[Bugfix] Fix triton import with local TritonPlaceholder (#17446)

Signed-off-by: Mengqing Cao <cmq0113@163.com>

benchmarks/kernels/benchmark_moe.py

@@ -10,12 +10,12 @@ from typing import Any, TypedDict
 
 import ray
 import torch
-import triton
 from ray.experimental.tqdm_ray import tqdm
 from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
+from vllm.triton_utils import triton
 from vllm.utils import FlexibleArgumentParser
 
 FP8_DTYPE = current_platform.fp8_dtype()

benchmarks/kernels/benchmark_rmsnorm.py

@@ -4,11 +4,11 @@ import itertools
 from typing import Optional, Union
 
 import torch
-import triton
 from flashinfer.norm import fused_add_rmsnorm, rmsnorm
 from torch import nn
 
 from vllm import _custom_ops as vllm_ops
+from vllm.triton_utils import triton
 
 
 class HuggingFaceRMSNorm(nn.Module):

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -6,13 +6,13 @@ import time
 # Import DeepGEMM functions
 import deep_gemm
 import torch
-import triton
 from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor
 
 # Import vLLM functions
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8, w8a8_block_fp8_matmul)
+from vllm.triton_utils import triton
 
 
 # Copied from