[Qwen-moe] use npu_add_rms_norm_quant operator #2362
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces an optimization for Qwen-MoE models on Ascend NPUs by utilizing the npu_add_rms_norm_quant fused operator. The changes are controlled by a new environment variable USE_ADD_RMSNORM_QUANT. The implementation correctly adds a conditional path to use this new operator when applicable.
However, I've identified a critical issue in the implementation of AddRMSNormW8A8Quant where the eps value from the model configuration is ignored due to an incorrect call to the superclass constructor. This will cause the layer to use a default epsilon, potentially leading to numerical inconsistencies and incorrect model outputs. Please see the detailed comment for the fix.
s-jiayang
force-pushed
the
use_add_quant
branch
from
9b07cda to
a708ebc
Compare
|
👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:
If CI fails, you can run linting and testing checks locally according Contributing and Testing. |
Signed-off-by: s30076806 <[email protected]>
s-jiayang
force-pushed
the
use_add_quant
branch
from
a708ebc to
ddb0641
Compare
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
| "Expected quant_config to be an instance of AscendQuantConfig" | ||
| if isinstance(self.self_attn.qkv_proj.quant_method.quant_method, | ||
| AscendW8A8LinearMethod): | ||
| self.input_layernorm = AddRMSNormW8A8Quant( |
There was a problem hiding this comment.
We will use the torch.fx rewriter to solve this kind of problem.
Please refer #2389
What this PR does / why we need it?
This PR the npu_add_rms_norm_quant operator is enabled
Does this PR introduce any user-facing change?
How was this patch tested?