Skip to content

[None][feat] AutoDeploy: Use the router gemm op for nemotron MOE #9500

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
nvchenghaoz merged 1 commit into from
Dec 1, 2025
Merged

[None][feat] AutoDeploy: Use the router gemm op for nemotron MOE #9500

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
13 changes: 7 additions & 6 deletions tensorrt_llm/_torch/auto_deploy/models/patches/nemotron_h.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,14 +93,15 @@ def _nemotron_h_topk_router_forward(self, hidden_states):
Forward pass for NemotronHTopkRouter using the optimized noaux_tc_op kernel.

This replaces the original forward method which used pure PyTorch operations
with a fused CUDA kernel that performs:
1. Sigmoid activation of logits
2. Group-based expert selection
3. Top-k selection within selected groups
4. Normalized weight computation
with optimized CUDA kernels:
"""
hidden_states = hidden_states.view(-1, self.config.hidden_size)
router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
if self.weight.dtype == torch.float32:
router_logits = F.linear(hidden_states.type(torch.float32), self.weight)
else:
router_logits = torch.ops.trtllm.dsv3_router_gemm_op(
hidden_states, self.weight.t(), bias=None, out_dtype=torch.float32
)

# Use the fused noaux_tc_op kernel which applies sigmoid internally
# and performs group-based top-k selection with normalization
Expand Down