[Track] progress in removing vLLM dependencies

[zhyncs]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

ref #1673

Create a separate issue to follow up on the two currently important PRs:

weight loader: #2220 @HandH1998
distributed: #2244 @yizhang2077

Related resources

No response

[zhyncs]

ref #2318

[zhyncs]

TLDR
Currently, aside from quantification, the main things left are distributed and rope.
cc @HandH1998 @yizhang2077 @james-p-xu @ispobock @merrymercy @Ying1123

current main

distributed/parallel_state.py
1261:    import vllm.distributed.parallel_state as vllm_parrlel_state

models/mllama.py
11:import vllm.distributed.parallel_state as ps

utils.py
497:    import vllm.distributed.device_communicators.custom_all_reduce_utils as tgt

layers/logits_processor.py
21:from vllm.distributed import (

layers/custom_op_util.py
15:from vllm.model_executor.custom_op import CustomOp

layers/layernorm.py
32:from vllm.model_executor.custom_op import CustomOp
128:    from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm

layers/linear.py
10:from vllm.distributed import (
20:from vllm.model_executor.layers.linear import LinearBase
21:from vllm.model_executor.parameter import (

layers/vocab_parallel_embedding.py
9:from vllm.distributed import (
15:from vllm.model_executor.parameter import BasevLLMParameter

layers/fused_moe_triton/fused_moe.py
14:from vllm import _custom_ops as ops

layers/fused_moe_triton/layer.py
8:from vllm.distributed import (
13:from vllm.model_executor.custom_op import CustomOp

layers/activation.py
28:from vllm.distributed import (
33:from vllm.model_executor.custom_op import CustomOp
156:    from vllm.model_executor.layers.activation import GeluAndMul, SiluAndMul

layers/quantization/__init__.py
6:from vllm.model_executor.layers.quantization.aqlm import AQLMConfig
7:from vllm.model_executor.layers.quantization.awq import AWQConfig
8:from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
9:from vllm.model_executor.layers.quantization.bitsandbytes import BitsAndBytesConfig
10:from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (
13:from vllm.model_executor.layers.quantization.deepspeedfp import DeepSpeedFPConfig
14:from vllm.model_executor.layers.quantization.experts_int8 import ExpertsInt8Config
15:from vllm.model_executor.layers.quantization.fbgemm_fp8 import FBGEMMFp8Config
16:from vllm.model_executor.layers.quantization.fp8 import Fp8Config, Fp8MoEMethod
17:from vllm.model_executor.layers.quantization.gguf import GGUFConfig
18:from vllm.model_executor.layers.quantization.gptq import GPTQConfig
19:from vllm.model_executor.layers.quantization.gptq_marlin import GPTQMarlinConfig
20:from vllm.model_executor.layers.quantization.gptq_marlin_24 import GPTQMarlin24Config
21:from vllm.model_executor.layers.quantization.marlin import MarlinConfig
22:from vllm.model_executor.layers.quantization.qqq import QQQConfig
23:from vllm.model_executor.layers.quantization.tpu_int8 import Int8TpuConfig
102:    from vllm.model_executor.layers.linear import LinearBase
103:    from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
104:    from vllm.model_executor.layers.quantization.utils.quant_utils import (

model_executor/cuda_graph_runner.py
23:from vllm.distributed.parallel_state import graph_capture
24:from vllm.model_executor.custom_op import CustomOp

model_executor/model_runner.py
23:from vllm.distributed import (
29:from vllm.distributed.parallel_state import in_the_same_node_as

models/qwen2_vl.py
33:from vllm.distributed import parallel_state
34:from vllm.distributed import utils as dist_utils
35:from vllm.logger import init_logger
36:from vllm.model_executor.layers.activation import QuickGELU

models/dbrx.py
22:from vllm.distributed import (
27:from vllm.model_executor.layers.rotary_embedding import get_rope
28:from vllm.transformers_utils.configs.dbrx import DbrxConfig

models/gpt2.py
25:from vllm.distributed.parallel_state import get_tensor_model_parallel_world_size
26:from vllm.model_executor.layers.activation import get_act_fn
27:from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding

models/olmo2.py
24:from vllm.distributed import (
30:from vllm.model_executor.layers.rotary_embedding import get_rope
31:from vllm.model_executor.model_loader.weight_utils import default_weight_loader

models/chatglm.py
24:from vllm.distributed import get_tensor_model_parallel_world_size
25:from vllm.model_executor.layers.rotary_embedding import get_rope
26:from vllm.transformers_utils.configs import ChatGLMConfig

models/grok.py
25:from vllm.distributed import get_tensor_model_parallel_world_size
26:from vllm.model_executor.layers.rotary_embedding import get_rope

models/internlm2.py
22:from vllm.distributed import get_tensor_model_parallel_world_size
23:from vllm.model_executor.layers.rotary_embedding import get_rope

models/llama.py
25:from vllm.distributed import get_tensor_model_parallel_world_size
26:from vllm.model_executor.layers.rotary_embedding import get_rope

