits_processor = LogitsProcessor(self.config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
         # like {8:0, 16:1, 24:2}, which stands for the captured deepstack features on
         # 8, 16, 24 layer will be merged to 0, 1, 2 layer of decoder output hidden_states
@@ -643,10 +650,7 @@ class Qwen3VLForConditionalGeneration(nn.Module):
         # deepstack
         self.deepstack_visual_indexes = self.visual.deepstack_visual_indexes
         self.num_deepstack_embeddings = len(self.deepstack_visual_indexes)
-
-    @property
-    def use_deepstack(self) -> bool:
-        return hasattr(self, "deepstack_visual_indexes")
+        self.use_deepstack = {Modality.IMAGE: True, Modality.VIDEO: True}
 
     def separate_deepstack_embeds(self, embedding):
         assert (

sglang/srt/models/qwen3_vl_moe.py

@@ -14,49 +14,23 @@
 # ==============================================================================
 """Inference-only Qwen3-VL model compatible with HuggingFace weights."""
 import logging
-from functools import lru_cache, partial
-from typing import Callable, Iterable, List, Literal, Optional, Tuple, TypedDict, Union
+from functools import lru_cache
+from typing import Iterable, Optional, Tuple, Union
 
-import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
-from einops import rearrange
-from transformers import BatchFeature
-from transformers.activations import ACT2FN
-from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
-    Qwen2_5_VisionRotaryEmbedding,
-)
 
-from sglang.srt.configs.qwen3_vl import Qwen3VLMoeConfig, Qwen3VLMoeVisionConfig
+from sglang.srt.configs.qwen3_vl import Qwen3VLMoeConfig, Qwen3VLMoeTextConfig
 from sglang.srt.distributed import (
     get_moe_expert_parallel_world_size,
-    get_pp_group,
     get_tensor_model_parallel_rank,
 )
-from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
-from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.layers.utils import get_layer_id
-from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
-from sglang.srt.managers.mm_utils import (
-    MultiModalityDataPaddingPatternMultimodalTokens,
-    general_mm_embed_routine,
-)
-from sglang.srt.managers.schedule_batch import (
-    MultimodalDataItem,
-    MultimodalInputs,
-    global_server_args_dict,
-)
 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.qwen3_moe import Qwen3MoeForCausalLM, Qwen3MoeModel
-from sglang.srt.models.qwen3_vl import (
-    Qwen3_VisionTransformer,
-    Qwen3VLForConditionalGeneration,
-)
-from sglang.srt.utils import add_prefix
+from sglang.srt.models.qwen3_moe import Qwen3MoeModel
+from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
 from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
@@ -68,28 +42,16 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
     def __init__(
         self,
         *,
-        config: Qwen3VLMoeConfig,
+        config: Qwen3VLMoeTextConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ):
         super().__init__(config=config, quant_config=quant_config, prefix=prefix)
-
         self.hidden_size = config.hidden_size
 
     def get_input_embeddings(self) -> nn.Embedding:
         return self.embed_tokens
 
-    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
-        # in qwen-vl, last dim is the same
-        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
-            self.visual.dtype
-        )
-        image_grid_thw = torch.concat([item.image_grid_thw for item in items], dim=0)
-        assert pixel_values.dim() == 2, pixel_values.dim()
-        assert image_grid_thw.dim() == 2, image_grid_thw.dim()
-        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
-        return image_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -114,7 +76,7 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
         for layer_idx, layer in enumerate(
             self.layers[self.start_layer : self.end_layer]
         ):
-            layer_idx = layer_idx + self.start_layer
+            layer_idx += self.start_layer
             if layer_idx in self.layers_to_capture:
                 aux_hidden_states.append(
                     hidden_states + residual if residual is not None else hidden_states
@@ -128,11 +90,10 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
             )
 
