sglang 0.5.3rc2__py3-none-any.whl → 0.5.4.post1__py3-none-any.whl

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 (

sglang/srt/models/longcat_flash.py

@@ -32,14 +32,10 @@
 
 import concurrent.futures
 import logging
-import os
-from enum import IntEnum, auto
-from typing import Any, Dict, Iterable, Optional, Tuple, Union
+from typing import Iterable, Optional, Tuple
 
 import torch
-import torch.nn.functional as F
 from torch import nn
-from tqdm import tqdm
 
 from sglang.srt.configs import LongcatFlashConfig
 from sglang.srt.distributed import (
@@ -48,9 +44,8 @@ from sglang.srt.distributed import (
 )
 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 import deep_gemm_wrapper
 from sglang.srt.layers.activation import SiluAndMul
-from sglang.srt.layers.amx_utils import PackWeightMethod
 from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
 from sglang.srt.layers.dp_attention import (
     get_attention_tp_rank,
@@ -68,7 +63,6 @@ from sglang.srt.layers.moe.ep_moe.kernels import zero_experts_compute_triton
 from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import StandardTopKOutput, TopK
-from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
 from sglang.srt.layers.quantization.fp8_utils import (
@@ -85,26 +79,21 @@ 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.weight_utils import default_weight_loader
 from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA
+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,
     is_npu,
-    is_sm100_supported,
 )
 
 _is_hip = is_hip()
@@ -117,13 +106,7 @@ _is_cpu = is_cpu()
 _device_sm = get_device_sm()
 
 if _is_cuda:
-    from sgl_kernel import (
-        awq_dequantize,
-        bmm_fp8,
-        dsv3_fused_a_gemm,
-        dsv3_router_gemm,
-        merge_state_v2,
-    )
+    from sgl_kernel import awq_dequantize
 elif _is_cpu and _is_cpu_amx_available:
     pass
 elif _is_hip:
@@ -595,7 +578,7 @@ class LongcatFlashForCausalLM(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)
 

sglang/srt/models/longcat_flash_nextn.py

@@ -32,17 +32,14 @@
 
 import concurrent.futures
 import logging
-import os
-from enum import IntEnum, auto
-from typing import Any, Dict, Iterable, Optional, Tuple, Union
+from typing import Iterable, Optional, Tuple
 
 import torch
-import torch.nn.functional as F
 from torch import nn
-from tqdm import tqdm
 
 from sglang.srt.configs import LongcatFlashConfig
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
+from sglang.srt.layers import deep_gemm_wrapper
 from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
 from sglang.srt.layers.dp_attention import (
     get_attention_tp_rank,
@@ -52,7 +49,6 @@ from sglang.srt.layers.dp_attention import (
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import ReplicatedLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
 from sglang.srt.layers.quantization.fp8_utils import (
@@ -75,7 +71,6 @@ from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA
 from sglang.srt.models.longcat_flash import LongcatFlashForCausalLM, LongcatFlashMLP
 from sglang.srt.utils import (
     BumpAllocator,
-    LazyValue,
     add_prefix,
     bind_or_assign,
     cpu_has_amx_support,
@@ -97,13 +92,7 @@ _is_cpu = is_cpu()
 _device_sm = get_device_sm()
 
 if _is_cuda:
-    from sgl_kernel import (
-        awq_dequantize,
-        bmm_fp8,
-        dsv3_fused_a_gemm,
-        dsv3_router_gemm,
-        merge_state_v2,
-    )
+    from sgl_kernel import awq_dequantize
 elif _is_cpu and _is_cpu_amx_available:
     pass
 elif _is_hip:

sglang/srt/models/mimo.py

@@ -1,28 +1,17 @@
 # Adapted from qwen2.py
 
-from functools import partial
-from typing import Any, Dict, Iterable, Optional, Tuple
+from typing import Iterable, Optional, Tuple
 
 import torch
 from torch import nn
 
-from sglang.srt.distributed import (
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-    split_tensor_along_last_dim,
-    tensor_model_parallel_all_gather,
-)
-from sglang.srt.layers.layernorm import RMSNorm
-from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.rotary_embedding import get_rope
 from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
 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 Qwen2DecoderLayer, Qwen2MLP, Qwen2Model
+from sglang.srt.models.qwen2 import Qwen2DecoderLayer, Qwen2Model
 from sglang.srt.utils import add_prefix
 
 MiMoConfig = None

sglang/srt/models/mimo_mtp.py

@@ -1,7 +1,6 @@
 # Adapted from https://github.com/vllm-project/vllm/pull/17433/files  and deepseek_nextn.py
 
-from functools import partial
-from typing import Any, Dict, Iterable, Optional, Tuple
+from typing import Iterable, Optional, Tuple
 
 import torch
 from torch import nn

sglang/srt/models/minicpmo.py

@@ -43,7 +43,6 @@ from sglang.srt.managers.mm_utils import (
     general_mm_embed_routine,
 )
 from sglang.srt.managers.schedule_batch import (
-    Modality,
     MultimodalDataItem,
     MultimodalInputs,
     flatten_nested_list,
@@ -59,8 +58,6 @@ from sglang.srt.utils import logger
 try:
     from transformers import LogitsWarper
     from vector_quantize_pytorch import GroupedResidualFSQ
-    from vocos import Vocos
-    from vocos.pretrained import instantiate_class
 
     _tts_deps = True
 except:
@@ -795,8 +792,10 @@ class ConditionalChatTTS(PreTrainedModel):
         force_no_stop=False,
         min_new_token=10,
         max_new_token=50,
-        logits_warpers: List[LogitsWarper] = [],
-        logits_processors: List[CustomRepetitionPenaltyLogitsProcessorRepeat] = [],
+        logits_warpers: Optional[List[LogitsWarper]] = None,
+        logits_processors: Optional[
+            List[CustomRepetitionPenaltyLogitsProcessorRepeat]
+        ] = None,
         show_tqdm=False,
     ):
         """Generate audio codes in streaming setting or non-streaming setting.
@@ -825,6 +824,9 @@ class ConditionalChatTTS(PreTrainedModel):
         assert input_ids.shape[0] == 1
         assert past_key_values is not None
 
+        logits_warpers = logits_warpers or []
+        logits_processors = logits_processors or []
+
         # fix: this should not be `input_ids.shape[1]`
         # start_idx = input_ids.shape[1]
         start_idx = (