apply_query_key_layer_scaling ignored in PyTorch ≥ 2.0 scaled_dot_product_attention path

System Info / 系統信息

None

Who can help? / 谁可以帮助到您？

No response

Information / 问题信息

• The official example scripts / 官方的示例脚本
• My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

Description

In the ChatGLM implementation, the configuration flag:

apply_query_key_layer_scaling = True

is ignored when running on PyTorch 2.0 or later.

Background

In the original THUDM ChatGLM3 implementation, layer-wise attention scaling is applied:

if self.apply_query_key_layer_scaling:
    coeff = layer_number
    self.norm_factor *= coeff

This effectively scales attention scores by 1 / (sqrt(head_dim) * layer_number), stabilizing deep layers.

⸻

Current Behavior For PyTorch < 2.0, the implementation uses baddbmm with alpha = 1 / norm_factor:

matmul_result = torch.baddbmm(
    ...,
    alpha=1.0 / self.norm_factor
)

In this path, layer-wise scaling works as intended. For PyTorch ≥ 2.0, the implementation uses torch.nn.functional.scaled_dot_product_attention:

context_layer = torch.nn.functional.scaled_dot_product_attention(
    query_layer, key_layer, value_layer, attention_mask, is_causal=True
)

This function applies a fixed scale of 1 / sqrt(head_dim) and does not accept a layer_number parameter. As a result, apply_query_key_layer_scaling=True is completely ignored.

⸻

Expected behavior / 期待表现

Suggested Fix • Expose a scaling or norm_factor parameter in the PyTorch 2.0 path • Multiply by layer_number when apply_query_key_layer_scaling=True