tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

GRPO fixes (peft) (#2676)

Browse Source 
* don't set peft_config on grpo to prevent double peft wrap

* remove overrides needed to support bug

* fix grpo tests

* require more CPU for multigpu to help with torch compile for vllm
This commit is contained in:
Wing Lian
2025-05-16 15:47:03 -04:00
committed by GitHub
parent 6cb07b9d12
commit a27b909c5c
4 changed files with 5 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/e2e/multigpu/solo/test_grpo.py
View File

@@ -166,7 +166,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
"""
)
@pytest.mark.skip(reason="flaky test")
@pytest.mark.parametrize(
"num_gpus",
[1, 2],
@@ -231,8 +230,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
"NCCL_P2P_LEVEL": "LOC",
**current_env,
"CUDA_VISIBLE_DEVICES": "1",
"VLLM_DISABLE_COMPILE_CACHE": "1",
# "VLLM_USE_V1": "0",
}
vllm_process = start_vllm(
cfg.base_model,
@@ -266,7 +263,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
finally:
recursive_kill(vllm_process)
@pytest.mark.skip(reason="flaky test")
@pytest.mark.parametrize(
"num_gpus",
[1, 2],
@@ -325,8 +321,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
"NCCL_P2P_LEVEL": "LOC", # nccl can be brittle, assume P2P isn't reliable
**current_env,
"CUDA_VISIBLE_DEVICES": "1",
"VLLM_DISABLE_COMPILE_CACHE": "1",
# "VLLM_USE_V1": "0",
}
vllm_process = start_vllm(
cfg.base_model,