Skip to content

FP4 grouped refactor #4934

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
cthi wants to merge 3 commits into from
Closed

FP4 grouped refactor #4934

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4cdb116
Fix typo in FP4 quantize (#4933)
cthi Sep 24, 2025
cf2dc81
Split grouped gemm metadata kernel into grouped_common.cuh (#4932)
cthi Sep 24, 2025
553b40e
FP4 grouped refactor
cthi Sep 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion fbgemm_gpu/experimental/gemm/triton_gemm/fp4_quantize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1461,7 +1461,7 @@ def triton_scale_nvfp4_quant(
stochastic_casting (bool): Whether to use stochastic casting.

Returns:
torch.Tensor: [M / 2] nvfp4 scaled tensor packed into in8
torch.Tensor: [M / 2] nvfp4 scaled tensor packed into int8
torch.Tensor: [M / group_size] nvfp4 shared exponents into int8

eg.
Expand Down
17 changes: 0 additions & 17 deletions fbgemm_gpu/experimental/gen_ai/src/quantize/cutlass_extensions/f4f4bf16_grouped.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,16 +8,6 @@

#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>
#include <cutlass/util/device_memory.h>
#include <cutlass/util/packed_stride.hpp>

// clang-format off
// The fixed ordering of the headers is required for CUTLASS 3.2+
#include <cute/tensor.hpp>
#include <cutlass/gemm/collective/collective_builder.hpp> // @manual
#include <cutlass/gemm/device/gemm_universal_adapter.h> // @manual
#include <cutlass/epilogue/collective/collective_builder.hpp> // @manual
// clang-format on

#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12080)
#include "f4f4bf16_grouped/f4f4bf16_grouped_manifest.cuh"
Expand Down Expand Up @@ -160,14 +150,10 @@ at::Tensor dispatch_fp4_grouped_kernel(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
std::optional<at::Tensor> zero_start_index_M = std::nullopt,
std::optional<at::Tensor> M_sizes = std::nullopt,
std::optional<at::Tensor> global_scale = std::nullopt,
std::optional<at::Tensor> starting_row_after_padding = std::nullopt,
bool use_mx = true) {
TORCH_CHECK(
zero_start_index_M.has_value() != M_sizes.has_value(),
"One of zero_start_index_M or M_sizes must be provided.");
TORCH_CHECK(M_sizes.has_value(), "M_sizes is assumed to be provided.");
TORCH_CHECK(
starting_row_after_padding.has_value(),
Expand All @@ -187,8 +173,6 @@ at::Tensor dispatch_fp4_grouped_kernel(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down Expand Up @@ -228,7 +212,6 @@ at::Tensor f4f4bf16_grouped_stacked(
x_scale,
w_scale,
Y,
std::nullopt,
M_sizes,
global_scale,
starting_row_after_padding,
Expand Down
4 changes: 0 additions & 4 deletions .../src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_128_128_256_1_1_1_f.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_128_128_256_1_1_1_f(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_128_128_256_1_1_1_f(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions .../src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_128_128_256_1_1_1_t.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_128_128_256_1_1_1_t(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_128_128_256_1_1_1_t(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions ...i/src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_128_64_256_1_1_1_f.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_128_64_256_1_1_1_f(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_128_64_256_1_1_1_f(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions ...i/src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_128_64_256_1_1_1_t.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_128_64_256_1_1_1_t(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_128_64_256_1_1_1_t(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions .../src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_256_128_256_2_1_1_f.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_256_128_256_2_1_1_f(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_256_128_256_2_1_1_f(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions .../src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_256_128_256_2_1_1_t.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_256_128_256_2_1_1_t(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_256_128_256_2_1_1_t(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions .../src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_256_256_256_2_1_1_f.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_256_256_256_2_1_1_f(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_256_256_256_2_1_1_f(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions .../src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_256_256_256_2_1_1_t.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_256_256_256_2_1_1_t(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_256_256_256_2_1_1_t(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions ...i/src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_256_64_256_2_1_1_f.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_256_64_256_2_1_1_f(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_256_64_256_2_1_1_f(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
4 changes: 0 additions & 4 deletions ...i/src/quantize/cutlass_extensions/f4f4bf16_grouped/f4f4bf16_grouped_256_64_256_2_1_1_t.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,6 @@ at::Tensor f4f4bf16_grouped_256_64_256_2_1_1_t(
at::Tensor x_scale,
at::Tensor w_scale,
at::Tensor output,
int64_t G,
std::optional<at::Tensor> zero_start_index_M,
std::optional<at::Tensor> M_sizes,
std::optional<at::Tensor> global_scale,
std::optional<at::Tensor> starting_row_after_padding) {
Expand All @@ -36,8 +34,6 @@ at::Tensor f4f4bf16_grouped_256_64_256_2_1_1_t(
x_scale,
w_scale,
output,
G,
zero_start_index_M,
M_sizes,
global_scale,
starting_row_after_padding);
Expand Down
Loading
Loading