[Hardware/NVIDIA/Kernel] Enable nvidia/DeepSeek-R1-FP4 Model (#16362)

This commit is contained in:
Pavani Majety
2025-05-09 16:24:41 -07:00
committed by GitHub
parent 3b602cdea7
commit 0c0fdae84f
16 changed files with 1994 additions and 112 deletions
+15
View File
@@ -363,6 +363,14 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
{stride_tag});
ops.impl("cutlass_scaled_fp4_mm", torch::kCUDA, &cutlass_scaled_fp4_mm);
// cutlass nvfp4 block scaled group GEMM
ops.def(
"cutlass_fp4_group_mm(Tensor! out, Tensor a, Tensor b,"
" Tensor a_blockscale, Tensor b_blockscales, Tensor alphas,"
" Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()",
{stride_tag});
ops.impl("cutlass_fp4_group_mm", torch::kCUDA, &cutlass_fp4_group_mm);
// CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
// quantization, as well as bias
ops.def(
@@ -492,6 +500,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" Tensor! output_scale, Tensor input_scale) -> ()");
ops.impl("scaled_fp4_quant", torch::kCUDA, &scaled_fp4_quant);
// Compute NVFP4 experts quantization.
ops.def(
"scaled_fp4_experts_quant(Tensor! output, Tensor! output_scale,"
"Tensor input, Tensor input_global_scale, Tensor input_offset_by_experts,"
"Tensor output_scale_offset_by_experts) -> ()");
ops.impl("scaled_fp4_experts_quant", torch::kCUDA, &scaled_fp4_experts_quant);
// Check if cutlass_scaled_mm_fp4 is supported for CUDA devices
// of the given capability
ops.def("cutlass_scaled_mm_supports_fp4(int cuda_device_capability) -> bool");