             # process deepstack
-            if input_deepstack_embeds is not None and layer_idx in range(3):
+            if input_deepstack_embeds is not None and layer_idx < 3:
                 sep = self.hidden_size * layer_idx
-                hidden_states = (
-                    hidden_states
-                    + input_deepstack_embeds[:, sep : sep + self.hidden_size]
+                hidden_states.add_(
+                    input_deepstack_embeds[:, sep : sep + self.hidden_size]
                 )
 
         if not self.pp_group.is_last_rank:
@@ -155,144 +116,56 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
         return hidden_states, aux_hidden_states
 
 
-class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
-    def __init__(
-        self,
-        *,
-        config: Qwen3VLMoeConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-        prefix: str = "",
-    ):
-        super(Qwen3VLForConditionalGeneration, self).__init__()
-        self.config = config
-
-        self.visual = Qwen3_VisionTransformer(
-            config.vision_config,
-            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
-            # NOTE: Qwen3-VL vision encoder currently supports BitsAndBytes 4-bit quantization.
-            # Other quantization methods (e.g., GPTQ, AWQ) are untested and may not be supported.
-            quant_config=quant_config,
-            prefix=add_prefix("visual", prefix),
-        )
-
-        self.model = Qwen3MoeLLMModel(
-            config=config,
-            quant_config=quant_config,
-            prefix=add_prefix("model", prefix),
-        )
-
-        if config.tie_word_embeddings:
-            self.lm_head = self.model.embed_tokens
-        else:
-            self.lm_head = ParallelLMHead(
-                config.vocab_size,
-                config.hidden_size,
-                quant_config=quant_config,
-                prefix=add_prefix("lm_head", prefix),
+def load_fused_expert_weights(
+    name: str,
+    params_dict: dict,
+    loaded_weight: torch.Tensor,
+    shard_id: str,
+    num_experts: int,
+):
+    param = params_dict[name]
+    # weight_loader = typing.cast(Callable[..., bool], param.weight_loader)
+    weight_loader = param.weight_loader
+    ep_rank = get_tensor_model_parallel_rank()
+    ep_size = get_moe_expert_parallel_world_size()
+    if ep_size == 1:
+        for expert_id in range(num_experts):
+            curr_expert_weight = loaded_weight[expert_id]
+            weight_loader(
+                param,
+                curr_expert_weight,
+                name,
+                shard_id,
+                expert_id,
             )
-        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
-
-        self.logits_processor = LogitsProcessor(config)
-        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
-
-        # deepstack
-        self.deepstack_visual_indexes = self.visual.deepstack_visual_indexes
-        self.num_deepstack_embeddings = len(self.deepstack_visual_indexes)
-
-    @property
-    def use_deepstack(self) -> bool:
-        return hasattr(self, "deepstack_visual_indexes")
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        positions: torch.Tensor,
-        forward_batch: ForwardBatch,
-        get_embedding: bool = False,
-    ):
-        """Run forward pass for Qwen3-VL.
-
-        Args:
-            input_ids: Flattened (concatenated) input_ids corresponding to a
-                batch.
-            positions: Flattened (concatenated) position ids corresponding to a
-                batch.
-                **NOTE**: If mrope is enabled (default setting for Qwen2-VL
-                opensource models), the shape will be `(3, seq_len)`,
-                otherwise it will be `(seq_len,).
-                (Use input_metadata.mrope_positions to replace it)
-        """
-        if self.is_mrope_enabled:
-            positions = forward_batch.mrope_positions
-
-        if not (
-            forward_batch.forward_mode.is_decode()
-            or not forward_batch.contains_image_inputs()
-        ):
-            if self.is_mrope_enabled:
-                assert positions.ndim == 2 and positions.size(0) == 3, (
-                    "multimodal section rotary embedding requires "
-                    f"(3, seq_len) positions, but got {positions.size()}"
-                )
-
-        hidden_states = general_mm_embed_routine(
-            input_ids=input_ids,
-            forward_batch=forward_batch,
-            language_model=self.model,
-            multimodal_model=self,
-            positions=positions,
-            use_deepstack=self.use_deepstack,
+    else:
+        experts_per_ep = num_experts // ep_size
+        start_expert = ep_rank * experts_per_ep
+        end_expert = (
+            (ep_rank + 1) * experts_per_ep if ep_rank != ep_size - 1 else num_experts
         )
 
