A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Compile] Fix Compile Warning SM100 Cutlass MLA (#23287)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Aug 20, 2025 at 23:09 UTC
f94bf9b924afe2e720b864590c9798b911e77e66
1 file changed
+2
-2
csrc/attention/mla/sm100_cutlass_mla_kernel.cu
+2
-2
@@ -167,7 +167,7 @@ typename T::Fmha::Arguments args_from_options(
167
// TODO(trevor-m): Change split_kv back to -1 when
168
// https://github.com/NVIDIA/cutlass/issues/2274 is fixed. Split_kv=1 will
169
// perform worse with larger context length and smaller batch sizes.
170
- num_kv_splits, // split_kv
170
+ static_cast<int>(num_kv_splits), // split_kv
171
nullptr, // is_var_split_kv
172
};
173
// TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
@@ -264,7 +264,7 @@ int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_ba
264
// Assumes device 0 when getting sm_count.
265
arguments.hw_info.sm_count =
266
sm_count <= 0 ? cutlass::KernelHardwareInfo::query_device_multiprocessor_count(/*device_id=*/0) : sm_count;
267
- arguments.split_kv = num_kv_splits;
267
+ arguments.split_kv = static_cast<int>(num_kv_splits);
268
MlaSm100Type::Fmha::set_split_kv(arguments);
269
270
return MlaSm100Type::Fmha::get_workspace_size(arguments);