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

[Kernel/Quant] Remove the original marlin format and qqq (#23204)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-08-20 15:13:36 -04:00
committed by GitHub
parent ebe56a0064
commit 0cdbf5e61c
26 changed files with 92 additions and 3698 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
csrc/torch_bindings.cpp
View File

@@ -241,14 +241,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
// custom types:
// https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA
// Marlin (Dense) Optimized Quantized GEMM for GPTQ.
ops.def(
"marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
"Tensor! workspace, SymInt size_m, SymInt size_n, SymInt size_k) -> "
"Tensor",
{stride_tag});
// conditionally compiled so impl in source file
// Marlin_24 (Sparse) Optimized Quantized GEMM for GPTQ.
ops.def(
"gptq_marlin_24_gemm(Tensor a, Tensor b_q_weight, Tensor b_meta, "
@@ -353,15 +345,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
ops.def("ggml_moe_get_block_size", &ggml_moe_get_block_size);
#ifndef USE_ROCM
// marlin_qqq_gemm for QQQ.
ops.def(
"marlin_qqq_gemm(Tensor a, Tensor b_q_weight, "
"Tensor s_tok, Tensor s_ch, Tensor s_group, "
"Tensor! workspace, SymInt size_m, SymInt size_n, "
"SymInt size_k) -> Tensor",
{stride_tag});
// conditionally compiled so impl registration is in source file
// CUTLASS nvfp4 block scaled GEMM
ops.def(
"cutlass_scaled_fp4_mm(Tensor! out, Tensor a, Tensor b,"