fix: 整理和改正 attention 构造问题

Signed-off-by: YdrMaster <[email protected]>

Author: YdrMaster

src/05computation/include/computation/operators/attention.h

@@ -6,14 +6,14 @@
 namespace refactor::computation {
 
     struct Attention final : public Operator {
-        dim_t maxSeqLen;
 
-        constexpr Attention(decltype(maxSeqLen) maxSeqLen_) noexcept
-            : Operator(), maxSeqLen(maxSeqLen_) {}
+        constexpr Attention() noexcept = default;
 
         static size_t typeId() noexcept;
         size_t opTypeId() const noexcept final;
         std::string_view name() const noexcept final;
+        kernel::CollectorBox candidateKernels(Target) const final;
+        std::string serialize() const noexcept final;
     };
 
 }// namespace refactor::computation

src/05computation/src/operators/attention.cc

@@ -1,4 +1,5 @@
 #include "computation/operators/attention.h"
+#include "kernel/collectors/attention.h"
 
 namespace refactor::computation {
     using Op = Attention;
@@ -9,5 +10,12 @@ namespace refactor::computation {
     }
     auto Op::opTypeId() const noexcept -> size_t { return typeId(); }
     auto Op::name() const noexcept -> std::string_view { return "Attention"; }
+    auto Op::candidateKernels(Target target) const -> kernel::CollectorBox {
+        using Collector_ = kernel::AttentionCollector;
+        return std::make_unique<Collector_>(target);
+    }
+    auto Op::serialize() const noexcept -> std::string {
+        return "Attention()";
+    }
 
 }// namespace refactor::computation

src/08-01llm/src/operators/attention.cc

@@ -9,7 +9,7 @@ namespace refactor::llm {
         : Operator(), maxSeqLen(maxSeqLen_) {}
 
     auto Op::build(ModelContext const &, std::string_view, Attributes attributes) -> OpBox {
-        auto maxSeqLen = attributes.getOrInsert("max_seq_len", {0}).float_();
+        auto maxSeqLen = attributes.getOrInsert("max_seq_len", {0}).int_();
         return OpBox(std::make_unique<Op>(maxSeqLen));
     }
     auto Op::typeId() -> size_t {
@@ -129,7 +129,7 @@ namespace refactor::llm {
 
     auto Op::lower(TensorRefs) const -> computation::OpBox {
         using Op_ = computation::Attention;
-        return std::make_unique<Op_>(maxSeqLen);
+        return std::make_unique<Op_>();
     }
 
 }// namespace refactor::llm