biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V1] Support LLM.apply_model (#18465)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-09-20 15:14:35 +08:00
committed by GitHub
parent be874c0201
commit 3d9a1d2de5
17 changed files with 194 additions and 169 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
tests/quantization/test_gptq_dynamic.py
View File

@@ -31,41 +31,46 @@ MODEL_QUANT = [
@pytest.mark.parametrize("model_id, use_marlin_kernel", MODEL_QUANT)
def test_gptq_with_dynamic(vllm_runner, model_id: str, use_marlin_kernel: bool,
monkeypatch):
# vllm_runner.apply_model() relies on V0 internals.
monkeypatch.setenv("VLLM_USE_V1", "0")
vllm_model = vllm_runner(model_id, dtype=torch.float16, max_model_len=2048)
# `LLM.apply_model` requires pickling a function.
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
linear_method_cls = GPTQMarlinLinearMethod if use_marlin_kernel else (
GPTQLinearMethod)
for name, submodule in (vllm_model.llm.llm_engine.model_executor.
driver_worker.model_runner.model.named_modules()):
if name == "lm_head":
assert isinstance(submodule.quant_method, linear_method_cls)
elif name == 'model.layers.0.self_attn.qkv_proj':
# The first layer is quantized using bits=4, group_size=128
# desc_act=True
assert isinstance(submodule.quant_method, linear_method_cls)
config = submodule.quant_method.quant_config
assert config.weight_bits == 4
assert config.group_size == 128
assert config.desc_act
elif name == 'model.layers.1.self_attn.qkv_proj':
# The second layer is quantized using bits=8, group_size=32
# desc_act=False
assert isinstance(submodule.quant_method, linear_method_cls)
config = submodule.quant_method.quant_config
assert get_dynamic_override(config, layer_name=name,
key="bits") == 8
assert get_dynamic_override(config,
layer_name=name,
key="group_size") == 32
assert not get_dynamic_override(
config, layer_name=name, key="desc_act")
elif (name == 'model.layers.2.self_attn.qkv_proj'
or name == 'model.layers.2.mlp.gate_up_proj'):
# All other layers (layer index >= 2) are not quantized
assert isinstance(submodule.quant_method, UnquantizedLinearMethod)
with vllm_runner(model_id, dtype=torch.float16, max_model_len=2048) as llm:
del vllm_model
def check_model(model):
for name, submodule in model.named_modules():
if name == "lm_head":
assert isinstance(submodule.quant_method,
linear_method_cls)
elif name == 'model.layers.0.self_attn.qkv_proj':
# The first layer is quantized using bits=4, group_size=128
# desc_act=True
assert isinstance(submodule.quant_method,
linear_method_cls)
config = submodule.quant_method.quant_config
assert config.weight_bits == 4
assert config.group_size == 128
assert config.desc_act
elif name == 'model.layers.1.self_attn.qkv_proj':
# The second layer is quantized using bits=8, group_size=32
# desc_act=False
assert isinstance(submodule.quant_method,
linear_method_cls)
config = submodule.quant_method.quant_config
assert get_dynamic_override(config,
layer_name=name,
key="bits") == 8
assert get_dynamic_override(config,
layer_name=name,
key="group_size") == 32
assert not get_dynamic_override(
config, layer_name=name, key="desc_act")
elif (name == 'model.layers.2.self_attn.qkv_proj'
or name == 'model.layers.2.mlp.gate_up_proj'):
# All other layers (layer index >= 2) are not quantized
assert isinstance(submodule.quant_method,
UnquantizedLinearMethod)
llm.apply_model(check_model)