apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-31 03:46:40 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

[Precision] Support lm_head layer running in float32 (#3597)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
Details
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
Details
CE Compile Job / Show Code Archive Output (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
Details
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Details
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source 
* support lm_head fp32 bf16 fp16

* support lm_head fp32 bf16 fp16

* add doc and check code

* lm_head_fp32 specify lm_head as fp32

* code check

* check doc
This commit is contained in:
chen
2025-08-27 11:34:53 +08:00
committed by GitHub
parent ad319a87cc
commit ce9c0917c5
15 changed files with 99 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
fastdeploy/model_executor/models/ernie4_5_moe.py
View File

@@ -511,10 +511,7 @@ class Ernie4_5_MoeForCausalLM(ModelForCasualLM):
"""
self.ernie.load_state_dict(state_dict)
if self.tie_word_embeddings:
if hasattr(self.lm_head, "linear"):
self.lm_head.linear.weight.set_value(self.ernie.embed_tokens.embeddings.weight.transpose([1, 0]))
else: # ep
self.lm_head.weight.set_value(self.ernie.embed_tokens.embeddings.weight.transpose([1, 0]))
self.lm_head.load_state_dict({self.lm_head.weight_key: self.ernie.embed_tokens.embeddings.weight})
else:
self.lm_head.load_state_dict(state_dict)
@@ -581,11 +578,11 @@ class Ernie4_5_MoeForCausalLM(ModelForCasualLM):
model_sublayer_name = re.sub(r"\.(up_gate_proj_weight|down_proj_weight|weight)$", "", model_param_name)
process_weights_after_loading_fn(model_sublayer_name, param)
if self.tie_word_embeddings:
self.lm_head.linear.weight.set_value(self.ernie.embed_tokens.embeddings.weight.transpose([1, 0]))
self.lm_head.load_state_dict({self.lm_head.weight_key: self.ernie.embed_tokens.embeddings.weight})
def compute_logits(self, hidden_states: paddle.Tensor):
logits = self.lm_head(hidden_states)
logits = paddle.cast(logits, paddle.float32)
logits = logits.astype(paddle.float32)
logits[:, self.ori_vocab_size :] = -float("inf")
return logits