Skip to content

[Bug] Fix precision of gate and e_score_correction_bias in Glm4Moe #2637

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
Sep 19, 2025
Merged

[Bug] Fix precision of gate and e_score_correction_bias in Glm4Moe #2637

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
df3b236
Fix precision of gate and e_score_correction_bias in Glm4Moe
DrownFish19 Sep 19, 2025
35635b5
add _keep_in_fp32_modules
DrownFish19 Sep 19, 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
9 changes: 8 additions & 1 deletion paddleformers/transformers/glm4_moe/modeling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,6 +304,9 @@ def __init__(self, config, num_experts, expert_hidden_size, **kwargs):
)
self.expert_usage.stop_gradient = True

# weight and e_score_correction_bias do not need to be cast to low precision
self._cast_to_low_precision = False

def forward(self, hidden_states):
"""
Args:
Expand Down Expand Up @@ -340,12 +343,15 @@ def __init__(self, config: Glm4MoeConfig):

self.weight = paddle.create_parameter(
shape=[self.n_routed_experts, config.hidden_size],
dtype="bfloat16",
dtype="float32",
default_initializer=paddle.nn.initializer.Uniform(),
)

self.register_buffer("e_score_correction_bias", paddle.zeros((self.n_routed_experts,), dtype=paddle.float32))

# weight and e_score_correction_bias do not need to be cast to low precision
self._cast_to_low_precision = False

@paddle.no_grad()
def get_topk_indices(self, scores):
scores_for_choice = scores.reshape([-1, self.n_routed_experts]) + self.e_score_correction_bias.unsqueeze(0)
Expand Down Expand Up @@ -588,6 +594,7 @@ class Glm4MoePreTrainedModel(PretrainedModel):
config: Glm4MoeConfig
config_class = Glm4MoeConfig
base_model_prefix = "model"
_keep_in_fp32_modules = ["mlp.gate.weight", "e_score_correction_bias"]
transpose_weight_keys = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]

@classmethod
Expand Down
Loading