Skip to content

Quick fix for DeepGemm requant to also cover MTP. #7378

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 2 commits into from
Jun 23, 2025
Merged

Quick fix for DeepGemm requant to also cover MTP. #7378

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3878138
Quick fix for DeepGemm requant to also cover MTP.
pyc96 Jun 20, 2025
e67e59d
Merge branch 'main' into requant
zhyncs Jun 20, 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
16 changes: 9 additions & 7 deletions python/sglang/srt/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1936,11 +1936,9 @@ def post_load_weights(self, is_nextn=False, weight_names=None):
and hasattr(self.quant_config, "weight_block_size")
and self.quant_config.weight_block_size is not None
):
self._weight_requant_ue8m0()
self._weight_requant_ue8m0(is_nextn)

def _weight_requant_ue8m0(self):
if self.config.architectures[0] == "DeepseekV3ForCausalLMNextN":
return
def _weight_requant_ue8m0(self, is_nextn=False):
weight_block_size = self.quant_config.weight_block_size

moe_layers = list(
Expand All @@ -1951,8 +1949,12 @@ def _weight_requant_ue8m0(self):
)
)

for layer_id in range(self.config.num_hidden_layers):
layer = self.model.layers[layer_id]
num_hidden_layers = 1 if is_nextn else self.config.num_hidden_layers
for layer_id in range(num_hidden_layers):
if is_nextn:
layer = self.model.decoder
else:
layer = self.model.layers[layer_id]
Comment on lines +1953 to +1957
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jun 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Consider adding a comment to explain why the decoder is used when is_nextn is true, and layers otherwise.

            if is_nextn:
                layer = self.model.decoder  # Use decoder for NextN
            else:
                layer = self.model.layers[layer_id]  # Use layers for normal layers


for module in [
layer.self_attn.fused_qkv_a_proj_with_mqa,
Expand All @@ -1964,7 +1966,7 @@ def _weight_requant_ue8m0(self):
module.weight, module.weight_scale_inv, weight_block_size
)

if layer_id in moe_layers:
if layer_id in moe_layers or is_nextn:
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jun 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Consider adding a comment to explain why is_nextn is included in this conditional.

            if layer_id in moe_layers or is_nextn:  # Also apply to NextN

shared_experts = getattr(layer.mlp, "shared_experts", None)
if shared_experts is not None:
for module in [
Expand Down
Loading