biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernel][Quantization] add w4a8 support for marlin kernel (#24722)

Browse Source 
Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Michael Goin <mgoin@redhat.com>
This commit is contained in:
Jinzhen Lin
2025-11-29 23:19:33 +08:00
committed by GitHub
parent fa59fe417f
commit 1656ad3704
46 changed files with 4371 additions and 2240 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu
View File

@@ -437,10 +437,10 @@ struct ComputeTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
for (int n_idx = 0; n_idx < WARP_NITER; ++n_idx) {
#pragma unroll
for (int k_idx = 0; k_idx < 2; ++k_idx) {
FType low16 =
ScalarType<FType>::float2num(C_frag[m_idx][n_idx][k_idx * 2]);
FType high16 =
ScalarType<FType>::float2num(C_frag[m_idx][n_idx][k_idx * 2 + 1]);
FType low16 = MarlinScalarType2<FType>::float2num(
C_frag[m_idx][n_idx][k_idx * 2]);
FType high16 = MarlinScalarType2<FType>::float2num(
C_frag[m_idx][n_idx][k_idx * 2 + 1]);
uint32_t tmp = (reinterpret_cast<uint32_t&>(low16) & 0xffff) |
(reinterpret_cast<uint32_t&>(high16) << 16);
int sts_offset =