已合并
Revert "_npu_paged_attention_splitfuse_v2的context_lens输入添加host数据绑定" #5457
hellokitty911创建于 7月13日
Revert "_npu_paged_attention_splitfuse_v2的context_lens输入添加host数据绑定" #5457
已合并
hellokitty911创建于 7月13日
1 个文件变更+0-6
Mop_plugin/ops/atb/PagedAttentionSplitfuseV2.cpp+0-6
@@ -18,9 +18,6 @@
18 18 
19using namespace std;19using namespace std;
20 20 
21-namespace {
22- constexpr int64_t CONTEXT_LENS_INDEX = 4;
23-}
24namespace atb {21namespace atb {
25using PagedAttentionParam = atb::infer::PagedAttentionParam;22using PagedAttentionParam = atb::infer::PagedAttentionParam;
26 23 
@@ -69,9 +66,6 @@ void _npu_paged_attention_splitfuse_v2(const at::Tensor &query, const at::Tensor
69 .Input(mask)66 .Input(mask)
70 .Input(seq_len, true)67 .Input(seq_len, true)
71 .Output(out);68 .Output(out);
72- at::Tensor context_lens_cpu = context_lens.cpu();
73- paramsetter.variant_pack_.inTensors[CONTEXT_LENS_INDEX].hostData = context_lens_cpu.data_ptr();
74- paramsetter.tensor_maintainer_.cpu_tensors.emplace_back(std::move(context_lens_cpu));
75 auto opPaged = pagedAttentionParamCache.getOperation(pagedparam, "PagedAttentionOperation");69 auto opPaged = pagedAttentionParamCache.getOperation(pagedparam, "PagedAttentionOperation");
76 RunAtbCmd(opPaged, paramsetter, "PagedAttentionOperation");70 RunAtbCmd(opPaged, paramsetter, "PagedAttentionOperation");
77 71