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

[CI/Test] improve robustness of test (vllm_runner) (#5357)

Browse Source 
[CI/Test] improve robustness of test by replacing del with context manager (vllm_runner) (#5357)
This commit is contained in:
youkaichao
2024-06-08 01:59:20 -07:00
committed by GitHub
parent 9fb900f90c
commit 8ea5e44a43
28 changed files with 431 additions and 470 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
tests/quantization/test_compressed_tensors.py
View File

@@ -12,42 +12,45 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso
def test_compressed_tensors_w8a8_static_setup(vllm_runner):
model_path = "nm-testing/tinyllama-one-shot-static-quant-test-compressed"
llm = vllm_runner(model_path, quantization="sparseml", enforce_eager=True)
model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model
layer = model.model.layers[0]
with vllm_runner(model_path, quantization="sparseml",
enforce_eager=True) as llm:
model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501
layer = model.model.layers[0]
qkv_proj = layer.self_attn.qkv_proj
o_proj = layer.self_attn.o_proj
gate_up_proj = layer.mlp.gate_up_proj
down_proj = layer.mlp.down_proj
qkv_proj = layer.self_attn.qkv_proj
o_proj = layer.self_attn.o_proj
gate_up_proj = layer.mlp.gate_up_proj
down_proj = layer.mlp.down_proj
assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(o_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(gate_up_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(down_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(o_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(gate_up_proj.quant_method,
CompressedTensorsLinearMethod)
assert isinstance(down_proj.quant_method,
CompressedTensorsLinearMethod)
assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8StaticTensor)
assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8StaticTensor)
assert qkv_proj.weight.dtype is torch.int8
assert o_proj.weight.dtype is torch.int8
assert gate_up_proj.weight.dtype is torch.int8
assert qkv_proj.weight.dtype is torch.int8
assert o_proj.weight.dtype is torch.int8
assert gate_up_proj.weight.dtype is torch.int8
assert qkv_proj.weight_scale.shard_splitter is not None
assert qkv_proj.weight_scale.logical_widths is not None
assert qkv_proj.input_scale.dtype is torch.float32
assert qkv_proj.weight_scale.shard_splitter is not None
assert qkv_proj.weight_scale.logical_widths is not None
assert qkv_proj.input_scale.dtype is torch.float32
def test_compressed_tensors_w8a8_dynanmic_per_token(vllm_runner):
model_path = "nm-testing/tinyllama-one-shot-dynamic-test"
llm = vllm_runner(model_path,
quantization="sparseml",
enforce_eager=True,
dtype=torch.float16)
model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model
layer = model.model.layers[0]
with vllm_runner(model_path,
quantization="sparseml",
enforce_eager=True,
dtype=torch.float16) as llm:
model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501
layer = model.model.layers[0]
qkv_proj = layer.self_attn.qkv_proj
qkv_proj = layer.self_attn.qkv_proj
assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8DynamicToken)
assert qkv_proj.weight.dtype is torch.int8
assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8DynamicToken)
assert qkv_proj.weight.dtype is torch.int8