[runtime][vm] Add LoRA adapter metadata to paged KV cache

src/runtime/vm/attn_utils.h

@@ -180,6 +180,8 @@ struct Sequence {
    * in the KV cache even when sliding window is enabled.
    */
   int last_block_attn_sink_size = 0;
+  /*! \brief The LoRA adapter id associated with the sequence. */
+  int32_t lora_adapter_id = 0;
 
   /*! \brief Whether the current appended tokens form a chain (not a tree). */
   bool is_chain = true;

src/runtime/vm/kv_state.cc

@@ -65,6 +65,12 @@ TVM_FFI_STATIC_INIT_BLOCK() {
                   &AttentionKVCacheObj::EnableSlidingWindowForSeq)
       .def_method("vm.builtin.attention_kv_cache_commit_accepted_token_tree_nodes",
                   &AttentionKVCacheObj::CommitAcceptedTokenTreeNodes)
+      .def_method("vm.builtin.attention_kv_cache_set_sequence_lora_adapter",
+                  &AttentionKVCacheObj::SetSequenceLoraAdapter)
+      .def_method("vm.builtin.attention_kv_cache_get_sequence_lora_adapter",
+                  &AttentionKVCacheObj::GetSequenceLoraAdapter)
+      .def_method("vm.builtin.attention_kv_cache_get_current_lora_adapter_ids",
+                  &AttentionKVCacheObj::GetCurrentLoraAdapterIds)
       .def_method("vm.builtin.attention_kv_cache_empty", &AttentionKVCacheObj::Empty)
       .def_method("vm.builtin.attention_kv_cache_get_num_available_pages",
                   &AttentionKVCacheObj::GetNumAvailablePages)

src/runtime/vm/kv_state.h

@@ -158,6 +158,26 @@ class AttentionKVCacheObj : public KVStateObj {
   virtual void CommitAcceptedTokenTreeNodes(const IntTuple& seq_ids,
                                             const IntTuple& leaf_indices) = 0;
 
+  /*!
+   * \brief Set the LoRA adapter id for the specified sequence.
+   * \param seq_id The id of the sequence to update.
+   * \param lora_adapter_id The LoRA adapter id to associate with the sequence.
+   */
+  virtual void SetSequenceLoraAdapter(int64_t seq_id, int64_t lora_adapter_id) = 0;
+
+  /*!
+   * \brief Get the LoRA adapter id for the specified sequence.
+   * \param seq_id The id of the sequence to query.
+   * \return The LoRA adapter id associated with the sequence.
+   */
+  virtual int64_t GetSequenceLoraAdapter(int64_t seq_id) = 0;
+
+  /*!
+   * \brief Get the LoRA adapter ids of the current batch specified in BeginForward.
+   * \return The LoRA adapter ids, in the same order as the current batch sequence ids.
+   */
+  virtual ffi::Shape GetCurrentLoraAdapterIds() = 0;
+
   /*! \brief Prepare for the disaggregation KV data receive for the specified sequence and length.*/
   virtual IntTuple DisaggPrepareRecv(int64_t seq_id, int length) = 0;
 

src/runtime/vm/paged_kv_cache.cc

@@ -206,6 +206,7 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
   std::vector<HostMemoryVector> k_rope_pos_offset_on_depths_host_;
   std::vector<HostMemoryVector> k_rope_pos_offset_sliding_window_on_depths_host_;
   HostMemoryVector k_ragged_rope_pos_offset_host_;
+  HostMemoryVector current_lora_adapter_ids_host_;
   HostMemoryVector q_rope_position_map_host_;
   HostMemoryVector append_position_map_host_;
   HostMemoryVector cur_append_lengths_indptr_host_;
@@ -414,6 +415,8 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
     }
     k_ragged_rope_pos_offset_host_ =
         HostMemoryVector(reserved_num_seqs, dtype_aux_, preferred_host_device);
+    current_lora_adapter_ids_host_ =
+        HostMemoryVector(reserved_num_seqs, dtype_aux_, preferred_host_device);
     q_rope_position_map_host_ =
         HostMemoryVector(prefill_chunk_size, dtype_aux_, preferred_host_device);
     append_position_map_host_ =
@@ -685,7 +688,10 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
       break;
     }
     // Create the child sequence with the child block.
-    seq_map_.insert({child_seq_id, Sequence(&global_block_pool_, child_block_idx)});
+    auto [child_it, inserted] =
+        seq_map_.insert({child_seq_id, Sequence(&global_block_pool_, child_block_idx)});
+    TVM_FFI_ICHECK(inserted);
+    child_it->second.lora_adapter_id = parent_it->second.lora_adapter_id;
     dirty_aux_data_device_ = true;
   }
 
@@ -874,11 +880,13 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
     sequences.reserve(cur_batch_size_);
     last_block_length_before_append.reserve(cur_batch_size_);
     k_ragged_rope_pos_offset_host_.clear();
+    current_lora_adapter_ids_host_.clear();
     for (int i = 0; i < cur_batch_size_; ++i) {
       auto it = seq_map_.find(seq_ids[i]);
       TVM_FFI_ICHECK(it != seq_map_.end())
           << "The sequence \"" << seq_ids[i] << "\" cannot be found in KV cache.";
       sequences.push_back(&it->second);
+      current_lora_adapter_ids_host_.push_back(it->second.lora_adapter_id);
       last_block_length_before_append.push_back(
           global_block_pool_[it->second.last_block_idx].seq_length);
       int k_rope_offset = it->second.seq_length;
@@ -1195,6 +1203,27 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
     }
   }
 