-        if not get_embedding:
-            return self.logits_processor(
-                input_ids, hidden_states, self.lm_head, forward_batch
+        for idx, expert_id in enumerate(range(start_expert, end_expert)):
+            curr_expert_weight = loaded_weight[expert_id]
+            weight_loader(
+                param,
+                curr_expert_weight,
+                name,
+                shard_id,
+                idx,
             )
-        else:
-            return self.pooler(hidden_states, forward_batch)
+    return True
 
-    def load_fused_expert_weights(
+
+class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
+    def __init__(
         self,
-        name: str,
-        params_dict: dict,
-        loaded_weight: torch.Tensor,
-        shard_id: str,
-        num_experts: int,
+        config: Qwen3VLMoeConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        language_model_cls=Qwen3MoeLLMModel,
     ):
-        param = params_dict[name]
-        # weight_loader = typing.cast(Callable[..., bool], param.weight_loader)
-        weight_loader = param.weight_loader
-        ep_rank = get_tensor_model_parallel_rank()
-        ep_size = get_moe_expert_parallel_world_size()
-        if ep_size == 1:
-            for expert_id in range(num_experts):
-                curr_expert_weight = loaded_weight[expert_id]
-                weight_loader(
-                    param,
-                    curr_expert_weight,
-                    name,
-                    shard_id,
-                    expert_id,
-                )
-        else:
-            experts_per_ep = num_experts // ep_size
-            start_expert = ep_rank * experts_per_ep
-            end_expert = (
-                (ep_rank + 1) * experts_per_ep
-                if ep_rank != ep_size - 1
-                else num_experts
-            )
-
-            for idx, expert_id in enumerate(range(start_expert, end_expert)):
-                curr_expert_weight = loaded_weight[expert_id]
-                weight_loader(
-                    param,
-                    curr_expert_weight,
-                    name,
-                    shard_id,
-                    idx,
-                )
-        return True
+        super().__init__(config, quant_config, prefix, language_model_cls)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
@@ -338,8 +211,7 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
             self._cached_params_dict = dict(self.named_parameters())
         params_dict = self._cached_params_dict
         for name, loaded_weight in weights:
-            if "language_model" in name:
-                name = name.replace(r"model.language_model.", r"model.")
+            name = name.replace(r"model.language_model.", r"model.")
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if "experts.gate_up_proj" in name or "experts.down_proj" in name:
@@ -393,14 +265,14 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
                         loaded_weight = loaded_weight.transpose(-1, -2)  # no bias
                         if "experts.gate_up_proj" in name:
                             loaded_weight = loaded_weight.chunk(2, dim=-2)
-                            self.load_fused_expert_weights(
+                            load_fused_expert_weights(
                                 name_mapped,
                                 params_dict,
                                 loaded_weight[0],
                                 "w1",
                                 num_experts,
                             )
-                            self.load_fused_expert_weights(
+                            load_fused_expert_weights(
                                 name_mapped,
                                 params_dict,
                                 loaded_weight[1],
@@ -408,7 +280,7 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
                                 num_experts,
                             )
                         else:
-                            self.load_fused_expert_weights(
+                            load_fused_expert_weights(
                                 name_mapped,
                                 params_dict,
                                 loaded_weight,

sglang/srt/models/roberta.py

@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 
-import itertools
+import os
 from typing import Iterable, Optional, Tuple
 
 import torch
@@ -8,10 +8,12 @@ from torch import nn
 
 from sglang.srt.layers.pooler import CrossEncodingPooler, Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.sparse_pooler import SparsePooler
 from sglang.srt.layers.vocab_parallel_embedding import 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.bert import BertEncoder
+from sglang.srt.utils.hf_transformers_utils import download_from_hf
 
