Preserve revision pins in secondary artifact loaders (#51446)
Signed-off-by: Clinton Thomas <1033162+KernelClint@users.noreply.github.com> Co-authored-by: Lucas Bourtoule <35483370+dhalf@users.noreply.github.com>
This commit is contained in:
co-authored by
Lucas Bourtoule
parent
adc1200597
commit
90fd4a333f
@@ -7,6 +7,7 @@ from copy import copy
|
||||
from pathlib import Path
|
||||
from types import SimpleNamespace
|
||||
from typing import cast
|
||||
from unittest.mock import Mock
|
||||
|
||||
import pytest
|
||||
from transformers import PretrainedConfig
|
||||
@@ -138,6 +139,32 @@ def test_head_size_falls_back_when_head_dim_is_zero():
|
||||
assert convertor.get_head_size() == 128
|
||||
|
||||
|
||||
def test_qk_rope_head_dim_recovery_uses_model_revision(monkeypatch: pytest.MonkeyPatch):
|
||||
hf_config = PretrainedConfig(
|
||||
model_type="deepseek_v2",
|
||||
qk_rope_head_dim=128,
|
||||
qk_nope_head_dim=128,
|
||||
)
|
||||
hf_config.name_or_path = "org/model"
|
||||
get_hf_file_to_dict = Mock(return_value={"qk_rope_head_dim": 64})
|
||||
|
||||
monkeypatch.setattr(
|
||||
"vllm.transformers_utils.repo_utils.get_hf_file_to_dict",
|
||||
get_hf_file_to_dict,
|
||||
)
|
||||
|
||||
convertor = ModelArchConfigConvertorBase(
|
||||
hf_config,
|
||||
hf_config,
|
||||
revision="pinned-revision",
|
||||
)
|
||||
|
||||
assert convertor._get_qk_rope_head_dim() == 64
|
||||
get_hf_file_to_dict.assert_called_once_with(
|
||||
"config.json", "org/model", "pinned-revision"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"attribute",
|
||||
[
|
||||
|
||||
@@ -0,0 +1,40 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import Mock
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.config.vllm import VllmConfig
|
||||
|
||||
pytestmark = [pytest.mark.cpu_test, pytest.mark.skip_global_cleanup]
|
||||
|
||||
|
||||
def test_quantization_config_forwards_model_revision(monkeypatch: pytest.MonkeyPatch):
|
||||
quant_config = Mock()
|
||||
quant_config.get_min_capability.return_value = 0
|
||||
quant_config.get_supported_act_dtypes.return_value = [torch.float32]
|
||||
model_config = SimpleNamespace(
|
||||
quantization="auto_gptq",
|
||||
model="repo/model",
|
||||
hf_config=object(),
|
||||
dtype=torch.float32,
|
||||
revision="pinned-revision",
|
||||
)
|
||||
|
||||
monkeypatch.setattr(
|
||||
"vllm.model_executor.model_loader.weight_utils.get_quant_config",
|
||||
lambda model_config, load_config: quant_config,
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
"vllm.platforms.current_platform.get_device_capability", lambda: None
|
||||
)
|
||||
|
||||
VllmConfig._get_quantization_config(model_config, SimpleNamespace())
|
||||
|
||||
quant_config.maybe_update_config.assert_called_once_with(
|
||||
"repo/model",
|
||||
hf_config=model_config.hf_config,
|
||||
revision="pinned-revision",
|
||||
)
|
||||
@@ -0,0 +1,51 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import contextlib
|
||||
import io
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import Mock
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.model_loader import tensorizer
|
||||
|
||||
pytestmark = [pytest.mark.cpu_test, pytest.mark.skip_global_cleanup]
|
||||
|
||||
|
||||
def test_serialize_vllm_model_forwards_model_revision(
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
):
|
||||
@contextlib.contextmanager
|
||||
def fake_open_stream(*args, **kwargs):
|
||||
yield io.BytesIO()
|
||||
|
||||
snapshot_download = Mock()
|
||||
monkeypatch.setattr(tensorizer, "open_stream", fake_open_stream)
|
||||
monkeypatch.setattr(tensorizer, "TensorSerializer", Mock())
|
||||
monkeypatch.setattr(
|
||||
tensorizer,
|
||||
"hf_api",
|
||||
lambda: SimpleNamespace(snapshot_download=snapshot_download),
|
||||
)
|
||||
|
||||
model = torch.nn.Module()
|
||||
tensorizer_config = SimpleNamespace(
|
||||
_construct_tensorizer_args=lambda: SimpleNamespace(
|
||||
tensorizer_uri="/tmp/model.tensors",
|
||||
tensorizer_dir="/tmp/out",
|
||||
stream_kwargs={},
|
||||
),
|
||||
encryption_keyfile=None,
|
||||
_is_sharded=False,
|
||||
serialization_kwargs=None,
|
||||
)
|
||||
model_config = SimpleNamespace(
|
||||
served_model_name="repo/model",
|
||||
revision="pinned-revision",
|
||||
)
|
||||
|
||||
tensorizer.serialize_vllm_model(model, tensorizer_config, model_config)
|
||||
|
||||
assert snapshot_download.call_args.args == ("repo/model",)
|
||||
assert snapshot_download.call_args.kwargs["revision"] == "pinned-revision"
|
||||
@@ -6,7 +6,15 @@ only get the `eos_token_id` from the tokenizer as defined by
|
||||
`BaseRenderer.get_eos_token_id`.
|
||||
"""
|
||||
|
||||
from types import SimpleNamespace
|
||||
from typing import cast
|
||||
from unittest.mock import patch
|
||||
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.config.model import ModelConfig
|
||||
from vllm.tokenizers import get_tokenizer
|
||||
from vllm.transformers_utils import config as config_module
|
||||
from vllm.transformers_utils.config import try_get_generation_config
|
||||
|
||||
|
||||
@@ -30,3 +38,42 @@ def test_get_blip2_eos_token():
|
||||
generation_config = try_get_generation_config(model_name, trust_remote_code=False)
|
||||
assert generation_config is not None
|
||||
assert generation_config.eos_token_id == 50118
|
||||
|
||||
|
||||
def test_model_config_generation_fallback_forwards_code_revision():
|
||||
model_config = cast(
|
||||
ModelConfig,
|
||||
SimpleNamespace(
|
||||
generation_config="auto",
|
||||
hf_config_path=None,
|
||||
model="org/model",
|
||||
trust_remote_code=True,
|
||||
revision="model-pin",
|
||||
code_revision="code-pin",
|
||||
config_format="auto",
|
||||
hf_token=None,
|
||||
),
|
||||
)
|
||||
|
||||
with (
|
||||
patch.object(
|
||||
config_module.GenerationConfig,
|
||||
"from_pretrained",
|
||||
side_effect=OSError,
|
||||
),
|
||||
patch.object(
|
||||
config_module,
|
||||
"get_config",
|
||||
return_value=PretrainedConfig(),
|
||||
) as get_config,
|
||||
):
|
||||
ModelConfig.try_get_generation_config(model_config)
|
||||
|
||||
get_config.assert_called_once_with(
|
||||
"org/model",
|
||||
trust_remote_code=True,
|
||||
revision="model-pin",
|
||||
code_revision="code-pin",
|
||||
config_format="auto",
|
||||
token=None,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user