+  void SetSequenceLoraAdapter(int64_t seq_id, int64_t lora_adapter_id) final {
+    TVM_FFI_ICHECK(lora_adapter_id >= 0) << "LoRA adapter id must be non-negative.";
+    TVM_FFI_ICHECK(lora_adapter_id <= std::numeric_limits<int32_t>::max())
+        << "LoRA adapter id exceeds int32 range.";
+    auto it = seq_map_.find(seq_id);
+    TVM_FFI_ICHECK(it != seq_map_.end())
+        << "The sequence \"" << seq_id << "\" cannot be found in KV cache.";
+    it->second.lora_adapter_id = static_cast<int32_t>(lora_adapter_id);
+  }
+
+  int64_t GetSequenceLoraAdapter(int64_t seq_id) final {
+    auto it = seq_map_.find(seq_id);
+    TVM_FFI_ICHECK(it != seq_map_.end())
+        << "The sequence \"" << seq_id << "\" cannot be found in KV cache.";
+    return it->second.lora_adapter_id;
+  }
+
+  ffi::Shape GetCurrentLoraAdapterIds() final {
+    return current_lora_adapter_ids_host_.as_int_tuple();
+  }
+
   ffi::Shape DisaggPrepareRecv(int64_t seq_id, int append_length) final {
     // No CPU to GPU copy is needed.
     // Essentially we

tests/python/relax/test_runtime_builtin_paged_attention_kv_cache_cpu.py

@@ -65,6 +65,9 @@
 fbegin_forward = None
 fend_forward = None
 fcommit_accepted_token_tree_nodes = None
+fset_sequence_lora_adapter = None
+fget_sequence_lora_adapter = None
+fget_current_lora_adapter_ids = None
 fattention_with_fuse_qkv = None
 fis_empty = None
 fdebug_get_kv = None
@@ -88,6 +91,7 @@
 def set_global_func(head_dim, dtype):
     global fclear, fadd_sequence, fremove_sequence, ffork_sequence, fenable_sliding_window_for_seq
     global fpopn, fbegin_forward, fend_forward, fcommit_accepted_token_tree_nodes
+    global fset_sequence_lora_adapter, fget_sequence_lora_adapter, fget_current_lora_adapter_ids
     global fattention_with_fuse_qkv, fis_empty, fdebug_get_kv
     global ftranspose_append, fcopy_cache, fattn_prefill, fattn_decode
     global \
@@ -110,6 +114,15 @@ def set_global_func(head_dim, dtype):
     fcommit_accepted_token_tree_nodes = tvm.get_global_func(
         "vm.builtin.attention_kv_cache_commit_accepted_token_tree_nodes"
     )
+    fset_sequence_lora_adapter = tvm.get_global_func(
+        "vm.builtin.attention_kv_cache_set_sequence_lora_adapter"
+    )
+    fget_sequence_lora_adapter = tvm.get_global_func(
+        "vm.builtin.attention_kv_cache_get_sequence_lora_adapter"
+    )
+    fget_current_lora_adapter_ids = tvm.get_global_func(
+        "vm.builtin.attention_kv_cache_get_current_lora_adapter_ids"
+    )
     fattention_with_fuse_qkv = tvm.get_global_func(
         "vm.builtin.attention_kv_cache_attention_with_fused_qkv"
     )
@@ -254,6 +267,27 @@ def verify_cached_kv(kv_cache, seq_ids, expected_k, expected_v):
         tvm.testing.assert_allclose(values.numpy(), values_expected, rtol=1e-3, atol=1e-3)
 
 
+def test_lora_adapter_metadata(kv_cache_and_config):
+    kv_cache, _, _ = kv_cache_and_config
+
+    fadd_sequence(kv_cache, 1)
+    fadd_sequence(kv_cache, 2)
+    assert fget_sequence_lora_adapter(kv_cache, 1) == 0
+    assert fget_sequence_lora_adapter(kv_cache, 2) == 0
+
+    fset_sequence_lora_adapter(kv_cache, 1, 7)
+    fset_sequence_lora_adapter(kv_cache, 2, 11)
+    assert fget_sequence_lora_adapter(kv_cache, 1) == 7
+    assert fget_sequence_lora_adapter(kv_cache, 2) == 11
+
+    ffork_sequence(kv_cache, 1, 3, -1)
+    assert fget_sequence_lora_adapter(kv_cache, 3) == 7
+
+    fbegin_forward(kv_cache, ShapeTuple([2, 3, 1]), ShapeTuple([1, 2, 1]))
+    assert list(fget_current_lora_adapter_ids(kv_cache)) == [11, 7, 7]
+    fend_forward(kv_cache)
+
+
 def f_apply_rotary(x, offset, scale, theta, offset_list: list[int] | None = None):
     # x: (N, H, D)
     assert len(x.shape) == 3