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

Fix GPTQMarlinMoE (#7697)

Browse Source
This commit is contained in:
Kyungmin Lee
2025-07-02 14:34:43 +09:00
committed by GitHub
parent 00aec6ad6c
commit b3fa5dc3c8
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/quantization/gptq.py
View File

@@ -344,6 +344,10 @@ class GPTQMarlinConfig(QuantizationConfig):
if (num_bits, sym) not in cls.TYPE_MAP: if (num_bits, sym) not in cls.TYPE_MAP:
return False return False
assert (
VLLM_AVAILABLE
), "vllm is not installed, to use gptq_marlin, please install vllm"
return check_marlin_supported( return check_marlin_supported(
quant_type=cls.TYPE_MAP[(num_bits, sym)], group_size=group_size quant_type=cls.TYPE_MAP[(num_bits, sym)], group_size=group_size
) )
@@ -726,6 +730,6 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
g_idx2=layer.w2_g_idx, g_idx2=layer.w2_g_idx,
sort_indices1=layer.w13_g_idx_sort_indices, sort_indices1=layer.w13_g_idx_sort_indices,
sort_indices2=layer.w2_g_idx_sort_indices, sort_indices2=layer.w2_g_idx_sort_indices,
num_bits=self.quant_config.quant_type.size_bits, quant_type_id=self.quant_config.quant_type.id,
is_k_full=self.is_k_full, is_k_full=self.is_k_full,
).to(orig_dtype) ).to(orig_dtype)