[Feature]cpu offload connector (#1659)

This PR implements cpu offload connector to enable NPU kv cache offload
to host DRAM.

- vLLM version: v0.10.2
- vLLM main:
5aeb925452

Signed-off-by: lidenghui <lidenghui1110@gmail.com>
Signed-off-by: AlvisGong <gwly0401@163.com>
Signed-off-by: CalvinXKY <kyxiezju@163.com>
Co-authored-by: AlvisGong <gwly0401@163.com>

vllm_ascend/attention/attention_v1.py

@@ -26,53 +26,21 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionLayer, AttentionType)
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.config import VllmConfig
-from vllm.distributed.kv_transfer import (get_kv_transfer_group,
-                                          has_kv_transfer_group,
-                                          is_v1_kv_transfer_group)
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.utils import cdiv, direct_register_custom_op
 from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import AttentionSpec
 
-from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
+from vllm_ascend.attention.utils import (AscendCommonAttentionMetadata,
+                                         maybe_save_kv_layer_to_connector,
+                                         wait_for_kv_layer_from_connector)
 from vllm_ascend.compilation.acl_graph import get_graph_params
 from vllm_ascend.ops.attention import vanilla_chunked_prefill
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, is_310p,
                                nd_to_nz_2d, nd_to_nz_spec)
 
 
-def wait_for_kv_layer_from_connector(layer_name: str):
-    if not has_kv_transfer_group() or not is_v1_kv_transfer_group():
-        return
-
-    connector = get_kv_transfer_group()
-
-    forward_context: ForwardContext = get_forward_context()
-    attn_metadata = forward_context.attn_metadata
-    if attn_metadata is None:
-        return
-    # TODO: assert ascendMetadata
-    connector.wait_for_layer_load(layer_name)
-
-
-def maybe_save_kv_layer_to_connector(
-    layer_name: str,
-    kv_cache_layer: List[torch.Tensor],
-):
-    if not has_kv_transfer_group() or not is_v1_kv_transfer_group():
-        return
-
-    connector = get_kv_transfer_group()
-
-    forward_context: ForwardContext = get_forward_context()
-    attn_metadata = forward_context.attn_metadata
-    if attn_metadata is None:
-        return
-    # TODO: assert ascendMetadata
-    connector.save_kv_layer(layer_name, kv_cache_layer, attn_metadata)
-
-
 class AscendAttentionBackend(AttentionBackend):
     accept_output_buffer: bool = True