Update vllm to v0.4.3 (#511)

Co-authored-by: Qubitium <417764+Qubitium@users.noreply.github.com>
Co-authored-by: ZX <zx@lbx.dev>

python/pyproject.toml

@@ -20,7 +20,7 @@ dependencies = [
 
 [project.optional-dependencies]
 srt = ["aiohttp", "fastapi", "psutil", "rpyc", "torch", "uvloop", "uvicorn",
-       "zmq", "vllm==0.4.2", "interegular", "pydantic", "pillow", "packaging", "huggingface_hub", "hf_transfer", "outlines>=0.0.34"]
+       "zmq", "vllm==0.4.3", "interegular", "pydantic", "pillow", "packaging", "huggingface_hub", "hf_transfer", "outlines>=0.0.34"]
 openai = ["openai>=1.0", "numpy", "tiktoken"]
 anthropic = ["anthropic>=0.20.0", "numpy"]
 all = ["sglang[srt]", "sglang[openai]", "sglang[anthropic]"]

python/sglang/srt/managers/controller/model_runner.py

@@ -11,7 +11,7 @@ import torch
 import torch.nn as nn
 from vllm.config import DeviceConfig, LoadConfig
 from vllm.config import ModelConfig as VllmModelConfig
-from vllm.distributed import initialize_model_parallel
+from vllm.distributed import initialize_model_parallel, init_distributed_environment
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.models import ModelRegistry
 
@@ -240,11 +240,11 @@ class ModelRunner:
         logger.info(f"[gpu_id={self.gpu_id}] Set cuda device.")
         torch.cuda.set_device(self.gpu_id)
         logger.info(f"[gpu_id={self.gpu_id}] Init nccl begin.")
-        torch.distributed.init_process_group(
+        init_distributed_environment(
             backend="nccl",
             world_size=self.tp_size,
             rank=self.tp_rank,
-            init_method=f"tcp://127.0.0.1:{self.nccl_port}",
+            distributed_init_method=f"tcp://127.0.0.1:{self.nccl_port}",
         )
         initialize_model_parallel(tensor_model_parallel_size=self.tp_size)
         total_gpu_memory = get_available_gpu_memory(
@@ -291,6 +291,7 @@ class ModelRunner:
             vision_language_config=None,
             parallel_config=None,
             scheduler_config=None,
+            cache_config=None,
         )
         logger.info(
             f"[gpu_id={self.gpu_id}] Load weight end. "

python/sglang/srt/models/commandr.py

@@ -30,6 +30,7 @@ import torch.utils.checkpoint
 from torch import nn
 from torch.nn.parameter import Parameter
 from transformers import PretrainedConfig
+from vllm.config import CacheConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -304,6 +305,7 @@ class CohereForCausalLM(nn.Module):
         self,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/dbrx.py

@@ -5,6 +5,7 @@ from typing import Iterable, Optional, Tuple
 
 import torch
 import torch.nn as nn
+from vllm.config import CacheConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -352,6 +353,7 @@ class DbrxForCausalLM(nn.Module):
         self,
         config: DbrxConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ):
         super().__init__()
         self.config = config

python/sglang/srt/models/gemma.py

@@ -6,7 +6,7 @@ from typing import Iterable, Optional, Tuple
 import torch
 from torch import nn
 from transformers import PretrainedConfig
-from vllm.config import LoRAConfig
+from vllm.config import LoRAConfig, CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import GeluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -264,6 +264,7 @@ class GemmaForCausalLM(nn.Module):
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
         lora_config: Optional[LoRAConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         del lora_config  # Unused.
         super().__init__()

python/sglang/srt/models/grok.py

@@ -11,6 +11,7 @@ from torch import nn
 from transformers import PretrainedConfig
 
 from vllm import _custom_ops as ops
+from vllm.config import CacheConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -538,6 +539,7 @@ class Grok1ModelForCausalLM(nn.Module):
         self,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/llama2.py

@@ -7,6 +7,7 @@ import torch
 import tqdm
 from torch import nn
 from transformers import LlamaConfig
+from vllm.config import CacheConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size
@@ -258,6 +259,7 @@ class LlamaForCausalLM(nn.Module):
         self,
         config: LlamaConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/llava.py

