[Fix] Prevent memory leak in MLA decode graph (#3743)

### What this PR does / why we need it?
The cache for MLA decode graph parameters was holding strong references
to tensors, preventing them from being garbage collected and leading to
increased memory usage.

This change wraps the cached tensors in weak references, allowing them
to be deallocated when no longer in use and reducing overall memory
pressure.

### Does this PR introduce _any_ user-facing change?
None.

### How was this patch tested?
None.

- vLLM version: v0.11.0rc3
- vLLM main:
vllm-project/vllm@c9461e0

---------

Signed-off-by: Yizhou Liu <[email protected]>

Author: yiz-liu

vllm_ascend/attention/attention_v1.py

@@ -562,7 +562,8 @@ def _forward_decode_only(
                             block_table=attn_metadata.block_tables,
                             context_lens=attn_metadata.seq_lens,
                             out=output)
-                        update_graph_params_workspaces(num_tokens, workspace)
+                        update_graph_params_workspaces(
+                            num_tokens, weak_ref_tensors(workspace))
 
                 # Handle graph capturing mode
                 stream = torch_npu.npu.current_stream()
@@ -578,7 +579,7 @@ def _forward_decode_only(
                     self.num_kv_heads,
                     self.num_heads,
                     self.scale,
-                    weak_ref_tensors(attn_metadata.block_tables),
+                    attn_metadata.block_tables,
                     attn_metadata.seq_lens,
                     weak_ref_tensors(output),
                 ))

vllm_ascend/attention/mla_v1.py

@@ -12,8 +12,6 @@
                                               AttentionMetadata,
                                               MLAAttentionImpl)
 from vllm.config import VllmConfig, get_current_vllm_config
-
-# isort: off
 from vllm.distributed import (get_dcp_group,
                               get_decode_context_model_parallel_rank,
                               get_decode_context_model_parallel_world_size,
@@ -35,19 +33,22 @@
                                          split_decodes_and_prefills,
                                          trans_rope_weight, transdata,
                                          wait_for_kv_layer_from_connector)
-from vllm_ascend.compilation.acl_graph import get_graph_params
+from vllm_ascend.compilation.acl_graph import (get_graph_params,
+                                               update_graph_params_workspaces)
 from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
 from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
-                               is_enable_nz, prefill_context_parallel_enable)
+                               is_enable_nz, prefill_context_parallel_enable,
+                               weak_ref_tensors)
 from vllm_ascend.worker.npu_input_batch import InputBatch
 
+# isort: off
 if prefill_context_parallel_enable():
     from vllm.distributed import (get_pcp_group,
                                   get_prefill_context_model_parallel_rank,
                                   get_prefill_context_model_parallel_world_size
                                   )
-# isort:on
+# isort: on
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
 
@@ -743,7 +744,7 @@ def get_and_maybe_dequant_weights(layer: LinearBase):
                 getattr(self.fused_qkv_a_proj.quant_method, 'quant_method',
                         None), AscendW8A8LinearMethod):
             self.enable_mlapo = False
-            logger.warning(
+            logger.warning_once(
                 "Currently mlapo only supports W8A8 quantization in MLA scenario."
                 "Some layers in your model are not quantized with W8A8,"
                 "thus mlapo is disabled for these layers.")
@@ -1115,19 +1116,22 @@ def _forward_decode(
             if workspace is None:
                 workspace = torch_npu._npu_fused_infer_attention_score_get_max_workspace(
                     q_nope, k_nope, k_nope, **common_kwargs)
-                graph_params.workspaces[num_tokens] = workspace
+                update_graph_params_workspaces(num_tokens,
+                                               weak_ref_tensors(workspace))
 
             attn_output = torch.empty_like(q_nope)
             softmax_lse = torch.empty(num_tokens,
                                       dtype=q_nope.dtype,
                                       device=q_nope.device)
 
             graph_params.attn_params[num_tokens].append(
-                (q_nope, k_nope, q_pe, k_pe, self.num_heads, self.num_kv_heads,
-                 input_layout, spec_attn_mask, sparse_mode, self.scale,
-                 decode_meta.block_table, block_size,
-                 decode_meta.seq_lens_list, actual_seq_lengths, workspace,
-                 attn_output, softmax_lse))
+                (weak_ref_tensors(q_nope), weak_ref_tensors(k_nope),
+                 weak_ref_tensors(q_pe), weak_ref_tensors(k_pe),
+                 self.num_heads, self.num_kv_heads, input_layout,
+                 weak_ref_tensors(spec_attn_mask) if spec_attn_mask is not None
+                 else None, sparse_mode, self.scale, decode_meta.block_table,
+                 block_size, decode_meta.seq_lens_list, actual_seq_lengths,
+                 weak_ref_tensors(attn_output), weak_ref_tensors(softmax_lse)))
 
             torch.npu.graph_task_group_begin(stream)
             torch_npu.npu_fused_infer_attention_score.out(

vllm_ascend/compilation/acl_graph.py

@@ -212,7 +212,6 @@ def update_attn_params(update_stream, forward_context, runtime_shape):
             seq_lens,
             output,
         ) = param
-        # block_table = forward_context.attn_metadata[key].block_tables
         seq_lens = forward_context.attn_metadata[key].seq_lens
         torch_npu_check = version_check()
 
@@ -258,8 +257,7 @@ def update_mla_attn_params(update_stream, forward_context, runtime_shape,
     ):
         (q_nope, k_nope, q_pe, k_pe, num_heads, num_kv_heads, input_layout,
          spec_attn_mask, sparse_mode, scale, block_table, block_size,
-         seq_lens_list, actual_seq_lengths, workspace, attn_output,
-         softmax_lse) = param
+         seq_lens_list, actual_seq_lengths, attn_output, softmax_lse) = param
         seq_lens_list = forward_context.attn_metadata[key].decode.seq_lens_list
         if speculative_config and speculative_config.method == "deepseek_mtp":
             actual_seq_lengths = forward_context.attn_metadata[
@@ -295,7 +293,7 @@ def update_mla_attn_params(update_stream, forward_context, runtime_shape,
                 block_size=block_size,
                 actual_seq_lengths_kv=seq_lens_list,
                 actual_seq_lengths=actual_seq_lengths,
-                workspace=workspace,
+                workspace=graph_params.workspaces.get(runtime_shape),
                 out=[attn_output, softmax_lse])
             torch.npu.graph_task_update_end(update_stream)
 
@@ -329,7 +327,7 @@ def set_graph_params(aclgraph_capture_sizes: set[int]):
     )
 
 
-def update_graph_params_workspaces(num_tokens: int, workspace: int):
+def update_graph_params_workspaces(num_tokens: int, workspace: Any):
     global _graph_params
     if _graph_params is not None:
         _graph_params.workspaces[num_tokens] = workspace

vllm_ascend/utils.py

@@ -697,6 +697,13 @@ def weak_ref_tensors(
     """
     Convenience function to create weak references to tensors,
     for single tensor, list of tensors or tuple of tensors.
+
+    This function should be used in the following scenario:
+    When a tensor is created during graph capture, and it's held by a method
+    that's not part of the graph, we don't really need to store it, but we
+    **do need** its buffer pointer. If we don't handle this, it cannot
+    be garbage collected, leading to a memory leak. To avoid this,
+    we should create a weak reference to the tensor.
     """
     if isinstance(tensors, torch.Tensor):
         return weak_ref_tensor(tensors)