sglang 0.5.3rc0__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/models/glm4_moe.py

@@ -12,7 +12,7 @@
 # limitations under the License.
 # ==============================================================================
 
-"""Inference-only GLM-4.5 model compatible with HuggingFace weights"""
+"""Inference-only GLM-4.5, GLM-4.6 model compatible with HuggingFace weights"""
 
 import logging
 from typing import Any, Dict, Iterable, Optional, Tuple
@@ -27,7 +27,6 @@ from sglang.srt.distributed import (
     get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
-    parallel_state,
     tensor_model_parallel_all_reduce,
 )
 from sglang.srt.layers.activation import SiluAndMul
@@ -44,30 +43,23 @@ from sglang.srt.layers.dp_attention import (
 )
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
-    ColumnParallelLinear,
     MergedColumnParallelLinear,
     QKVParallelLinear,
-    ReplicatedLinear,
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe import get_deepep_mode, get_moe_a2a_backend
+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.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.layers.quantization.fp8_kernel import (
-    is_fp8_fnuz,
-    per_tensor_quant_mla_fp8,
-    per_token_group_quant_mla_deep_gemm_masked_fp8,
-)
+from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
 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.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
 from sglang.srt.model_loader.weight_utils import default_weight_loader
@@ -77,21 +69,17 @@ from sglang.srt.models.deepseek_v2 import (
     DeepseekV2Model,
     DeepseekV2MoE,
 )
-from sglang.srt.two_batch_overlap import MaybeTboDeepEPDispatcher
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     BumpAllocator,
     LazyValue,
     add_prefix,
-    bind_or_assign,
     cpu_has_amx_support,
     get_bool_env_var,
     get_device_sm,
-    get_int_env_var,
     is_cpu,
     is_cuda,
-    is_flashinfer_available,
     is_hip,
-    is_non_idle_and_non_empty,
     log_info_on_rank0,
     use_intel_amx_backend,
 )
@@ -395,7 +383,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
         self.n_shared_experts = config.n_shared_experts
         self.num_fused_shared_experts = (
             0
-            if global_server_args_dict["disable_shared_experts_fusion"]
+            if get_global_server_args().disable_shared_experts_fusion
             else config.n_shared_experts
         )
         self.config = config
@@ -432,7 +420,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
         self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.n_routed_experts
             + self.num_fused_shared_experts
-            + global_server_args_dict["ep_num_redundant_experts"],
+            + get_global_server_args().ep_num_redundant_experts,
             num_fused_shared_experts=self.num_fused_shared_experts,
             top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
             hidden_size=config.hidden_size,
@@ -471,12 +459,12 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
 
         self.top_k = config.num_experts_per_tok
 
-        if get_moe_a2a_backend().is_deepep():
+        if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake():
             # TODO: we will support tp < ep in the future
             self.ep_size = get_moe_expert_parallel_world_size()
             self.num_experts = (
                 config.n_routed_experts
-                + global_server_args_dict["ep_num_redundant_experts"]
+                + get_global_server_args().ep_num_redundant_experts
             )
             self.renormalize = config.norm_topk_prob
             self.topk_group = config.topk_group
@@ -487,20 +475,9 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
                 else None
             )
 
-            self.deepep_dispatcher = MaybeTboDeepEPDispatcher(
-                group=parallel_state.get_tp_group().device_group,
-                router_topk=self.top_k,
-                permute_fusion=True,
-                num_experts=self.num_experts,
-                num_local_experts=config.n_routed_experts // self.tp_size,
-                hidden_size=config.hidden_size,
-                params_dtype=config.torch_dtype,
-                deepep_mode=get_deepep_mode(),
-                async_finish=True,
-                return_recv_hook=True,
-            )
-
-        self._enable_deepep_moe = get_moe_a2a_backend().is_deepep()
+        self._enable_a2a_moe = (
+            get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake()
+        )
 
     def forward_normal_dual_stream(
         self,
@@ -664,7 +641,7 @@ class Glm4MoeDecoderLayer(DeepseekV2DecoderLayer):
             layer_scatter_modes=self.layer_scatter_modes,
             input_layernorm=self.input_layernorm,
             post_attention_layernorm=self.post_attention_layernorm,
-            allow_reduce_scatter=True,
+            allow_reduce_scatter=False,
         )
 
     def forward(
@@ -758,7 +735,7 @@ class Glm4MoeForCausalLM(DeepseekV2ForCausalLM):
             config.hidden_size,
             quant_config=quant_config,
             prefix=add_prefix("lm_head", prefix),
-            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+            use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
         )
         self.logits_processor = LogitsProcessor(config)
 
