A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
Remove last `level` references not removed in #26355 (#27260)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Oct 22, 2025 at 10:18 UTC
8f18feb191d731dd4be000377a18ffb84ed8dc18
6 files changed
+13
-13
tests/compile/piecewise/test_toy_llama.py
+2
-2
@@ -355,13 +355,13 @@ def test_toy_llama(
355
)
356
357
compile_config_no_compile = CompilationConfig(
358
- level=CompilationMode.NONE,
358
+ mode=CompilationMode.NONE,
359
cudagraph_mode=CUDAGraphMode.NONE,
360
backend="eager",
361
)
362
363
compile_config_no_split = CompilationConfig(
364
- level=CompilationMode.VLLM_COMPILE,
364
+ mode=CompilationMode.VLLM_COMPILE,
365
use_inductor_graph_partition=use_inductor_graph_partition,
366
cudagraph_mode=CUDAGraphMode.PIECEWISE,
367
backend=backend,
tests/compile/test_aot_compile.py
+1
-1
@@ -38,7 +38,7 @@ class CompiledMod(torch.nn.Module):
38
def make_vllm_config() -> VllmConfig:
39
return VllmConfig(
40
compilation_config=CompilationConfig(
41
- level=CompilationMode.VLLM_COMPILE,
41
+ mode=CompilationMode.VLLM_COMPILE,
42
)
43
)
44
tests/compile/test_config.py
+3
-3
@@ -168,7 +168,7 @@ def test_splitting_ops_dynamic():
168
if is_torch_equal_or_newer("2.9.0.dev"):
169
config = VllmConfig(
170
compilation_config=CompilationConfig(
171
- level=CompilationMode.VLLM_COMPILE,
171
+ mode=CompilationMode.VLLM_COMPILE,
172
use_inductor_graph_partition=True,
173
splitting_ops=["vllm::unified_attention"],
174
)
@@ -180,7 +180,7 @@ def test_splitting_ops_dynamic():
180
# When attn_fusion pass enabled, splitting_ops now default to attention ops.
181
config = VllmConfig(
182
compilation_config=CompilationConfig(
183
- level=CompilationMode.VLLM_COMPILE,
183
+ mode=CompilationMode.VLLM_COMPILE,
184
pass_config={"enable_attn_fusion": True, "enable_noop": True},
185
custom_ops=["+quant_fp8"],
186
cudagraph_mode=CUDAGraphMode.PIECEWISE,
@@ -195,7 +195,7 @@ def test_splitting_ops_dynamic():
195
if is_torch_equal_or_newer("2.9.0.dev"):
196
config = VllmConfig(
197
compilation_config=CompilationConfig(
198
- level=CompilationMode.VLLM_COMPILE,
198
+ mode=CompilationMode.VLLM_COMPILE,
199
use_inductor_graph_partition=True,
200
pass_config={"enable_attn_fusion": True, "enable_noop": True},
201
custom_ops=["+quant_fp8"],
tests/compile/test_full_graph.py
+1
-1
@@ -198,7 +198,7 @@ def run_model(compile_config: int | CompilationConfig, model: str, **model_kwarg
198
compilation_config = (
199
compile_config
200
if isinstance(compile_config, CompilationConfig)
201
- else CompilationConfig(level=compile_config)
201
+ else CompilationConfig(mode=compile_config)
202
)
203
204
prompts = [
tests/compile/test_fusions_e2e.py
+3
-3
@@ -151,7 +151,7 @@ def test_attn_quant(
151
cudagraph_mode=mode,
152
splitting_ops=splitting_ops,
153
# Common
154
- level=CompilationMode.VLLM_COMPILE,
154
+ mode=CompilationMode.VLLM_COMPILE,
155
pass_config=PassConfig(enable_attn_fusion=True, enable_noop=True),
156
# Inductor caches custom passes by default as well via uuid
157
inductor_compile_config={"force_disable_caches": True},
@@ -236,7 +236,7 @@ def test_tp2_attn_quant_allreduce_rmsnorm(
236
custom_ops=custom_ops_list,
237
splitting_ops=splitting_ops,
238
# Common
239
- level=CompilationMode.VLLM_COMPILE,
239
+ mode=CompilationMode.VLLM_COMPILE,
240
pass_config=PassConfig(
241
enable_attn_fusion=True,
242
enable_noop=True,
@@ -273,7 +273,7 @@ def run_model(compile_config: int | CompilationConfig, model: str, **model_kwarg
273
compilation_config = (
274
compile_config
275
if isinstance(compile_config, CompilationConfig)
276
- else CompilationConfig(level=compile_config)
276
+ else CompilationConfig(mode=compile_config)
277
)
278
279
prompts = [
tests/model_executor/test_enabled_custom_ops.py
+3
-3
@@ -36,7 +36,7 @@ class Relu3(ReLUSquaredActivation):
36
37
38
@pytest.mark.parametrize(
39
- "env, torch_level, backend, ops_enabled, default_on",
39
+ "env, compilation_mode, backend, ops_enabled, default_on",
40
[
41
# Default values based on compile level
42
# - All by default (no Inductor compilation)
@@ -77,7 +77,7 @@ class Relu3(ReLUSquaredActivation):
77
)
78
def test_enabled_ops(
79
env: str | None,
80
- torch_level: int,
80
+ compilation_mode: int,
81
backend: str,
82
ops_enabled: list[int],
83
default_on: bool,
@@ -85,7 +85,7 @@ def test_enabled_ops(
85
custom_ops = env.split(",") if env else []
86
vllm_config = VllmConfig(
87
compilation_config=CompilationConfig(
88
- backend=backend, level=torch_level, custom_ops=custom_ops
88
+ backend=backend, mode=compilation_mode, custom_ops=custom_ops
89
)
90
)
91
with set_current_vllm_config(vllm_config):