@@ -7,6 +7,7 @@ import torch
 from torch import nn
 from transformers import CLIPVisionModel, CLIPVisionConfig, LlavaConfig, Qwen2Config, MistralConfig 
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
+from vllm.config import CacheConfig
 from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
@@ -27,6 +28,7 @@ class LlavaLlamaForCausalLM(nn.Module):
         self,
         config: LlavaConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -294,8 +296,9 @@ class LlavaQwenForCausalLM(LlavaLlamaForCausalLM):
         self,
         config: LlavaConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
-        super().__init__(config, quant_config=quant_config)
+        super().__init__(config, quant_config=quant_config, cache_config=cache_config)
         self.config = config
         self.vision_tower = None
         if getattr(self.config, "vision_config", None) is None:
@@ -356,8 +359,9 @@ class LlavaMistralForCausalLM(LlavaLlamaForCausalLM):
         self,
         config: LlavaConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
-        super().__init__(config, quant_config=quant_config)
+        super().__init__(config, quant_config=quant_config, cache_config=cache_config)
         self.config = config
         self.vision_tower = None
         if getattr(self.config, "vision_config", None) is None:

python/sglang/srt/models/llavavid.py

@@ -7,6 +7,7 @@ import torch
 from torch import nn
 from transformers import CLIPVisionModel, LlavaConfig
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
+from vllm.config import CacheConfig
 from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
@@ -25,6 +26,7 @@ class LlavaVidForCausalLM(nn.Module):
         self,
         config: LlavaConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/mixtral.py

@@ -9,6 +9,7 @@ import torch.nn.functional as F
 from torch import nn
 from transformers import MixtralConfig
 from vllm import _custom_ops as ops
+from vllm.config import CacheConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -414,6 +415,7 @@ class MixtralForCausalLM(nn.Module):
         self,
         config: MixtralConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/mixtral_quant.py

@@ -8,6 +8,7 @@ import torch
 import torch.nn.functional as F
 from torch import nn
 from transformers import MixtralConfig
+from vllm.config import CacheConfig
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -313,6 +314,7 @@ class QuantMixtralForCausalLM(nn.Module):
         self,
         config: MixtralConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/qwen.py

@@ -5,6 +5,7 @@ from typing import Any, Dict, Optional, Iterable, Tuple
 import torch
 from torch import nn
 from transformers import PretrainedConfig
+from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -227,6 +228,7 @@ class QWenLMHeadModel(nn.Module):
         self,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ):
         super().__init__()
         self.config = config

python/sglang/srt/models/qwen2.py

@@ -5,6 +5,7 @@ from typing import Any, Dict, Optional, Tuple, Iterable
 
 import torch
 from torch import nn
+from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -251,6 +252,7 @@ class Qwen2ForCausalLM(nn.Module):
         self,
         config: Qwen2Config,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/stablelm.py

@@ -7,6 +7,7 @@ from typing import Optional, Tuple, Iterable
 import torch
 from torch import nn
 from transformers import PretrainedConfig
+from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.linear import (
@@ -225,6 +226,7 @@ class StableLmForCausalLM(nn.Module):
         self,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config

python/sglang/srt/models/yivl.py

 """Inference-only Yi-VL model."""
 
-from typing import Tuple, Iterable
+from typing import Tuple, Iterable, Optional
 
 import torch
 import torch.nn as nn
 from transformers import CLIPVisionModel, LlavaConfig
+from vllm.config import CacheConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.models.llava import (
     LlavaLlamaForCausalLM,
     monkey_path_clip_vision_embed_forward,
@@ -15,9 +17,12 @@ from sglang.srt.models.llava import (
 
 class YiVLForCausalLM(LlavaLlamaForCausalLM):
     def __init__(
-        self, config, quant_config = None,
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
-        super().__init__(config, quant_config)
+        super().__init__(config, quant_config, cache_config)
 
         self.multi_modal_projector = YiVLMultiModalProjector(self.config)
         self.vision_tower_subfolder = self.config.mm_vision_tower.replace(

python/sglang/srt/utils.py

@@ -421,9 +421,10 @@ def suppress_other_loggers():
     from vllm.logger import logger as vllm_default_logger
 
     vllm_default_logger.setLevel(logging.WARN)
-    logging.getLogger("vllm.utils").setLevel(logging.WARN)
-    logging.getLogger("vllm.selector").setLevel(logging.WARN)
     logging.getLogger("vllm.config").setLevel(logging.ERROR)
+    logging.getLogger("vllm.distributed.device_communicators.pynccl").setLevel(logging.WARN)
+    logging.getLogger("vllm.selector").setLevel(logging.WARN)
+    logging.getLogger("vllm.utils").setLevel(logging.WARN)
 
 
 def assert_pkg_version(pkg: str, min_version: str):