Skip to content

Added the verification of num_max_dispatch_tokens_per_rank to the decode operator adaptation layer. #330

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

Merged
Yael-X merged 1 commit into from
Jan 21, 2026
Merged

Added the verification of num_max_dispatch_tokens_per_rank to the decode operator adaptation layer. #330

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 5 additions & 3 deletions csrc/deepep/deep_ep.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -938,11 +938,12 @@ Buffer::low_latency_dispatch(const at::Tensor &x, const at::Tensor &topk_idx,
this->new_topk_idx = torch::cat(topk_blocks, 0);
}

EP_HOST_ASSERT(num_max_dispatch_tokens_per_rank >= new_x.size(0));

auto num_tokens = static_cast<int>(new_x.size(0)), hidden = static_cast<int>(new_x.size(1));
auto num_scales = hidden / 128, num_topk = static_cast<int>(new_topk_idx.size(1));
auto num_local_experts = num_experts / (num_ranks - shared_expert_rank_num);

int64_t global_bs = std::max(new_topk_idx.size(0), num_max_dispatch_tokens_per_rank) * num_ranks;
int64_t global_bs = num_max_dispatch_tokens_per_rank * num_ranks;
auto num_max_tokens = 0;
if (rank < shared_expert_rank_num) {
num_max_tokens = global_bs / shared_expert_rank_num;
Expand Down Expand Up @@ -1059,6 +1060,7 @@ std::tuple<at::Tensor, std::optional<EventHandle>, std::optional<std::function<v
}
// Tensor checks
EP_HOST_ASSERT(x.dim() == 2 and x.is_contiguous() and x.scalar_type() == at::kBFloat16);
EP_HOST_ASSERT(num_max_dispatch_tokens_per_rank >= new_idx.size(0));
// EP_HOST_ASSERT(x.size(0) == num_experts / num_ranks);

// get ep & tp name
Expand All @@ -1082,7 +1084,7 @@ std::tuple<at::Tensor, std::optional<EventHandle>, std::optional<std::function<v
int64_t tp_world_size = 1;
int64_t tp_rankId = 0;
int64_t expert_shared_type = 0;
int64_t global_bs = std::max(new_idx.size(0), num_max_dispatch_tokens_per_rank) * num_ranks;
int64_t global_bs = num_max_dispatch_tokens_per_rank * num_ranks;
int64_t out_dtype = 0;
int64_t comm_quant_mode = 0;
int64_t group_list_type = 0;
Expand Down