Skip to content

fix: int32 overflow in trtllm_fp4_block_scale_moe causing "Unsupported hidden state scale shape" for EP32+ configs #2853

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
yzh119 merged 3 commits into from
Mar 24, 2026
Merged

fix: int32 overflow in trtllm_fp4_block_scale_moe causing "Unsupported hidden state scale shape" for EP32+ configs #2853

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1e59ce5
fix: int32 overflow in trtllm_fp4_block_scale_moe for large EP config…
Mar 22, 2026
07d7aaf
chore: also change hidden_states_scale_vec_size to int64_t for consis…
Mar 23, 2026
69e587e
style: apply clang-format
Mar 23, 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
11 changes: 6 additions & 5 deletions csrc/trtllm_fused_moe_kernel_launcher.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1986,15 +1986,16 @@ Array<Tensor> trtllm_fp4_block_scale_moe(
int hidden_size = hidden_states.size(1);
if (hidden_states.dtype() == dl_uint8) hidden_size *= 2;

int hidden_states_scale_vec_size = -1;
int64_t hidden_states_scale_vec_size = -1;
if (hidden_states_scale.has_value()) {
hidden_states_scale_vec_size = (num_tokens * hidden_size) / hidden_states_scale.value().numel();
hidden_states_scale_vec_size =
(static_cast<int64_t>(num_tokens) * hidden_size) / hidden_states_scale.value().numel();
}
int64_t intermediate_size_factor =
isGatedActivation(static_cast<ActivationType>(act_type)) ? 2 : 1;
int weight_scale_vec_size =
(local_num_experts * intermediate_size * intermediate_size_factor * hidden_size) /
gemm1_weights_scale.numel();
int64_t weight_scale_vec_size = (static_cast<int64_t>(local_num_experts) * intermediate_size *
intermediate_size_factor * hidden_size) /
gemm1_weights_scale.numel();

TVM_FFI_ICHECK(weight_scale_vec_size == 16 || weight_scale_vec_size == 32)
<< "unsupported weight_scale_vec_size.";
Expand Down
Loading