@@ -774,7 +751,7 @@ class Glm4MoeForCausalLM(DeepseekV2ForCausalLM):
         self, architecture: str = "Glm4MoeForCausalLM"
     ):
         self.num_fused_shared_experts = 0
-        if global_server_args_dict["disable_shared_experts_fusion"]:
+        if get_global_server_args().disable_shared_experts_fusion:
             return
 
         # Only Deepseek V3/R1 can use shared experts fusion optimization now.
@@ -785,12 +762,12 @@ class Glm4MoeForCausalLM(DeepseekV2ForCausalLM):
             or self.config.architectures[0] != architecture
             or self.config.n_shared_experts != 1
         ):
-            disable_reason = "Only GLM-4.5 on NV-platform with capability >= 80 can use shared experts fusion optimization."
+            disable_reason = "Only GLM-4.5 or GLM-4.6 on NV-platform with capability >= 80 can use shared experts fusion optimization."
         elif get_moe_expert_parallel_world_size() > 1:
-            disable_reason = "Deepseek and GLM-4.5 can not use shared experts fusion optimization under expert parallelism."
+            disable_reason = "Deepseek and GLM-4.5 or GLM-4.6 can not use shared experts fusion optimization under expert parallelism."
 
         if disable_reason is not None:
-            global_server_args_dict["disable_shared_experts_fusion"] = True
+            get_global_server_args().disable_shared_experts_fusion = True
             self.num_fused_shared_experts = 0
             log_info_on_rank0(
                 logger,

sglang/srt/models/glm4_moe_nextn.py

@@ -12,7 +12,7 @@
 # limitations under the License.
 # ==============================================================================
 
-"""Inference-only GLM-4.5 NextN Speculative Decoding."""
+"""Inference-only GLM-4.5, GLM-4.6 NextN Speculative Decoding."""
 import logging
 from typing import Iterable, Optional, Tuple
 
@@ -30,9 +30,9 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.glm4_moe import Glm4MoeDecoderLayer, Glm4MoeForCausalLM
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import BumpAllocator, add_prefix
 
 logger = logging.getLogger(__name__)
@@ -48,7 +48,7 @@ class Glm4MoeModelNextN(nn.Module):
         super().__init__()
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
             logger.warning(
-                "Overriding Glm4MoeForCausalLMNextN quant config for modelopt_fp4 GLM-4.5 model."
+                "Overriding Glm4MoeForCausalLMNextN quant config for modelopt_fp4 GLM-4.5 / GLM-4.6 model."
             )
             quant_config = None
 
@@ -145,7 +145,7 @@ class Glm4MoeForCausalLMNextN(Glm4MoeForCausalLM):
             config.hidden_size,
             quant_config=quant_config,
             prefix=add_prefix("model.shared_head.head", prefix),
-            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+            use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
         )
         self.logits_processor = LogitsProcessor(config)
 

sglang/srt/models/glm4v.py

@@ -7,9 +7,9 @@ import torch.nn as nn
 import torch.nn.functional as F
 from transformers.models.glm4v.configuration_glm4v import Glm4vConfig, Glm4vVisionConfig
 
-from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.attention import vision_utils
+from sglang.srt.layers.dp_attention import get_attention_tp_size
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     ColumnParallelLinear,
@@ -28,6 +28,7 @@ from sglang.srt.models.qwen2_5_vl import (
     Qwen2_5_VLForConditionalGeneration,
 )
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 
@@ -434,7 +435,7 @@ class Glm4vVisionModel(nn.Module):
         cu_seqlens = torch.repeat_interleave(
             grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
         ).cumsum(dim=0, dtype=torch.int32)
-        cu_seqlens = F.pad(cu_seqlens, (1, 0), "constant", 0)
+        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
 
         seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
         x = self.embeddings(

sglang/srt/models/glm4v_moe.py

@@ -10,18 +10,18 @@ from sglang.srt.distributed import (
     get_moe_expert_parallel_world_size,
     get_tensor_model_parallel_world_size,
 )
-from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.attention import vision_utils
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.glm4_moe import Glm4MoeModel
 from sglang.srt.models.glm4v import Glm4vForConditionalGeneration, Glm4vVisionModel
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, is_cuda, log_info_on_rank0
+from sglang.srt.utils.hf_transformers_utils import get_processor
 
