minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[CPU] Fix argument issues in qkv_proj_with_rope_fused_weight and bmm… (#21367)

Browse Source 
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
This commit is contained in:
blzheng
2026-04-13 09:59:13 +08:00
committed by GitHub
co-authored by Ma Mingfei
parent da6b8e1448
commit 934e19a610
2 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/cpu_graph_runner.py
+1
View File
@@ -222,6 +222,7 @@ def register_fake_ops():
use_fp8_w8a16, use_fp8_w8a16,
qkv_a_proj_scale, qkv_a_proj_scale,
q_b_proj_scale, q_b_proj_scale,
w_scale,
is_vnni, is_vnni,
block_size, block_size,
q_lora_rank, q_lora_rank,
python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_fused_rope_cpu.py
+2 -2
View File
@@ -100,7 +100,7 @@ class DeepseekMLACpuForwardMixin:
else None else None
) )
), ),
self.w_scale, self.w_scale if self.qkv_proj_with_rope_is_fp8 else None,
True, # is_vnni True, # is_vnni
self.weight_block_size, self.weight_block_size,
self.q_lora_rank, self.q_lora_rank,
@@ -145,7 +145,7 @@ class DeepseekMLACpuForwardMixin:
attn_output.transpose(0, 1), attn_output.transpose(0, 1),
self.w_vc, self.w_vc,
True, # is_vnni True, # is_vnni
self.w_scale, # scale self.w_scale if self.qkv_proj_with_rope_is_fp8 else None, # scale
) )
attn_output = output attn_output = output
output, _ = self.o_proj(attn_output) output, _ = self.o_proj(attn_output)