EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

fix pagedattention to support fullgraph. (#3436)

Browse Source 
### What this PR does / why we need it?
Calculate in advance the workspace memory size needed for the
PagedAttention operator to avoid deadlocks during resource cleanup. This
PR requires torch_npu version 0920 or newer.
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
This commit is contained in:
XiaoxinWang
2025-10-14 16:10:09 +08:00
committed by GitHub
parent 22a1d91cf5
commit 9eb62935b8
5 changed files with 271 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
vllm_ascend/compilation/acl_graph.py
View File

@@ -18,6 +18,8 @@ from vllm.forward_context import BatchDescriptor, get_forward_context
from vllm.logger import logger
from vllm.platforms import current_platform
from vllm_ascend.attention.utils import version_check
from ..utils import weak_ref_tensors
@@ -212,18 +214,32 @@ def update_attn_params(update_stream, forward_context, runtime_shape):
) = param
# block_table = forward_context.attn_metadata[key].block_tables
seq_lens = forward_context.attn_metadata[key].seq_lens
torch_npu_check = version_check()
with torch.npu.stream(update_stream):
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu._npu_paged_attention(query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output)
if torch_npu_check:
torch_npu._npu_paged_attention(
query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output,
workspace=graph_params.workspaces.get(runtime_shape))
else:
torch_npu._npu_paged_attention(query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
@@ -302,5 +318,11 @@ def set_graph_params(aclgraph_capture_sizes: set[int]):
)
def update_graph_params_workspaces(num_tokens: int, workspace: int):
global _graph_params
if _graph_params is not None:
_graph_params.workspaces[num_tokens] = workspace
def get_graph_params():
return _graph_params