 _is_cuda = is_cuda()
 
@@ -47,7 +47,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
         self.determine_num_fused_shared_experts("Glm4MoeForCausalLM")
         self.num_fused_shared_experts = (
             0
-            if global_server_args_dict["disable_shared_experts_fusion"]
+            if get_global_server_args().disable_shared_experts_fusion
             else config.n_shared_experts
         )
 
@@ -68,7 +68,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
             config.hidden_size,
             quant_config=quant_config,
             prefix=add_prefix("lm_head", prefix),
-            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+            use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
         )
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
@@ -81,7 +81,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
         self, architecture: str = "Glm4MoeForCausalLM"
     ):
         self.num_fused_shared_experts = 0
-        if global_server_args_dict["disable_shared_experts_fusion"]:
+        if get_global_server_args().disable_shared_experts_fusion:
             return
 
         # Only Deepseek V3/R1 can use shared experts fusion optimization now.
@@ -97,7 +97,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
             disable_reason = "Deepseek and GLM-4.5 can not use shared experts fusion optimization under expert parallelism."
 
         if disable_reason is not None:
-            global_server_args_dict["disable_shared_experts_fusion"] = True
+            get_global_server_args().disable_shared_experts_fusion = True
             self.num_fused_shared_experts = 0
             log_info_on_rank0(
                 logger,

sglang/srt/models/gpt_oss.py

@@ -63,9 +63,13 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import 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.utils import (
+    create_fused_set_kv_buffer_arg,
+    enable_fused_set_kv_buffer,
+)
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     LazyValue,
     add_prefix,
@@ -81,7 +85,7 @@ _is_sm100_supported = is_cuda() and is_sm100_supported()
 
 
 if _is_cuda:
-    from sgl_kernel import FusedSetKVBufferArg
+    from sgl_kernel import FusedSetKVBufferArg  # noqa: F401
 
 
 class GptOssConfig(PretrainedConfig):
@@ -134,7 +138,7 @@ class GptOssSparseMoeBlock(nn.Module):
             }
         self.experts = experts_type(
             num_experts=config.num_local_experts
-            + global_server_args_dict["ep_num_redundant_experts"],
+            + get_global_server_args().ep_num_redundant_experts,
             top_k=config.num_experts_per_tok,
             layer_id=layer_id,
             hidden_size=config.hidden_size,
@@ -193,33 +197,6 @@ class GptOssSparseMoeBlock(nn.Module):
         return ans
 
 
-def _enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
-    """Enable fused set_kv_buffer only on CUDA with bfloat16 KV cache."""
-    return _is_cuda and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
-
-
-# TODO maybe move to a model-common utils
-def _create_fused_set_kv_buffer_arg(
-    value: torch.Tensor,
-    layer: RadixAttention,
-    forward_batch: ForwardBatch,
-):
-    layer_id = layer.layer_id
-    token_to_kv_pool = forward_batch.token_to_kv_pool
-
-    k_buffer = token_to_kv_pool.get_key_buffer(layer_id)
-    v_buffer = token_to_kv_pool.get_value_buffer(layer_id)
-
-    return FusedSetKVBufferArg(
-        value=value,
-        k_buffer=k_buffer.view(k_buffer.shape[0], -1),
-        v_buffer=v_buffer.view(v_buffer.shape[0], -1),
-        k_scale=layer.k_scale,
-        v_scale=layer.v_scale,
-        cache_loc=forward_batch.out_cache_loc,
-    )
-
-
 class GptOssAttention(nn.Module):
     def __init__(
         self,
@@ -282,7 +259,7 @@ class GptOssAttention(nn.Module):
 
