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

[feature] support pcp + mtp in full graph (#4572)

Browse Source 
1. support pcp + mtp in full graph
2. pcp/dcp related mtp bugfix
3. support pcp + mtpx

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: zhangsicheng5 <zhangsicheng5@huawei.com>
This commit is contained in:
zhangsicheng5
2025-12-22 16:13:39 +08:00
committed by GitHub
parent 12d581605b
commit 78aa7f2693
10 changed files with 478 additions and 94 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm_ascend/compilation/acl_graph.py
View File

@@ -440,7 +440,10 @@ def update_attn_dcp_pcp_params(update_stream, forward_context, runtime_shape):
def update_mla_attn_dcp_pcp_params(update_stream, forward_context,
runtime_shape):
graph_params = get_graph_params()
if forward_context.is_mtp_model:
graph_params = get_mtp_graph_params()
else:
graph_params = get_graph_params()
# FIXME: Behold! We are using a temporary hack here to update the args
# for each layer's attention op in the graph.
with torch.npu.stream(update_stream):