chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

[Feature] Add LoRA support for embedding layers (#14177)

Browse Source 
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
Co-authored-by: Beichen-Ma <bm685@cornell.edu>
This commit is contained in:
Ethan (Yusheng) Su
2025-12-09 15:53:33 -08:00
committed by GitHub
co-authored by Baizhou Zhang Beichen-Ma
parent 9ad02b799d
commit 0c63fb9420
18 changed files with 1492 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/lora/test_lora_update.py
+35 -3
View File
@@ -218,7 +218,17 @@ BASIC_TESTS = [
base="meta-llama/Llama-3.1-8B-Instruct",
enable_lora=True,
max_lora_rank=256,
lora_target_modules=["all"],
# Need to list all lora modules, or "all" might include lora modules without assigning lora weights
# lora_target_modules=["all"],
lora_target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
max_loras_per_batch=4,
all_adapters=[
"philschmid/code-llama-3-1-8b-text-to-sql-lora",
@@ -337,6 +347,8 @@ TARGET_MODULE_TESTS = [
description="Test explicitly specified lora-target-modules.",
base="meta-llama/Llama-3.1-8B-Instruct",
max_loras_per_batch=3,
# Need to list all lora modules, or "all" might include lora modules without assigning lora weights
# lora_target_modules=["all"],
lora_target_modules=[
"q_proj",
"k_proj",
@@ -751,7 +763,17 @@ EVICTION_TESTS = [
],
enable_lora=True,
max_lora_rank=256,
lora_target_modules=["all"],
# Need to list all lora modules, or "all" might include lora modules without assigning lora weights
# lora_target_modules=["all"],
lora_target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
op_sequence=[
Operation(
type=OperationType.LOAD,
@@ -1503,7 +1525,17 @@ class TestLoRADynamicUpdate(CustomTestCase):
lora_paths=[],
max_loras_per_batch=2,
max_lora_rank=256,
lora_target_modules=["all"],
# Need to list all lora modules, or "all" might include lora modules without assigning lora weights
# lora_target_modules=["all"],
lora_target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
enable_lora=True,
) as session:
# Test with no adapters loaded