Improve configs - the rest! (#17562)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -9,7 +9,7 @@ import torch
|
||||
|
||||
from tests.quantization.utils import is_quant_method_supported
|
||||
from vllm import LLM, SamplingParams
|
||||
from vllm.config import CompilationConfig, CompilationLevel
|
||||
from vllm.config import CompilationConfig, CompilationLevel, PassConfig
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
from ..utils import create_new_process_for_each_test
|
||||
@@ -95,9 +95,6 @@ def test_full_graph(
|
||||
run_model(optimization_level, model, model_kwargs)
|
||||
|
||||
|
||||
PassConfig = CompilationConfig.PassConfig
|
||||
|
||||
|
||||
# TODO(luka) add other supported compilation config scenarios here
|
||||
@pytest.mark.parametrize(
|
||||
"compilation_config, model_info",
|
||||
|
||||
@@ -11,7 +11,7 @@ from vllm.compilation.fusion import (FUSED_OPS, FusionPass, QuantKey,
|
||||
kFp8DynamicTokenSym, kFp8StaticTensorSym)
|
||||
from vllm.compilation.fx_utils import find_auto_fn, find_auto_fn_maybe, is_func
|
||||
from vllm.compilation.noop_elimination import NoOpEliminationPass
|
||||
from vllm.config import CompilationConfig, VllmConfig
|
||||
from vllm.config import CompilationConfig, PassConfig, VllmConfig
|
||||
|
||||
from .backend import TestBackend
|
||||
|
||||
@@ -53,9 +53,8 @@ def test_fix_functionalization(model: str, quant_key: QuantKey,
|
||||
torch.set_default_device("cuda")
|
||||
|
||||
vllm_config = VllmConfig()
|
||||
vllm_config.compilation_config = CompilationConfig(pass_config= \
|
||||
CompilationConfig.PassConfig(enable_fusion=do_fusion,
|
||||
enable_noop=True))
|
||||
vllm_config.compilation_config = CompilationConfig(
|
||||
pass_config=PassConfig(enable_fusion=do_fusion, enable_noop=True))
|
||||
noop_pass = NoOpEliminationPass(vllm_config)
|
||||
fusion_pass = FusionPass.instance(vllm_config)
|
||||
act_quant_fusion_pass = ActivationQuantFusionPass(vllm_config)
|
||||
|
||||
@@ -9,7 +9,8 @@ from vllm.compilation.fusion import (FUSED_OPS, QUANT_OPS, FusedRMSQuantKey,
|
||||
FusionPass, QuantKey)
|
||||
from vllm.compilation.fx_utils import find_auto_fn, find_auto_fn_maybe
|
||||
from vllm.compilation.noop_elimination import NoOpEliminationPass
|
||||
from vllm.config import CompilationConfig, CompilationLevel, VllmConfig
|
||||
from vllm.config import (CompilationConfig, CompilationLevel, PassConfig,
|
||||
VllmConfig)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||
CUTLASS_FP8_SUPPORTED, Fp8LinearOp, maybe_create_device_identity)
|
||||
@@ -78,8 +79,7 @@ def test_fusion_rmsnorm_quant(dtype, hidden_size, num_tokens, eps, static,
|
||||
vllm_config = VllmConfig(compilation_config=CompilationConfig(
|
||||
level=CompilationLevel.PIECEWISE, custom_ops=["+rms_norm"]))
|
||||
vllm_config.compilation_config.pass_config = \
|
||||
CompilationConfig.PassConfig(enable_fusion=True,
|
||||
enable_noop=True)
|
||||
PassConfig(enable_fusion=True, enable_noop=True)
|
||||
with vllm.config.set_current_vllm_config(vllm_config):
|
||||
# Reshape pass is needed for the fusion pass to work
|
||||
noop_pass = NoOpEliminationPass(vllm_config)
|
||||
|
||||
@@ -10,7 +10,7 @@ from vllm.compilation.fx_utils import (find_auto_fn, find_auto_fn_maybe,
|
||||
find_specified_fn_maybe, is_func)
|
||||
from vllm.compilation.sequence_parallelism import SequenceParallelismPass
|
||||
from vllm.config import (CompilationConfig, DeviceConfig, ModelConfig,
|
||||
VllmConfig)
|
||||
PassConfig, VllmConfig)
|
||||
from vllm.distributed import tensor_model_parallel_all_reduce
|
||||
from vllm.distributed.parallel_state import (init_distributed_environment,
|
||||
initialize_model_parallel)
|
||||
@@ -126,9 +126,8 @@ def sequence_parallelism_pass_on_test_model(local_rank: int, world_size: int,
|
||||
|
||||
# configure vllm config for SequenceParallelismPass
|
||||
vllm_config = VllmConfig()
|
||||
vllm_config.compilation_config = CompilationConfig(
|
||||
pass_config=CompilationConfig.PassConfig(
|
||||
enable_sequence_parallelism=True, ), )
|
||||
vllm_config.compilation_config = CompilationConfig(pass_config=PassConfig(
|
||||
enable_sequence_parallelism=True))
|
||||
vllm_config.device_config = DeviceConfig(device=torch.device("cuda"))
|
||||
|
||||
# this is a fake model name to construct the model config
|
||||
|
||||
@@ -6,7 +6,7 @@ import vllm.envs as envs
|
||||
from vllm._custom_ops import scaled_fp8_quant
|
||||
from vllm.compilation.activation_quant_fusion import ActivationQuantFusionPass
|
||||
from vllm.compilation.fx_utils import find_auto_fn, find_auto_fn_maybe
|
||||
from vllm.config import CompilationConfig, VllmConfig
|
||||
from vllm.config import CompilationConfig, PassConfig, VllmConfig
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
|
||||
from .backend import TestBackend
|
||||
@@ -36,8 +36,7 @@ def test_fusion_silu_and_mul_quant(num_tokens, hidden_size):
|
||||
# Reshape pass is needed for the fusion pass to work
|
||||
config = VllmConfig()
|
||||
config.compilation_config = CompilationConfig(
|
||||
pass_config=CompilationConfig.PassConfig(enable_fusion=True,
|
||||
enable_reshape=True))
|
||||
pass_config=PassConfig(enable_fusion=True, enable_reshape=True))
|
||||
fusion_pass = ActivationQuantFusionPass(config)
|
||||
|
||||
backend = TestBackend(fusion_pass)
|
||||
|
||||
Reference in New Issue
Block a user