Skip to content

[BugFix] Fix EP MoE prefill function #4101

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
merged 4 commits into from
Sep 24, 2025
Merged

[BugFix] Fix EP MoE prefill function #4101

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
28c20d9
fix wint8 ep
Sunny-bot1 Sep 15, 2025
4577afd
Merge branch 'feature/experimental_feature_20250908' into exp_wint8
Sunny-bot1 Sep 15, 2025
ea7aa57
Merge branch 'feature/experimental_feature_20250908' into exp_wint8
Sunny-bot1 Sep 16, 2025
7e8c5b6
Merge branch 'feature/experimental_feature_20250908' into exp_wint8
Sunny-bot1 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
192 changes: 63 additions & 129 deletions custom_ops/gpu_ops/moe/ep_moe_prefill_func.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -453,137 +453,71 @@ void EPMoeDispatchKernel(const paddle::Tensor& input,
auto place = input.place();
const int gridx = min(132 * 8, num_rows);
if (moe_quant_type == "w4a8") {
if (num_experts_per_rank == 8) {
permute_x_kernel<data_t, int8_t, 8><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<int8_t>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
127.0,
-127.0
);
} else if (num_experts_per_rank == 16) {
permute_x_kernel<data_t, int8_t, 16><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<int8_t>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
127.0,
-127.0
);
}
DISPATCH_NUM_EXPERTS_PER_RANK(num_experts_per_rank, NUM_EXPERTS_PER_RANK,
permute_x_kernel<data_t, int8_t, NUM_EXPERTS_PER_RANK><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<int8_t>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
127.0,
-127.0
);)
} else if (moe_quant_type == "w4afp8") {
if (num_experts_per_rank == 8) {
permute_x_kernel<data_t, data_t_fp8, 8, 512><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<data_t_fp8>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
448.0f,
-448.0f
);
} else if (num_experts_per_rank == 16) {
permute_x_kernel<data_t, data_t_fp8, 16, 512><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<data_t_fp8>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
448.0f,
-448.0f
);
}
DISPATCH_NUM_EXPERTS_PER_RANK(num_experts_per_rank, NUM_EXPERTS_PER_RANK,
permute_x_kernel<data_t, data_t_fp8, NUM_EXPERTS_PER_RANK, 512><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<data_t_fp8>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
448.0f,
-448.0f
);)
} else {
if (num_experts_per_rank == 8) {
permute_x_kernel<data_t, data_t, 8><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<data_t>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
127.0,
-127.0
);
} else if (num_experts_per_rank == 16) {
permute_x_kernel<data_t, data_t, 16><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<data_t>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
127.0,
-127.0
);
}
DISPATCH_NUM_EXPERTS_PER_RANK(num_experts_per_rank, NUM_EXPERTS_PER_RANK,
permute_x_kernel<data_t, data_t, NUM_EXPERTS_PER_RANK><<<gridx, 512, 0, stream>>>(
input.data<data_t>(),
topk_ids.data<int64_t>(),
topk_weights.data<float>(),
token_nums_per_expert.data<int>(),
up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
moe_topk,
num_rows,
token_nums_this_rank,
hidden_size,
permute_input->data<data_t>(),
permute_indices_per_token->data<int>(),
dst_weights->data<float>(),
dst_indices->data<int>(),
cumsum_idx_gpu->data<int>(),
token_nums_per_expert_cumsum->data<int64_t>(),
expert_idx_per_token->data<int64_t>(),
127.0,
-127.0
);)
}
}

Expand Down