mirror of
https://github.com/vllm-project/vllm.git
synced 2026-08-16 02:38:12 +00:00
Fix RunAI streamer tensor buffer reuse during weight loading (#43464)
Signed-off-by: bbartels <[email protected]>
This commit is contained in:
@@ -6,6 +6,7 @@ import tempfile
|
||||
|
||||
import huggingface_hub.constants
|
||||
import torch
|
||||
from safetensors.torch import save_file
|
||||
|
||||
from vllm.model_executor.model_loader.weight_utils import (
|
||||
download_weights_from_hf,
|
||||
@@ -14,6 +15,27 @@ from vllm.model_executor.model_loader.weight_utils import (
|
||||
)
|
||||
|
||||
|
||||
def test_runai_safetensors_weights_iterator_clones_reused_buffers(
|
||||
tmp_path, monkeypatch
|
||||
):
|
||||
monkeypatch.setenv("RUNAI_STREAMER_MEMORY_LIMIT", "0")
|
||||
weights_file = tmp_path / "model.safetensors"
|
||||
expected_tensors = {
|
||||
"first": torch.tensor([1.0, 2.0]),
|
||||
"second": torch.tensor([3.0, 4.0]),
|
||||
}
|
||||
save_file(expected_tensors, weights_file)
|
||||
|
||||
actual_tensors = dict(
|
||||
runai_safetensors_weights_iterator([str(weights_file)], False)
|
||||
)
|
||||
|
||||
assert actual_tensors.keys() == expected_tensors.keys()
|
||||
assert actual_tensors["first"].data_ptr() != actual_tensors["second"].data_ptr()
|
||||
for name, expected_tensor in expected_tensors.items():
|
||||
assert torch.equal(actual_tensors[name], expected_tensor)
|
||||
|
||||
|
||||
def test_runai_model_loader():
|
||||
with tempfile.TemporaryDirectory() as tmpdir:
|
||||
huggingface_hub.constants.HF_HUB_OFFLINE = False
|
||||
|
||||
@@ -1090,7 +1090,8 @@ def runai_safetensors_weights_iterator(
|
||||
mininterval=2,
|
||||
)
|
||||
|
||||
yield from tensor_iter
|
||||
for name, tensor in tensor_iter:
|
||||
yield name, tensor.clone()
|
||||
|
||||
|
||||
def _init_fastsafetensors_loader(
|
||||
|
||||
Reference in New Issue
Block a user