WIP

smirnov-alexey · Jan 9, 2025 · e2c2d38 · e2c2d38
1 parent 1186e32
commit e2c2d38
Show file tree

Hide file tree

Showing 4 changed files with 14 additions and 20 deletions.
diff --git a/src/plugins/intel_npu/src/plugin/npuw/compiled_model.cpp b/src/plugins/intel_npu/src/plugin/npuw/compiled_model.cpp
@@ -559,7 +559,12 @@ void ov::npuw::CompiledModel::CompiledModelDesc::serialize(
         write(stream, is_remote);
         write_weightless(stream, scales, const_to_offset);
         write_weightless(stream, zerops, const_to_offset);
-        write_weightless_closure(stream, closure, closure_uid, const_to_offset);
+
+        // !!! FIXME !!!
+        // to serialize closures we need:
+        // 1) serialize CPU closures as scales/zerops
+        // 2) rest should be serialized from lazy_tensor
+        // 3) then LT should be evaluated and allocated (use bank for that) during deserialization
     }
 
     LOG_DEBUG("DONE.");
@@ -632,7 +637,12 @@ ov::npuw::CompiledModel::CompiledModelDesc ov::npuw::CompiledModel::CompiledMode
 
         read_weightless(stream, desc.scales, weights_stream);
         read_weightless(stream, desc.zerops, weights_stream);
-        read_weightless_closure(stream, desc.closure, weights_stream);
+
+        // !!! FIXME !!!
+        // to deserialize closures we need:
+        // 1) deserialize CPU closures as scales/zerops
+        // 2) rest should be deserialized via lazy_tensor
+        // 3) then LT should be evaluated and allocated (use bank for that)
     }
 
     LOG_DEBUG("DONE.");

diff --git a/src/plugins/intel_npu/src/plugin/npuw/serialization.cpp b/src/plugins/intel_npu/src/plugin/npuw/serialization.cpp
@@ -208,11 +208,6 @@ void ov::npuw::s11n::write_weightless(std::ostream& stream,
     }
 }
 
-void ov::npuw::s11n::write_weightless_closure(std::ostream& stream,
-                                              const std::vector<ov::Tensor>& var,
-                                              const std::vector<int64_t>& uids,
-                                              const std::unordered_map<const void*, std::size_t>& const_to_offset) {}
-
 void ov::npuw::s11n::read_weightless(std::istream& stream,
                                      std::vector<ov::Tensor>& var,
                                      std::ifstream& weights_stream) {
@@ -251,7 +246,3 @@ void ov::npuw::s11n::read_weightless(std::istream& stream,
         }
     }
 }
-
-void ov::npuw::s11n::read_weightless_closure(std::istream& stream,
-                                             std::vector<ov::Tensor>& var,
-                                             std::ifstream& weights_stream) {}
diff --git a/src/plugins/intel_npu/src/plugin/npuw/serialization.hpp b/src/plugins/intel_npu/src/plugin/npuw/serialization.hpp
@@ -10,8 +10,8 @@
 #include <optional>
 #include <string>
 #include <tuple>
-#include <unordered_set>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 const constexpr char* NPUW_NAME_IDENTIFIER = "NPUW_serialized_";
@@ -68,14 +68,7 @@ void read(std::istream& stream, std::vector<std::shared_ptr<ov::Node>>& var);
 void write_weightless(std::ostream& stream,
                       const std::vector<ov::Tensor>& var,
                       const std::unordered_map<const void*, std::size_t>& const_to_offset);
-void write_weightless_closure(std::ostream& stream,
-                              const std::vector<ov::Tensor>& var,
-                              const std::vector<int64_t>& uids,
-                              const std::unordered_map<const void*, std::size_t>& const_to_offset);
-// No allocation needed
 void read_weightless(std::istream& stream, std::vector<ov::Tensor>& var, std::ifstream& weights_stream);
-// With allocation where required
-void read_weightless_closure(std::istream& stream, std::vector<ov::Tensor>& var, std::ifstream& weights_stream);
 
 // Forward declaration
 template <typename T1, typename T2>

diff --git a/src/plugins/intel_npu/src/plugin/npuw/weights_bank.cpp b/src/plugins/intel_npu/src/plugin/npuw/weights_bank.cpp
@@ -233,7 +233,7 @@ void Bank::read_and_add_tensor(std::istream& stream, int64_t uid, const std::str
         return;
     }
 
-    // Need to allocate on device and copy deserialized tensor to that memory
+    // Need to allocate on device and read deserialized tensor into that memory
     ov::SoPtr<ov::ITensor> remote_tensor;
     ov::Tensor allocated_tensor;