Skip to content

[NVIDIA] Fix num_experts in modelopt_quant #8811

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
ch-wan merged 7 commits into from
Aug 6, 2025
Merged

[NVIDIA] Fix num_experts in modelopt_quant #8811

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
44e54b3
Fix num_experts in modelopt_quant
wenscarl Aug 5, 2025
c5a1942
Remove verbose print
wenscarl Aug 5, 2025
bfa1bfd
Avoid change nmodification of num_experts in create_weights
wenscarl Aug 5, 2025
06a225e
Merge branch 'main' into fix_trtllm_moe_create_weights
ishandhanani Aug 5, 2025
5b7a7e1
Merge branch 'main' into fix_trtllm_moe_create_weights
zhyncs Aug 5, 2025
85e68e2
Fix trtllm_fp4_block_scale_moe API change
wenscarl Aug 6, 2025
7a3b228
Merge branch 'main' into fix_trtllm_moe_create_weights
ishandhanani Aug 6, 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
5 changes: 5 additions & 0 deletions python/sglang/srt/layers/moe/fused_moe_triton/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1063,10 +1063,15 @@ def forward(self, hidden_states: torch.Tensor, topk_output):
gemm1_weights_scale=self.gemm1_scales_fp4_shuffled.data.view(
torch.float8_e4m3fn
),
gemm1_bias=None,
gemm1_alpha=None,
gemm1_beta=None,
gemm1_clamp_limit=None,
gemm2_weights=self.gemm2_weights_fp4_shuffled.data,
gemm2_weights_scale=self.gemm2_scales_fp4_shuffled.data.view(
torch.float8_e4m3fn
),
gemm2_bias=None,
output1_scale_scalar=self.g1_scale_c.data,
output1_scale_gate_scalar=self.g1_alphas.data,
output2_scale_scalar=self.g2_alphas.data,
Expand Down
6 changes: 2 additions & 4 deletions python/sglang/srt/layers/quantization/modelopt_quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -764,8 +764,6 @@ def create_weights(
)

# TODO(ch-wan): check if this is needed
layer.num_experts = num_experts
layer.num_local_experts = num_experts
layer.intermediate_size_per_partition = intermediate_size_per_partition
layer.params_dtype = params_dtype
layer.quant_config = self.quant_config
Expand Down Expand Up @@ -1106,7 +1104,7 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
layer.w13_weight_scale,
)

print("Applied flashinfer weight processing for both w13 and w2")
logger.info_once("Applied flashinfer weight processing for both w13 and w2")

else:
# CUTLASS processing - handle w13 and w2 separately
Expand All @@ -1126,7 +1124,7 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
layer.w2_weight = Parameter(layer.w2_weight.data, requires_grad=False)

# Both flashinfer cutlass and regular cutlass use same processing for w2
print("Applied weight processing for both w13 and w2")
logger.info_once("Applied weight processing for both w13 and w2")

# Set up CUTLASS MoE parameters
device = layer.w13_weight.device
Expand Down
Loading