Skip to content

Make warp segment threshold consitent with host function #148

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

Open
avbokovoy wants to merge 2 commits into
base: abokovoi/upstream
Choose a base branch
from
+19 −4
Open

Make warp segment threshold consitent with host function #148

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b55ad9c
Make warp segment threshold consitent with host function
avbokovoy Mar 24, 2026
b7fed8f
Extract is_supported_cdna()
avbokovoy Apr 9, 2026
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
23 changes: 19 additions & 4 deletions fbgemm_gpu/codegen/training/pt2/embedding_split_host_pt2_autograd_template.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,10 @@
#include "torch/csrc/autograd/record_function_ops.h"
#include "torch/csrc/autograd/record_function_ops.h"

#ifdef USE_ROCM
#include "fbgemm_gpu/rocm/cdna_guard.h"
#endif

{%- if ssd %}
#include "pt2_arg_utils_ssd.h"
{%- else %}
Expand Down Expand Up @@ -1067,13 +1071,24 @@ static torch::autograd::variable_list backward(
(not ssd) and
(not is_index_select) and
(not dense) %}
int32_t total_L = indices.numel();
const int32_t total_L = indices.numel();
const auto T = weights_offsets.sym_numel();
auto total_B = (offsets.size(0) - 1);
const auto B = total_B / T;
const bool cached = weights_uvm.numel() > 0 || weights_lxu_cache.numel() > 0;
const bool use_hip_kernel = config::is_feature_enabled(
config::FeatureGateName::TBE_ROCM_HIP_BACKWARD_KERNEL);
const auto weights_type = weights_dev.scalar_type();
const auto grad_type = grad_outputs[0].scalar_type();
const bool supported_weights_type =
(weights_type == at::kFloat) || (weights_type == at::kHalf);
const bool supported_grad_type =
(grad_type == at::kFloat) || (grad_type == at::kHalf);
const bool same_precision = (weights_type == grad_type);
const static bool supported_platform = rocm::is_supported_cdna();
{%- for kDimSize in [64, 128, 160, 192, 256, 320] %}
if(!mixed_D && total_L / total_B > 1 && (max_D == {{ kDimSize }}))
{
if (use_hip_kernel && total_L / total_B > 1 && !mixed_D && !cached && supported_weights_type &&
supported_grad_type && same_precision && supported_platform &&
(max_D == {{ kDimSize }})) {
max_segment_length_per_warp = 16384;
}
{%- endfor %}
Expand Down