         # Choose dtype of sinks based on attention backend: trtllm_mha requires float32,
         # others can use bfloat16
-        attn_backend = global_server_args_dict.get("attention_backend")
+        attn_backend = get_global_server_args().attention_backend
         sinks_dtype = torch.float32 if attn_backend == "trtllm_mha" else torch.bfloat16
         self.sinks = nn.Parameter(
             torch.empty(self.num_heads, dtype=sinks_dtype), requires_grad=False
@@ -337,12 +314,12 @@ class GptOssAttention(nn.Module):
             q,
             k,
             fused_set_kv_buffer_arg=(
-                _create_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)
+                if enable_fused_set_kv_buffer(forward_batch)
                 else None
             ),
         )
@@ -356,7 +333,7 @@ class GptOssAttention(nn.Module):
         attn_output = self.attn(
             *inner_state,
             sinks=self.sinks,
-            save_kv_cache=not _enable_fused_set_kv_buffer(forward_batch),
+            save_kv_cache=not enable_fused_set_kv_buffer(forward_batch),
         )
         output, _ = self.o_proj(attn_output)
         return output
@@ -614,7 +591,7 @@ class GptOssForCausalLM(nn.Module):
             config.hidden_size,
             # quant_config=quant_config,
             prefix=add_prefix("lm_head", prefix),
-            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+            use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
         )
         self.logits_processor = LogitsProcessor(config)
         self.capture_aux_hidden_states = False

sglang/srt/models/grok.py

@@ -28,7 +28,6 @@ from torch import nn
 from transformers import PretrainedConfig
 
 from sglang.srt.distributed import (
-    get_moe_expert_parallel_world_size,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_gather,
@@ -36,7 +35,6 @@ from sglang.srt.distributed import (
 )
 from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.elementwise import (
-    experts_combine_triton,
     fused_dual_residual_rmsnorm,
     fused_rmsnorm,
     gelu_and_mul_triton,
@@ -49,7 +47,6 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe.ep_moe.layer import EPMoE
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.router import fused_moe_router_shim
 from sglang.srt.layers.moe.topk import TopK
@@ -65,10 +62,10 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.loader import DefaultModelLoader
 from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, dispose_tensor, dump_to_file
 
 logger = logging.getLogger(__name__)
@@ -76,9 +73,6 @@ logger = logging.getLogger(__name__)
 
 # Dump tensors for debugging
 debug_tensor_dump_output_folder = None
-debug_tensor_dump_prefill_only = False
-# Skip all the other tensor dumps, only dump the target logits
-debug_tensor_dump_only_target_logprobs = False
 debug_tensor_dump_inject = False
 debug_tensor_dump_layers = None
 debug_tensor_dump_test = False
@@ -176,17 +170,7 @@ class Grok1MoE(nn.Module):
             custom_routing_function=custom_routing_function,
         )
 
-        kwargs = {}
-        if get_moe_expert_parallel_world_size() > 1:
-            MoEImpl = EPMoE
-        else:
-            MoEImpl = FusedMoE
-            kwargs["reduce_results"] = reduce_results
-            kwargs["use_presharded_weights"] = use_presharded_weights
-            kwargs["inplace"] = inplace
-            kwargs["no_combine"] = no_combine
-
-        self.experts = MoEImpl(
+        self.experts = FusedMoE(
             num_experts=num_experts,
             top_k=top_k,
             layer_id=layer_id,
@@ -195,7 +179,10 @@ class Grok1MoE(nn.Module):
             params_dtype=params_dtype,
             quant_config=quant_config,
             activation="gelu",
-            **kwargs,
+            reduce_results=reduce_results,
+            use_presharded_weights=use_presharded_weights,
+            inplace=inplace,
+            no_combine=no_combine,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -877,10 +864,10 @@ class Grok1ForCausalLM(nn.Module):
 
         # Dump tensors for debugging
         global debug_tensor_dump_output_folder, debug_tensor_dump_inject
-        debug_tensor_dump_output_folder = global_server_args_dict[
-            "debug_tensor_dump_output_folder"
-        ]
-        debug_tensor_dump_inject = global_server_args_dict["debug_tensor_dump_inject"]
+        debug_tensor_dump_output_folder = (
+            get_global_server_args().debug_tensor_dump_output_folder
+        )
+        debug_tensor_dump_inject = get_global_server_args().debug_tensor_dump_inject
         warnings.filterwarnings("ignore", category=FutureWarning)
 
         if get_tensor_model_parallel_rank() == 0:

