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

[NPU] Update C++ L0 (#12458)

Browse source 
* update

* fix style
This commit is contained in:
Ruonan Wang 2024-11-27 06:08:48 -08:00 committed by GitHub
parent a2272b70d3
commit b29da30205
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 7 additions and 23 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

22
python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py
View file

@ -426,11 +426,9 @@ def convert_llm_for_deploy(model: torch.nn.Module,
os.mkdir(save_directory)
weight_dir = os.path.join(save_directory, "model_weights")
os.mkdir(weight_dir)
use_level_zero = os.environ.get("IPEX_LLM_NPU_USE_LEVEL0", "0") == "1"
layernorm_const = os.environ.get("IPEX_LLM_NPU_LAYERNORM_CONST", "1") == "1"
if model.config.model_type == "qwen2":
layernorm_const = os.environ.get("IPEX_LLM_NPU_LAYERNORM_CONST", "0") == "1"
if model.config.hidden_size == 1536:
# Qwen2-1.5B-Instruct
fused_layers = 1
@ -449,28 +447,16 @@ def convert_llm_for_deploy(model: torch.nn.Module,
"weight_num": 7,
"weight_idx": 8,
"n_splits_linear": n_splits_linear,
"n_splits_down_proj": n_splits_down_proj,
"use_level_zero": use_level_zero}
"n_splits_down_proj": n_splits_down_proj}
model.config.update(update_dict)
model.config.save_pretrained(save_directory)
from .qwen import convert_qwen_layer, convert_fused_qwen_layer
from .qwen import convert_lm_head_and_embedding
if not use_level_zero:
# save fused_layers blobs of fused decoder layers
convert_fused_qwen_layer(model, fused_layers, n_splits_linear, n_splits_down_proj,
save_directory, weight_dir, transpose_value_cache, kv_len,
group_size, layernorm_const, "decode")
else:
# save layer_num blobs of each decoder layer
layer_num = len(model.model.layers)
param_list = []
for layer_idx in range(0, layer_num):
param_list.append((model, layer_idx, n_splits_linear, n_splits_down_proj,
save_directory, weight_dir, transpose_value_cache, kv_len,
group_size, layernorm_const))
with Pool() as pool:
result = pool.starmap(convert_qwen_layer, param_list)
# save blob of single prefill layer
convert_qwen_layer(model, 0, n_splits_linear, n_splits_down_proj,
save_directory, weight_dir, transpose_value_cache, max_prompt_len,
@ -512,8 +498,7 @@ def convert_llm_for_deploy(model: torch.nn.Module,
"embedding_post": embedding_post,
"cos_sin_input": cos_sin_input,
"n_splits_linear": n_splits_linear,
"n_splits_down_proj": n_splits_down_proj,
"use_level_zero": use_level_zero}
"n_splits_down_proj": n_splits_down_proj}
model.config.update(update_dict)
model.config.save_pretrained(save_directory)
@ -549,8 +534,7 @@ def convert_llm_for_deploy(model: torch.nn.Module,
"model_type": "minicpm",
"embedding_post": True,
"n_splits_linear": n_splits_linear,
"n_splits_down_proj": n_splits_down_proj,
"use_level_zero": use_level_zero}
"n_splits_down_proj": n_splits_down_proj}
model.config.update(update_dict)
model.config.save_pretrained(save_directory)