[torch.compile] Add torch inductor pass for fusing silu_and_mul with subsequent scaled_fp8_quant operations (#10867)

Signed-off-by: Sage Moore <sage@neuralmagic.com>
This commit is contained in:
Sage Moore
2025-05-01 07:59:28 -07:00
committed by GitHub
parent 28566d73b3
commit 460a2b1100
11 changed files with 406 additions and 9 deletions
+14 -3
View File
@@ -5,6 +5,7 @@ import torch
import vllm.envs as envs
from vllm import LLM, SamplingParams
from vllm.compilation.activation_quant_fusion import ActivationQuantFusionPass
from vllm.compilation.fix_functionalization import FixFunctionalizationPass
from vllm.compilation.fusion import (FUSED_OPS, FusionPass, QuantKey,
kFp8DynamicTokenSym, kFp8StaticTensorSym)
@@ -17,7 +18,6 @@ from .backend import TestBackend
OPS_IN_MODEL = [
torch.ops._C.rotary_embedding.default,
torch.ops._C.fused_add_rms_norm.default,
torch.ops._C.silu_and_mul.default,
]
RMS_OP = torch.ops._C.rms_norm.default
@@ -29,6 +29,9 @@ RMS_QUANT_OPS = {
],
}
SILU_MUL_OP = torch.ops._C.silu_and_mul.default
SILU_MUL_QUANT_OP = torch.ops._C.silu_and_mul_quant.default
prompts = [
"Hello, my name is",
"The president of the United States is",
@@ -55,8 +58,10 @@ def test_fix_functionalization(model: str, quant_key: QuantKey,
enable_noop=True))
noop_pass = NoOpEliminationPass(vllm_config)
fusion_pass = FusionPass.instance(vllm_config)
act_quant_fusion_pass = ActivationQuantFusionPass(vllm_config)
passes = [noop_pass, fusion_pass] if do_fusion else [noop_pass]
passes = [noop_pass, fusion_pass, act_quant_fusion_pass
] if do_fusion else [noop_pass]
func_pass = FixFunctionalizationPass(vllm_config)
backend_func = TestBackend(*passes, func_pass)
backend_no_func = TestBackend(*passes)
@@ -79,6 +84,7 @@ def test_fix_functionalization(model: str, quant_key: QuantKey,
model_runner.model = torch.compile(orig_model,
fullgraph=True,
backend=backend_no_func)
gen_no_func = llm.generate(prompts, sampling_params)
for output_func, output_no_func in zip(gen_func, gen_no_func):
@@ -88,7 +94,12 @@ def test_fix_functionalization(model: str, quant_key: QuantKey,
# and replaced by fused quantized ops in RMS_QUANT_OPS.
rms_ops = [FUSED_OPS[(quant_key, True)], FUSED_OPS[(quant_key, False)]
] if do_fusion else [RMS_OP]
ops = OPS_IN_MODEL + rms_ops
silu_mul_ops = [SILU_MUL_QUANT_OP] if do_fusion and \
quant_key == kFp8StaticTensorSym else [
SILU_MUL_OP
]
ops = OPS_IN_MODEL + rms_ops + silu_mul_ops
for op in ops:
find_auto_fn(backend_no_func.graph_post_pass.nodes, op)
@@ -0,0 +1,74 @@
# SPDX-License-Identifier: Apache-2.0
import pytest
import torch
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.model_executor.layers.activation import SiluAndMul
from .backend import TestBackend
class TestModel(torch.nn.Module):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.silu_and_mul = SiluAndMul()
self.scale = torch.rand(1, dtype=torch.float32)
def forward(self, x):
y = self.silu_and_mul(x)
x2 = scaled_fp8_quant(y, self.scale)
return x2
@pytest.mark.parametrize("num_tokens", [256])
@pytest.mark.parametrize("hidden_size", [64])
@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE != "cuda",
reason="Only test on CUDA")
def test_fusion_silu_and_mul_quant(num_tokens, hidden_size):
torch.set_default_device("cuda")
torch.set_default_dtype(torch.float16)
# 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))
fusion_pass = ActivationQuantFusionPass(config)
backend = TestBackend(fusion_pass)
model = TestModel()
# First dimension dynamic
x = torch.rand(num_tokens, hidden_size)
torch._dynamo.mark_dynamic(x, 0)
result = model(x)
model2 = torch.compile(model, backend=backend)
result2 = model2(x)
# Check that it gives the same answer
torch.testing.assert_close(result[0].to(dtype=torch.float16),
result2[0].to(dtype=torch.float16),
atol=1e-3,
rtol=1e-3)
# Check substitution worked
pre_nodes = backend.graph_pre_pass.nodes
post_nodes = backend.graph_post_pass.nodes
silu_and_mul_quant = torch.ops._C.silu_and_mul_quant.default
fp8_quant = torch.ops._C.static_scaled_fp8_quant.default
# In pre-nodes, fp8 quant should be present and fused kernels should not
assert find_auto_fn_maybe(pre_nodes, silu_and_mul_quant) is None
find_auto_fn(pre_nodes, fp8_quant)
# In post-nodes, fused kernels should be present and fp8 quant should not
find_auto_fn(post_nodes, silu_and_mul_quant)
assert find_auto_fn_maybe(post_nodes, fp8_quant) is None
@@ -0,0 +1,69 @@
# SPDX-License-Identifier: Apache-2.0
import pytest
import torch
import vllm._custom_ops as ops
from tests.kernels.utils import opcheck
from vllm.model_executor.layers.activation import SiluAndMul
DTYPES = [torch.bfloat16, torch.float16]
QUANT_DTYPES = [torch.float8_e4m3fn]
NUM_TOKENS = [1, 17, 86, 1234, 3045] # Arbitrary values for testing
HIDDEN_SIZES = [16, 48, 128, 1562, 4096] # Arbitrary values for testing
SEEDS = [0]
CUDA_DEVICES = [
f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
]
def ref_impl(silu_and_mul: SiluAndMul, x: torch.Tensor,
scale: torch.Tensor) -> torch.Tensor:
silu_and_mul_out = silu_and_mul.forward_native(x)
out, scales = ops.scaled_fp8_quant(silu_and_mul_out, scale)
return out
def ops_impl(x: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
out_shape = (x.shape[0], x.shape[1] // 2)
out = torch.empty(out_shape,
dtype=torch.torch.float8_e4m3fn,
device=x.device)
torch.ops._C.silu_and_mul_quant(out, x, scale)
return out
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("quant_dtype", QUANT_DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@torch.inference_mode()
def test_silu_and_mul(
num_tokens: int,
hidden_size: int,
dtype: torch.dtype,
quant_dtype: torch.dtype,
seed: int,
device: str,
) -> None:
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.set_default_device(device)
layer = SiluAndMul()
# Make inputs
scale = (torch.randn((1), device=device, dtype=torch.float32))
x = torch.randn(num_tokens, hidden_size, dtype=dtype)
ref_out = ref_impl(layer, x, scale)
ops_out = ops_impl(x, scale)
assert ref_out.dtype == quant_dtype
assert ops_out.dtype == quant_dtype
assert ref_out.shape == ops_out.shape
assert torch.allclose(ref_out.to(dtype=torch.float32),
ops_out.to(dtype=torch.float32))
opcheck(torch.ops._C.silu_and_mul_quant, (ops_out, x, scale))