 RobertaConfig = None
 
@@ -206,12 +208,29 @@ class XLMRobertaModel(nn.Module):
         config: RobertaConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        sparse_head: Optional[str] = None,
+        model_path: Optional[str] = None,
     ):
         super().__init__()
         self.roberta = XLMRobertaBaseModel(
             config=config, quant_config=quant_config, prefix=prefix
         )
-        self.pooler = Pooler(pooling_type=PoolingType.CLS, normalize=True)
+        if sparse_head is not None:
+            self._is_sparse = True
+            self._model_path = model_path
+            self._sparse_head = sparse_head
+            self.pooler = SparsePooler(config=config)
+            # Zero out special tokens
+            self._special_tokens = [
+                config.bos_token_id,
+                config.eos_token_id,
+                config.pad_token_id,
+                # self.config.unk_token_id # not available in the XLMRobertaConfig
+            ]
+            self._special_tokens = [t for t in self._special_tokens if t is not None]
+        else:
+            self._is_sparse = False
+            self.pooler = Pooler(pooling_type=PoolingType.CLS, normalize=True)
 
     def forward(
         self,
@@ -224,11 +243,44 @@ class XLMRobertaModel(nn.Module):
         hidden_states = self.roberta(
             input_ids, positions, forward_batch, input_embeds, get_embedding
         )
-        return self.pooler(hidden_states, forward_batch)
+        embeddings = self.pooler(hidden_states, forward_batch)
+
+        if self._is_sparse:
+            for token_id in self._special_tokens:
+                embeddings.embeddings[:, token_id] = 0.0
+            embeddings.embeddings = embeddings.embeddings.to_sparse()
+
+        return embeddings
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         self.roberta.load_weights(weights)
 
+        if self._is_sparse:
+            sparse_dict = XLMRobertaModel._load_sparse_linear(
+                self._model_path, self._sparse_head
+            )
+            self.pooler.load_weights(sparse_dict)
+
+    @staticmethod
+    def _load_sparse_linear(model_path_or_dir: str, sparse_head: str) -> dict:
+        """
+        Load sparse_head from local dir or HF Hub.
+        Returns a state_dict suitable for nn.Linear.load_state_dict().
+        """
+        if os.path.isdir(model_path_or_dir):
+            path = os.path.join(model_path_or_dir, sparse_head)
+            if not os.path.exists(path):
+                raise FileNotFoundError(
+                    f"'{sparse_head}' not found in {model_path_or_dir}"
+                )
+        else:
+            # remote → use SGLang HF utility
+            local_dir = download_from_hf(model_path_or_dir, allow_patterns=sparse_head)
+            path = os.path.join(local_dir, sparse_head)
+
+        state_dict = torch.load(path)
+        return state_dict
+
 
 class XLMRobertaForSequenceClassification(nn.Module):
     def __init__(

sglang/srt/models/sarashina2_vision.py

@@ -17,7 +17,6 @@ import logging
 from typing import Iterable, List, Optional, Tuple
 
 import torch
-import torch.nn.functional as F
 from torch import nn
 from transformers import LlamaConfig
 

sglang/srt/models/step3_vl.py

@@ -1,8 +1,7 @@
 import logging
 import math
-from collections.abc import Iterable
 from math import sqrt
-from typing import Any, Dict, Iterable, List, Literal, Optional, Tuple, TypedDict, Union
+from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -57,7 +56,6 @@ from sglang.srt.managers.schedule_batch import (
     Modality,
     MultimodalDataItem,
     MultimodalInputs,
-    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
@@ -300,7 +298,7 @@ class Step3TextDecoderLayer(nn.Module):
         # self.n_shared_experts = 1
         # self.num_fused_shared_experts = (
         #     0
-        #     if global_server_args_dict["disable_shared_experts_fusion"]
+        #     if global_server_args.disable_shared_experts_fusion
         #     else self.n_shared_experts
         # )
         self.num_fused_shared_experts = 0
@@ -774,7 +772,7 @@ class Step3VLForConditionalGeneration(nn.Module):
         # self.n_shared_experts = 1
         # self.num_fused_shared_experts = (
         #     0
-        #     if global_server_args_dict["disable_shared_experts_fusion"]
+        #     if global_server_args.disable_shared_experts_fusion
         #     else self.n_shared_experts
         # )
         self.num_fused_shared_experts = 0