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

[2/n]decouple quantization implementation from vLLM dependency (#8112)

Browse Source 
Co-authored-by: walker-ai <yiyun.wyt@antgroup.com>
Co-authored-by: leoneo <1320612015@qq.com>
This commit is contained in:
Peng Zhang
2025-08-14 18:19:03 +08:00
committed by GitHub
parent 4dbf43601d
commit 5aa1ebd242
32 changed files with 6506 additions and 202 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
sgl-kernel/include/scalar_type.hpp
View File

@@ -298,6 +298,7 @@ static inline constexpr auto kS8 = ScalarType::int_(8);
static inline constexpr auto kU8 = ScalarType::uint(8);
static inline constexpr auto kU8B128 = ScalarType::uint(8, 128);
static inline constexpr auto kFE2M1f = ScalarType::float_(2, 1, true, ScalarType::NAN_NONE);
static inline constexpr auto kFE3M2f = ScalarType::float_(3, 2, true, ScalarType::NAN_NONE);
static inline constexpr auto kFE4M3fn = ScalarType::float_(4, 3, true, ScalarType::NAN_EXTD_RANGE_MAX_MIN);
static inline constexpr auto kFE5M2 = ScalarType::float_IEEE754(5, 2);
@@ -313,6 +314,7 @@ static inline constexpr auto kInt8 = kS8;
static inline constexpr auto kUint8 = kU8;
static inline constexpr auto kUint8b128 = kU8B128;
static inline constexpr auto kFloat4_e2m1f = kFE2M1f;
static inline constexpr auto kFloat6_e3m2f = kFE3M2f;
static inline constexpr auto kFloat8_e4m3fn = kFE4M3fn;
static inline constexpr auto kFloat8_e5m2 = kFE5M2;

43
sgl-kernel/include/sgl_kernel_ops.h
View File

@@ -224,6 +224,40 @@ void dsv3_router_gemm(torch::Tensor& output, const torch::Tensor& mat_a, const t
void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a, torch::Tensor const& mat_b);
torch::Tensor gptq_marlin_gemm(
torch::Tensor& a,
std::optional<torch::Tensor> c_or_none,
torch::Tensor& b_q_weight,
torch::Tensor& b_scales,
std::optional<torch::Tensor> const& global_scale_or_none,
std::optional<torch::Tensor> const& b_zeros_or_none,
std::optional<torch::Tensor> const& g_idx_or_none,
std::optional<torch::Tensor> const& perm_or_none,
torch::Tensor& workspace,
sglang::ScalarTypeId const& b_q_type_id,
int64_t size_m,
int64_t size_n,
int64_t size_k,
bool is_k_full,
bool use_atomic_add,
bool use_fp32_reduce,
bool is_zp_float);
torch::Tensor gptq_gemm(
torch::Tensor a,
torch::Tensor b_q_weight,
torch::Tensor b_gptq_qzeros,
torch::Tensor b_gptq_scales,
torch::Tensor b_g_idx,
bool use_shuffle,
int64_t bit);
void gptq_shuffle(torch::Tensor q_weight, torch::Tensor q_perm, int64_t bit);
torch::Tensor
gptq_marlin_repack(torch::Tensor& b_q_weight, torch::Tensor& perm, int64_t size_k, int64_t size_n, int64_t num_bits);
torch::Tensor awq_marlin_repack(torch::Tensor& b_q_weight, int64_t size_k, int64_t size_n, int64_t num_bits);
/*
* From csrc/moe
*/
@@ -340,15 +374,6 @@ void scaled_fp4_experts_quant(
torch::Tensor const& input_offset_by_experts,
torch::Tensor const& output_scale_offset_by_experts);
namespace marlin_moe_wna16 {
torch::Tensor
gptq_marlin_repack(torch::Tensor& b_q_weight, torch::Tensor& perm, int64_t size_k, int64_t size_n, int64_t num_bits);
torch::Tensor awq_marlin_repack(torch::Tensor& b_q_weight, int64_t size_k, int64_t size_n, int64_t num_bits);
} // namespace marlin_moe_wna16
/*
* From csrc/speculative
*/