[Kernel] Add ModelOpt FP4 Checkpoint Support (#12520)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>
This commit is contained in:
Pavani Majety
2025-03-12 05:13:11 +00:00
committed by GitHub
parent 5c538c37b2
commit debd6bbf09
10 changed files with 388 additions and 30 deletions
+5 -3
View File
@@ -160,14 +160,16 @@ torch::Tensor ggml_moe_a8(torch::Tensor X, torch::Tensor W,
int64_t ggml_moe_get_block_size(int64_t type);
#ifndef USE_ROCM
bool cutlass_scaled_mm_supports_fp4(int64_t cuda_device_capability);
bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability);
void cutlass_scaled_fp4_mm(torch::Tensor& D, torch::Tensor const& A,
torch::Tensor const& B, torch::Tensor const& A_sf,
torch::Tensor const& B_sf,
torch::Tensor const& alpha);
bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability);
void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b, torch::Tensor const& a_scales,
torch::Tensor const& b_scales,