EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix wrong gemm branch cause 250us slower (#7969)

Browse Source
This commit is contained in:
fzyzcjy
2025-07-12 10:45:09 +08:00
committed by GitHub
parent aa2056091a
commit 2a2d3478af
1 changed files with 1 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/models/deepseek_v2.py
View File

@@ -2193,7 +2193,6 @@ class DeepseekV2ForCausalLM(nn.Module):
# This may affect the accuracy of fp8 model. # This may affect the accuracy of fp8 model.
# Fix deepseek v3 blockwise bmm by using deep_gemm # Fix deepseek v3 blockwise bmm by using deep_gemm
use_deep_gemm_bmm = False use_deep_gemm_bmm = False
model_dtype = torch.get_default_dtype()
if w.dtype in ( if w.dtype in (
torch.float8_e4m3fn, torch.float8_e4m3fn,
@@ -2219,7 +2218,6 @@ class DeepseekV2ForCausalLM(nn.Module):
_is_cuda _is_cuda
and weight_block_size[0] == 128 and weight_block_size[0] == 128
and weight_block_size[1] == 128 and weight_block_size[1] == 128
and model_dtype == torch.bfloat16
): ):
if ( if (
deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
@@ -2233,7 +2231,7 @@ class DeepseekV2ForCausalLM(nn.Module):
weight, weight,
weight_scale, weight_scale,
weight_block_size, weight_block_size,
model_dtype, torch.bfloat16,
) )
else: else:
w, scale = block_quant_to_tensor_quant( w, scale = block_quant_to_tensor_quant(