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

add new accuracy test case for aclgraph (#3390)

Browse Source 
### What this PR does / why we need it?
Add new accuracy test case Deepseek-V2-Lite-W8A8 for aclgraph

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?
ut

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

---------

Signed-off-by: lilinsiman <lilinsiman@gmail.com>
This commit is contained in:
lilinsiman
2025-10-20 20:04:04 +08:00
committed by GitHub
parent b9e2896eb1
commit 70bef33f13
3 changed files with 83 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm_ascend/attention/mla_v1.py
View File

@@ -976,17 +976,20 @@ class AscendMLAImpl(MLAAttentionImpl):
# Use TND layout for pure SpecDecoding and SpecDecoding in ChunkedPrefill
input_layout = "TND"
# [bs * q_seq_len, num_heads_per_rank, dim]
q_nope = q_nope.view(num_tokens, self.num_heads, -1)
# TODO: If the driver is upgraded later, the contiguous function can be deleted.
q_nope = q_nope.view(num_tokens, self.num_heads, -1).contiguous()
q_pe = q_pe.view(num_tokens, self.num_heads, -1)
sparse_mode = 3
spec_attn_mask = attn_metadata.decode.attn_mask # type:ignore
actual_seq_lengths = decode_meta.actual_seq_lengths_q
else:
if self.enable_kv_nz:
q_nope = q_nope.view(num_tokens, 1, self.num_heads, -1)
q_nope = q_nope.view(num_tokens, 1, self.num_heads,
-1).contiguous()
q_pe = q_pe.view(num_tokens, 1, self.num_heads, -1)
else:
q_nope = q_nope.view(num_tokens, self.num_heads, 1, -1)
q_nope = q_nope.view(num_tokens, self.num_heads, 1,
-1).contiguous()
q_pe = q_pe.view(num_tokens, self.num_heads, 1, -1)
sparse_mode = 0
spec_attn_mask = None