Merge tag 'v0.9.0' into v0.9.0-ori

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -6,13 +6,15 @@ 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)
+    per_token_group_quant_fp8,
+    w8a8_block_fp8_matmul,
+)
+from vllm.triton_utils import triton
 
 
 # Copied from

csrc/activation_kernels.cu

@@ -70,6 +70,9 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   int64_t num_tokens = input.numel() / input.size(-1);                   \
   dim3 grid(num_tokens);                                                 \
   dim3 block(std::min(d, 1024));                                         \
+  if (num_tokens == 0) {                                                 \
+    return;                                                              \
+  }                                                                      \
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
   VLLM_DISPATCH_FLOATING_TYPES(                                          \