sglang/srt/models/hunyuan.py

@@ -12,18 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only HunYuan model compatible with HuggingFace weights."""
-import logging
 import re
-from dataclasses import dataclass
-from enum import Enum, auto
-from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
 from torch import nn
 from transformers import PretrainedConfig
 
 from sglang.srt.distributed import (
-    get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
@@ -46,7 +42,6 @@ from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
 from sglang.srt.layers.sampler import Sampler
 from sglang.srt.layers.vocab_parallel_embedding import (
-    DEFAULT_VOCAB_PADDING_SIZE,
     ParallelLMHead,
     VocabParallelEmbedding,
 )
@@ -56,7 +51,7 @@ from sglang.srt.model_loader.weight_utils import (
     kv_cache_scales_loader,
     maybe_remap_kv_scale_name,
 )
-from sglang.srt.utils import add_prefix, is_hip
+from sglang.srt.utils import is_hip
 
 expert_distribution_recorder = ExpertDistributionRecorder()
 

sglang/srt/models/interns1.py

@@ -5,7 +5,6 @@ from torch import nn
 from transformers import PretrainedConfig
 
 from sglang.srt.layers.attention import vision_utils
-from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.managers.mm_utils import (

sglang/srt/models/kimi_vl.py

@@ -43,10 +43,8 @@
 
 import copy
 import logging
-import math
-from collections.abc import Mapping
 from dataclasses import dataclass
-from typing import Any, Iterable, List, Optional, Tuple
+from typing import Iterable, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -56,10 +54,6 @@ from sglang.srt.configs import KimiVLConfig
 from sglang.srt.configs.deepseekvl2 import DeepseekV2Config
 from sglang.srt.configs.kimi_vl import KimiVLConfig
 from sglang.srt.configs.kimi_vl_moonvit import MoonViTConfig
-from sglang.srt.distributed import (
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-)
 from sglang.srt.layers.activation import QuickGELU
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig

sglang/srt/models/kimi_vl_moonvit.py

@@ -49,7 +49,7 @@ from typing import List, Optional, Sequence, Tuple, Union
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from transformers.activations import ACT2FN, PytorchGELUTanh
+from transformers.activations import ACT2FN
 from transformers.modeling_utils import PreTrainedModel
 
 try:
@@ -596,6 +596,8 @@ class MoonVitPretrainedModel(PreTrainedModel):
     _supports_sdpa = True
 
     def __init__(self, config: MoonViTConfig, *inputs, **kwargs):
+        from transformers.activations import GELUTanh
+
         super().__init__(config, *inputs, **kwargs)
         config = deepcopy(config)
         self.merge_kernel_size = config.merge_kernel_size
@@ -614,7 +616,7 @@ class MoonVitPretrainedModel(PreTrainedModel):
                 "num_heads": config.num_attention_heads,
                 "hidden_dim": config.hidden_size,
                 "mlp_dim": config.intermediate_size,
-                "activation": PytorchGELUTanh(),
+                "activation": GELUTanh(),
                 "attn_bias": True,
                 "attn_implementation": config._attn_implementation,
             },

sglang/srt/models/llama.py

@@ -45,13 +45,13 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import 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,
     kv_cache_scales_loader,
     maybe_remap_kv_scale_name,
 )
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, make_layers
 from sglang.utils import get_exception_traceback
 
@@ -385,6 +385,10 @@ class LlamaModel(nn.Module):
                     "Self attention has no KV cache scaling " "factor attribute!"
                 )
 
+    def get_input_embeddings(self) -> nn.Embedding:
+        """Get input embeddings from the model."""
+        return self.embed_tokens
+
 
 class LlamaForCausalLM(nn.Module):
     # BitandBytes specific attributes
@@ -429,7 +433,7 @@ class LlamaForCausalLM(nn.Module):
                 config.hidden_size,
                 quant_config=quant_config,
                 prefix=add_prefix("lm_head", prefix),
-                use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+                use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
             )
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)

sglang/srt/models/llama_eagle3.py

@@ -27,7 +27,7 @@ from transformers import LlamaConfig
 
 from sglang.srt.distributed import get_pp_group
 from sglang.srt.layers.layernorm import RMSNorm
-from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
+from sglang.srt.layers.linear import QKVParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.vocab_parallel_embedding import (