Consolidate Llama model usage in tests (#13094)
This commit is contained in:
@@ -46,9 +46,9 @@ def test_filter_subtensors():
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def llama_2_7b_files():
|
||||
def llama_3p2_1b_files():
|
||||
with TemporaryDirectory() as cache_dir:
|
||||
input_dir = snapshot_download("meta-llama/Llama-3.2-1B",
|
||||
input_dir = snapshot_download("meta-llama/Llama-3.2-1B-Instruct",
|
||||
cache_dir=cache_dir,
|
||||
ignore_patterns=["*.bin*", "original/*"])
|
||||
|
||||
@@ -81,13 +81,13 @@ def _run_generate(input_dir, queue: mp.Queue, **kwargs):
|
||||
@pytest.mark.parametrize("enable_lora", [False, True])
|
||||
@pytest.mark.parametrize("tp_size", [1, 2])
|
||||
def test_sharded_state_loader(enable_lora, tp_size, num_gpus_available,
|
||||
llama_2_7b_files):
|
||||
llama_3p2_1b_files):
|
||||
if num_gpus_available < tp_size:
|
||||
pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
|
||||
|
||||
weights_patterns = ("*.safetensors", )
|
||||
gpu_memory_utilization = 0.8
|
||||
input_dir = llama_2_7b_files
|
||||
input_dir = llama_3p2_1b_files
|
||||
ctx = mp.get_context("spawn")
|
||||
|
||||
# Run in separate processes for memory & CUDA isolation
|
||||
|
||||
Reference in New Issue
Block a user