ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

[NPU] Fix regression caused by layer_norm change (#12627)

Browse source
This commit is contained in:
binbin Deng 2024-12-27 14:08:49 +08:00 committed by GitHub
parent 90f6709486
commit 46eeab4479
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 5 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
python/llm/src/ipex_llm/transformers/npu_models/mp_models_base.py
View file

@ -472,7 +472,11 @@ class LLMBaseNNFactory(NNFactory):
) )
eps = self.constant(self.rms_norm_eps) eps = self.constant(self.rms_norm_eps)
hidden_states = self.eltwise_div(hidden_states, self.sqrt(self.eltwise_add(variance, eps))) hidden_states = self.eltwise_div(hidden_states, self.sqrt(self.eltwise_add(variance, eps)))
hidden_states = self.convert_to_fp16(hidden_states) if os.environ.get("IPEX_LLM_NPU_DRIVER_VERSION", None) in ["5716", "5733"]:
# to support special drivers
hidden_states = self.convert_to_fp16(hidden_states)
else:
layernorm_weight = self.convert_to_fp32(layernorm_weight)
hidden_states = self.eltwise_mul(layernorm_weight, hidden_states) hidden_states = self.eltwise_mul(layernorm_weight, hidden_states)
hidden_states = self.convert_to_fp16(hidden_states) hidden_states = self.convert_to_fp16(hidden_states)
return hidden_states return hidden_states