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

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
tests/compile/test_full_graph.py
View File

@@ -14,8 +14,7 @@ from tests.quantization.utils import is_quant_method_supported
from vllm import LLM, SamplingParams
from vllm.attention.backends.registry import _Backend
from vllm.attention.selector import global_force_attn_backend_context_manager
from vllm.config import (CompilationConfig, CompilationLevel, CUDAGraphMode,
PassConfig)
from vllm.config import CompilationConfig, CompilationLevel, CUDAGraphMode, PassConfig
from vllm.platforms import current_platform
from vllm.utils import is_torch_equal_or_newer
@@ -25,43 +24,54 @@ from ..utils import create_new_process_for_each_test
def models_list(*, all: bool = True, keywords: Optional[list[str]] = None):
TEST_MODELS: list[tuple[str, dict[str, Any]]] = [
("facebook/opt-125m", {}),
("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", {
"dtype": torch.float16,
}),
("neuralmagic/Llama-3.2-1B-Instruct-FP8-dynamic", {
"dtype": torch.float16,
}),
(
"nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change",
{
"dtype": torch.float16,
},
),
(
"neuralmagic/Llama-3.2-1B-Instruct-FP8-dynamic",
{
"dtype": torch.float16,
},
),
("neuralmagic/Llama-3.2-1B-Instruct-quantized.w8a8", {}),
("meta-llama/Llama-3.2-1B-Instruct", {}),
]
if all:
# TODO: figure out why this fails.
if False and is_quant_method_supported("gguf"): # noqa: SIM223
TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", {
"quantization": "gguf"
}))
TEST_MODELS.append(
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", {"quantization": "gguf"})
)
if is_quant_method_supported("gptq"):
TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", {
"quantization": "gptq"
}))
TEST_MODELS.append(
("TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", {"quantization": "gptq"})
)
if is_quant_method_supported("gptq_marlin"):
TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", {
"quantization": "gptq_marlin"
}))
TEST_MODELS.append(
(
"TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ",
{"quantization": "gptq_marlin"},
)
)
if is_quant_method_supported("gptq_marlin_24"):
TEST_MODELS.append(("alexm-nm/tinyllama-24-marlin24-4bit-g128", {
"quantization": "gptq_marlin_24"
}))
TEST_MODELS.append(
(
"alexm-nm/tinyllama-24-marlin24-4bit-g128",
{"quantization": "gptq_marlin_24"},
)
)
if not current_platform.is_rocm() and is_quant_method_supported("awq"):
TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", {
"quantization": "AWQ"
}))
TEST_MODELS.append(
("TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", {"quantization": "AWQ"})
)
if keywords is None:
return TEST_MODELS
@@ -95,22 +105,34 @@ def test_full_graph(
"compilation_config, model_info",
[
# additional compile sizes, only some of the models
(CompilationConfig(level=CompilationLevel.PIECEWISE,
compile_sizes=[1, 2]), model)
(
CompilationConfig(level=CompilationLevel.PIECEWISE, compile_sizes=[1, 2]),
model,
)
for model in models_list(all=False)
] + [
]
+ [
# RMSNorm + quant fusion, only 8-bit quant models
(CompilationConfig(level=CompilationLevel.PIECEWISE,
custom_ops=["+rms_norm"],
pass_config=PassConfig(enable_fusion=True,
enable_noop=True)), model)
(
CompilationConfig(
level=CompilationLevel.PIECEWISE,
custom_ops=["+rms_norm"],
pass_config=PassConfig(enable_fusion=True, enable_noop=True),
),
model,
)
for model in models_list(keywords=["FP8-dynamic", "quantized.w8a8"])
] + [
]
+ [
# Test depyf integration works
(CompilationConfig(level=CompilationLevel.PIECEWISE,
debug_dump_path=tempfile.gettempdir()),
("facebook/opt-125m", {})),
] + [
(
CompilationConfig(
level=CompilationLevel.PIECEWISE, debug_dump_path=tempfile.gettempdir()
),
("facebook/opt-125m", {}),
),
]
+ [
# graph inductor partition
(
CompilationConfig(
@@ -119,20 +141,24 @@ def test_full_graph(
# torch._C.Tag.cudagraph_unsafe to specify splitting ops
use_inductor_graph_partition=True,
cudagraph_mode=CUDAGraphMode.PIECEWISE,
compile_sizes=[1, 2]),
model) for model in models_list(all=False)
compile_sizes=[1, 2],
),
model,
)
for model in models_list(all=False)
if is_torch_equal_or_newer("2.9.0.dev")
])
],
)
# only test some of the models
@create_new_process_for_each_test()
def test_custom_compile_config(
compilation_config: CompilationConfig,
model_info: tuple[str, dict[str, Any]],
):
if (compilation_config.use_inductor_graph_partition
and not is_torch_equal_or_newer("2.9.0.dev")):
pytest.skip("inductor graph partition is only available "
"in PyTorch 2.9+")
if compilation_config.use_inductor_graph_partition and not is_torch_equal_or_newer(
"2.9.0.dev"
):
pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
model, model_kwargs = model_info
print(f"MODEL={model}")
@@ -156,8 +182,7 @@ def test_fp8_kv_scale_compile(optimization_level: int):
def test_inductor_graph_partition_attn_fusion(caplog_vllm):
if not is_torch_equal_or_newer("2.9.0.dev"):
pytest.skip("inductor graph partition is only available "
"in PyTorch 2.9+")
pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
model = "nvidia/Llama-4-Scout-17B-16E-Instruct-FP8"
compilation_config = CompilationConfig(
@@ -171,14 +196,16 @@ def test_inductor_graph_partition_attn_fusion(caplog_vllm):
"kv_cache_dtype": "fp8",
"max_model_len": 1024,
}
with caplog_vllm.at_level(
logging.DEBUG), global_force_attn_backend_context_manager(
_Backend.FLASHINFER):
with (
caplog_vllm.at_level(logging.DEBUG),
global_force_attn_backend_context_manager(_Backend.FLASHINFER),
):
run_model(compilation_config, model, model_kwargs)
try:
assert ("Fused quantization onto 48 attention nodes"
in caplog_vllm.text), caplog_vllm.text
assert "Fused quantization onto 48 attention nodes" in caplog_vllm.text, (
caplog_vllm.text
)
except AssertionError:
# Note: this message is only triggered when the compilation goes
# through the custom pass. Due to multiple layers of cache on
@@ -189,8 +216,11 @@ def test_inductor_graph_partition_attn_fusion(caplog_vllm):
assert "Fused quantization" not in caplog_vllm.text
def run_model(compile_config: Union[int, CompilationConfig], model: str,
model_kwargs: dict[str, Any]):
def run_model(
compile_config: Union[int, CompilationConfig],
model: str,
model_kwargs: dict[str, Any],
):
prompts = [
"Hello, my name is",
"The president of the United States is",