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

[Core] Rename PassConfig flags as per RFC #27995 (#29646)

Browse Source 
Signed-off-by: arpitkh101 <arpit5khandelwal@gmail.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
(cherry picked from commit d7284a2604)
This commit is contained in:
Arpit Khandelwal
2025-12-02 22:38:55 -05:00
committed by Kevin H. Luu
parent a1d627e40f
commit 4fd9d6a85c
22 changed files with 318 additions and 123 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/compile/distributed/test_async_tp.py
View File

@@ -326,7 +326,7 @@ def async_tp_pass_on_test_model(
vllm_config = VllmConfig()
vllm_config.compilation_config = CompilationConfig(
pass_config=PassConfig(
enable_async_tp=True,
fuse_gemm_comms=True,
),
)
vllm_config.device_config = DeviceConfig(device=torch.device("cuda"))
@@ -413,7 +413,7 @@ def test_async_tp_pass_correctness(
"mode": CompilationMode.VLLM_COMPILE,
"compile_sizes": [2, 4, 8],
"splitting_ops": [],
"pass_config": {"enable_async_tp": async_tp_enabled},
"pass_config": {"fuse_gemm_comms": async_tp_enabled},
}
async_tp_args = [

2
tests/compile/distributed/test_fusion_all_reduce.py
View File

@@ -295,7 +295,7 @@ def all_reduce_fusion_pass_on_test_model(
)
)
vllm_config.compilation_config.pass_config = PassConfig(
enable_fi_allreduce_fusion=True, enable_noop=True
fuse_allreduce_rms=True, eliminate_noops=True
)
vllm_config.device_config = DeviceConfig(device=torch.device("cuda"))
vllm_config.parallel_config.rank = local_rank # Setup rank for debug path

16
tests/compile/distributed/test_fusions_e2e.py
View File

@@ -192,7 +192,7 @@ def test_attn_quant(
splitting_ops=splitting_ops,
# Common
mode=CompilationMode.VLLM_COMPILE,
pass_config=PassConfig(enable_attn_fusion=True, enable_noop=True),
pass_config=PassConfig(fuse_attn_quant=True, eliminate_noops=True),
# Inductor caches custom passes by default as well via uuid
inductor_compile_config={"force_disable_caches": True},
)
@@ -282,9 +282,9 @@ def test_tp2_attn_quant_allreduce_rmsnorm(
# Common
mode=CompilationMode.VLLM_COMPILE,
pass_config=PassConfig(
enable_attn_fusion=True,
enable_noop=True,
enable_fi_allreduce_fusion=True,
fuse_attn_quant=True,
eliminate_noops=True,
fuse_allreduce_rms=True,
),
# Inductor caches custom passes by default as well via uuid
inductor_compile_config={"force_disable_caches": True},
@@ -384,10 +384,10 @@ def test_tp2_attn_quant_async_tp(
# Common
level=CompilationMode.VLLM_COMPILE,
pass_config=PassConfig(
enable_attn_fusion=True,
enable_noop=True,
enable_sequence_parallelism=True,
enable_async_tp=True,
fuse_attn_quant=True,
eliminate_noops=True,
enable_sp=True,
fuse_gemm_comms=True,
),
# Inductor caches custom passes by default as well via uuid
inductor_compile_config={"force_disable_caches": True},

18
tests/compile/distributed/test_sequence_parallelism.py
View File

@@ -153,7 +153,7 @@ class TestAllReduceRMSNormStaticQuantFP8Model(torch.nn.Module):
]
def ops_in_model(self):
if self.vllm_config.compilation_config.pass_config.enable_fusion:
if self.vllm_config.compilation_config.pass_config.fuse_norm_quant:
return [torch.ops._C.fused_add_rms_norm_static_fp8_quant.default]
elif RMSNorm.enabled():
return [
@@ -183,7 +183,7 @@ class TestAllReduceRMSNormStaticQuantFP8Model(torch.nn.Module):
@pytest.mark.parametrize("seq_len", [16])
@pytest.mark.parametrize("hidden_size", [16])
@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
@pytest.mark.parametrize("enable_fusion", [True, False])
@pytest.mark.parametrize("fuse_norm_quant", [True, False])
@pytest.mark.parametrize("dynamic", [False, True])
@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"], reason="Only test on CUDA")
def test_sequence_parallelism_pass(
@@ -193,7 +193,7 @@ def test_sequence_parallelism_pass(
seq_len: int,
hidden_size: int,
dtype: torch.dtype,
enable_fusion: bool,
fuse_norm_quant: bool,
dynamic: bool,
):
num_processes = 2
@@ -211,7 +211,7 @@ def test_sequence_parallelism_pass(
seq_len,
hidden_size,
dtype,
enable_fusion,
fuse_norm_quant,
dynamic,
),
nprocs=nprocs,
@@ -229,7 +229,7 @@ def sequence_parallelism_pass_on_test_model(
seq_len: int,
hidden_size: int,
dtype: torch.dtype,
enable_fusion: bool,
fuse_norm_quant: bool,
dynamic: bool,
):
current_platform.seed_everything(0)
@@ -260,9 +260,9 @@ def sequence_parallelism_pass_on_test_model(
cudagraph_mode=CUDAGraphMode.NONE, # avoid piecewise warnings
custom_ops=custom_ops_list,
pass_config=PassConfig(
enable_sequence_parallelism=True,
enable_fusion=enable_fusion,
enable_noop=True,
enable_sp=True,
fuse_norm_quant=fuse_norm_quant,
eliminate_noops=True,
),
) # NoOp needed for fusion
device_config = DeviceConfig(device=torch.device("cuda"))
@@ -297,7 +297,7 @@ def sequence_parallelism_pass_on_test_model(
sequence_parallelism_pass,
]
if enable_fusion:
if fuse_norm_quant:
fusion_pass = RMSNormQuantFusionPass(vllm_config)
passes_for_backend.append(fusion_pass)