models/olmoe.py
26:from vllm.distributed import (
30:from vllm.model_executor.layers.linear import (
36:from vllm.model_executor.layers.rotary_embedding import get_rope

models/xverse.py
24:from vllm.distributed import get_tensor_model_parallel_world_size
25:from vllm.model_executor.layers.activation import SiluAndMul
26:from vllm.model_executor.layers.layernorm import RMSNorm
27:from vllm.model_executor.layers.linear import (
32:from vllm.model_executor.layers.rotary_embedding import get_rope

models/minicpm.py
21:from vllm.distributed import get_tensor_model_parallel_world_size
22:from vllm.model_executor.layers.rotary_embedding import get_rope

models/deepseek.py
24:from vllm.distributed import (
29:from vllm.model_executor.layers.rotary_embedding import get_rope

models/mllama.py
17:from vllm.distributed import get_tensor_model_parallel_world_size

models/stablelm.py
27:from vllm.distributed import get_tensor_model_parallel_world_size
28:from vllm.model_executor.layers.rotary_embedding import get_rope

models/gemma2.py
23:from vllm.distributed import get_tensor_model_parallel_world_size
48:from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
146:        # from vLLM: TODO(woosuk): Use the `get_rope` interface.

models/qwen.py
23:from vllm.distributed import get_tensor_model_parallel_world_size
24:from vllm.model_executor.layers.rotary_embedding import get_rope

models/mixtral_quant.py
26:from vllm.distributed import (
31:from vllm.model_executor.layers.rotary_embedding import get_rope

models/minicpm3.py
22:from vllm.distributed import get_tensor_model_parallel_world_size
23:from vllm.model_executor.layers.linear import (
29:from vllm.model_executor.layers.rotary_embedding import get_rope

models/gpt_bigcode.py
24:from vllm.distributed import get_tensor_model_parallel_world_size

models/commandr.py
47:from vllm.distributed import (
51:from vllm.model_executor.layers.rotary_embedding import get_rope

models/torch_native_llama.py
50:from vllm.distributed import (
54:from vllm.model_executor.layers.rotary_embedding import get_rope

models/phi3_small.py
8:from vllm.distributed import get_tensor_model_parallel_world_size
9:from vllm.model_executor.layers.rotary_embedding import get_rope

models/xverse_moe.py
21:from vllm.distributed import (
26:from vllm.model_executor.layers.activation import SiluAndMul
27:from vllm.model_executor.layers.layernorm import RMSNorm
28:from vllm.model_executor.layers.linear import (
34:from vllm.model_executor.layers.rotary_embedding import get_rope

models/gemma.py
24:from vllm.distributed import get_tensor_model_parallel_world_size
25:from vllm.model_executor.layers.rotary_embedding import get_rope

models/baichuan.py
27:from vllm.distributed import (
31:from vllm.model_executor.layers.linear import (
36:from vllm.model_executor.layers.rotary_embedding import get_rope

models/exaone.py
23:from vllm.distributed import get_tensor_model_parallel_world_size
24:from vllm.model_executor.layers.rotary_embedding import get_rope

models/qwen2.py
23:from vllm.distributed import get_tensor_model_parallel_world_size
24:from vllm.model_executor.layers.rotary_embedding import get_rope

models/deepseek_v2.py
24:from vllm.distributed import (
30:from vllm.model_executor.layers.rotary_embedding import get_rope

models/olmo.py
23:from vllm.distributed import get_tensor_model_parallel_world_size
24:from vllm.model_executor.layers.rotary_embedding import get_rope

models/mixtral.py
24:from vllm.distributed import get_tensor_model_parallel_world_size
25:from vllm.model_executor.layers.rotary_embedding import get_rope

models/qwen2_moe.py
25:from vllm.distributed import (
29:from vllm.model_executor.layers.rotary_embedding import get_rope

utils.py
430:    from vllm.logger import logger as vllm_default_logger
502:    from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
513:    from vllm.distributed.parallel_state import GroupCoordinator
554:    from vllm.model_executor.layers.linear import LinearBase
555:    from vllm.model_executor.layers.quantization.gguf import (
792:    from vllm.distributed import get_tensor_model_parallel_rank

model_loader/weight_utils.py
22:from vllm.distributed import get_tensor_model_parallel_rank

lora/lora.py
30:from vllm.model_executor.layers.vocab_parallel_embedding import (

hf_transformers_utils.py
34:    from vllm.transformers_utils.configs import ChatGLMConfig, DbrxConfig

model_loader/loader.py
24:from vllm.distributed import (
499:        from vllm.distributed import get_tensor_model_parallel_rank
559:        from vllm.distributed import get_tensor_model_parallel_rank

[james-p-xu]

Hi @zhyncs, the RoPE PR (#2016) is still blocked by a version upgrade on flashinfer. Happy to discuss this offline whenever you're available. In addition, #2016 also uses the vllm.model_executor.custom_op, which may need to be removed as a separate change.