[4/n] Migrate FP4/W4A8 CUTLASS kernels to torch stable ABI (#37503)

Signed-off-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
This commit is contained in:
mikaylagawarecki
2026-03-31 10:21:13 -07:00
committed by GitHub
parent 0dd25a44ea
commit 7c080dd3c5
27 changed files with 1205 additions and 1016 deletions
-44
View File
@@ -152,12 +152,6 @@ void silu_and_mul(torch::Tensor& out, torch::Tensor& input);
void silu_and_mul_quant(torch::Tensor& out, torch::Tensor& input,
torch::Tensor& scale);
#ifndef USE_ROCM
void silu_and_mul_nvfp4_quant(torch::Tensor& out,
torch::Tensor& output_block_scale,
torch::Tensor& input,
torch::Tensor& input_global_scale);
#endif
void persistent_masked_m_silu_mul_quant(
const at::Tensor& input, // (E, T, 2*H)
const at::Tensor& counts, // (E)
@@ -225,44 +219,6 @@ torch::Tensor ggml_moe_a8_vec(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);
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);
void cutlass_fp4_group_mm(
torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
const torch::Tensor& expert_offsets, const torch::Tensor& sf_offsets);
std::tuple<torch::Tensor, torch::Tensor> scaled_fp4_quant_func(
torch::Tensor const& input, torch::Tensor const& input_scale,
bool is_sf_swizzled_layout);
void scaled_fp4_quant_out(torch::Tensor const& input,
torch::Tensor const& input_scale,
bool is_sf_swizzled_layout, torch::Tensor& output,
torch::Tensor& output_scale);
void scaled_fp4_experts_quant(
torch::Tensor& output, torch::Tensor& output_scale,
torch::Tensor const& input, torch::Tensor const& input_global_scale,
torch::Tensor const& input_offset_by_experts,
torch::Tensor const& output_scale_offset_by_experts);
void silu_and_mul_scaled_fp4_experts_quant(
torch::Tensor& output, torch::Tensor& output_scale,
torch::Tensor const& input, torch::Tensor const& input_global_scale,
torch::Tensor const& input_offset_by_experts,
torch::Tensor const& output_scale_offset_by_experts);
#endif
void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
torch::Tensor const& scale,
std::optional<torch::Tensor> const& azp);