The inductor pattern matcher rejects match due to argument named `out` in the 2.5 nightly

[bnellnm]

🐛 Describe the bug

The inductor pattern matcher is rejecting a valid match due to an argument being named out. The final return statement in torch/_inductor/pattern_matcher.py is where this happens:

def is_mutation_op(node: torch.fx.Node) -> bool:
    if node.op == "call_function":
        if _mutation_op_re.search(node.target.__name__):  # type: ignore[union-attr]                                                                             
            return True
    elif node.op == "call_method":
        if _mutation_op_re.search(node.target):  # type: ignore[union-attr, arg-type]                                                                            
            return True
    return node.kwargs.get("out") is not None

The following script tries to do a replacement with custom ops that have out as an argument name. Even though they have been functionalized at the point of replacement, the matcher skips over these.

from typing import List, Callable, Optional, Tuple

import torch
from torch._inductor.pattern_matcher import PatternMatcherPass, register_replacement, fwd_only
from torch._higher_order_ops.auto_functionalize import auto_functionalized

torch.set_default_device("cuda")


@torch.library.custom_op("vllm::fused_rms_norm_quant_static", mutates_args=['out'])
def fused_rms_norm_quant_static(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor, scale: torch.Tensor,
                                azp: torch.Tensor, epsilon: float) -> None:
    print("vllm::fused_rms_norm_quant_static")
    # bogus implementation doesn't matter
    result_rms = torch.mul(input, weight) + epsilon
    out = torch.mul(result_rms, scale).to(torch.int8)


@torch.library.custom_op("vllm::rms_norm", mutates_args=['out'])
def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
              epsilon: float) -> None:
    # bogus implementation doesn't matter
    out = torch.mul(input, weight) + epsilon


@torch.library.custom_op("vllm::static_scaled_int8_quant", mutates_args=['out'])
def static_scaled_int8_quant(out: torch.Tensor,
                             input: torch.Tensor,
                             scale: torch.Tensor,
                             azp: Optional[torch.Tensor] = None) -> None:
    # bogus implementation doesn't matter
    out = torch.mul(input, scale).to(torch.int8)


def rms_pattern_static(out: torch.Tensor, result_rms: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
                       scale: torch.Tensor):
    at1 = auto_functionalized(torch.ops.vllm.rms_norm.default, out=result_rms, input=input, weight=weight, epsilon=1e-6)
    at2 = auto_functionalized(torch.ops.vllm.static_scaled_int8_quant.default, out=out, input=at1[1], scale=scale,
                              azp=None)

    # result
    return at2[1]


def rms_replacement_static(result: torch.Tensor, result_rms: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
                           scale: torch.Tensor):
    at = auto_functionalized(torch.ops.vllm.fused_rms_norm_quant_static.default, out=result, input=input, weight=weight,
                             epsilon=1e-6, scale=scale, azp=None)

    # result
    return at[1]

def empty_bf16(*args, **kwargs):
    return torch.empty(*args, **kwargs, dtype=torch.bfloat16)


def empty_int8(*args, **kwargs):
    return torch.empty(*args, **kwargs, dtype=torch.int8)


my_patterns = PatternMatcherPass()
inputs = [empty_int8(5, 4), empty_bf16(5, 4), empty_bf16(5, 4), empty_bf16(5, 1), torch.empty(1, 1)]
register_replacement(rms_pattern_static, rms_replacement_static, inputs, fwd_only, my_patterns)

def custom_pass(graph: torch.fx.Graph) -> torch.fx.Graph:
    #print("Pre-pass:")
    #print(graph)
    count = my_patterns.apply(graph)
    print(f"Count: {count}")
    #print(graph)
    return graph

def custom_backend(graph: torch.fx.GraphModule, example_inputs: List[torch.Tensor]) -> Callable:
    from torch._inductor import config
    current_config = config.shallow_copy_dict()
    from torch._inductor.compile_fx import compile_fx
    current_config['post_grad_custom_post_pass'] = custom_pass
    return compile_fx(graph, example_inputs, config_patches=current_config)

@torch.compile(backend=custom_backend)
def my_func_static(x, w, epsilon):
    result = torch.empty_like(x)
    torch.ops.vllm.rms_norm(result, x, w, epsilon)
    scale = torch.ones((1, 1))
    quant_result = torch.empty_like(x, dtype=torch.int8)
    torch.ops.vllm.static_scaled_int8_quant(quant_result, result, scale, None)
    return quant_result


print("Run my_func_static")
inputs = [torch.empty((5, 4)), torch.empty((5, 1)), 1e-6]
my_func_static(*inputs)

This script is the same as above but with out renamed to result. The matcher works as expected here:

from typing import List, Callable, Optional, Tuple

import torch
from torch._inductor.pattern_matcher import PatternMatcherPass, register_replacement, fwd_only
from torch._higher_order_ops.auto_functionalize import auto_functionalized

torch.set_default_device("cuda")


@torch.library.custom_op("vllm::fused_rms_norm_quant_static", mutates_args=['result'])
def fused_rms_norm_quant_static(result: torch.Tensor, input: torch.Tensor, weight: torch.Tensor, scale: torch.Tensor,
                                azp: torch.Tensor, epsilon: float) -> None:
    print("vllm::fused_rms_norm_quant_static")
    # bogus implementation doesn't matter
    result_rms = torch.mul(input, weight) + epsilon
    result = torch.mul(result_rms, scale).to(torch.int8)


@torch.library.custom_op("vllm::rms_norm", mutates_args=['result'])
def rms_norm(result: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
              epsilon: float) -> None:
    # bogus implementation doesn't matter
    result = torch.mul(input, weight) + epsilon


@torch.library.custom_op("vllm::static_scaled_int8_quant", mutates_args=['result'])
def static_scaled_int8_quant(result: torch.Tensor,
                              input: torch.Tensor,
                              scale: torch.Tensor,
                              azp: Optional[torch.Tensor] = None) -> None:
    # bogus implementation doesn't matter
    result = torch.mul(input, scale).to(torch.int8)


def rms_pattern_static(result: torch.Tensor, result_rms: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
                       scale: torch.Tensor):
    at1 = auto_functionalized(torch.ops.vllm.rms_norm.default, result=result_rms, input=input, weight=weight, epsilon=1e-6)
    at2 = auto_functionalized(torch.ops.vllm.static_scaled_int8_quant.default, result=result, input=at1[1], scale=scale,
                              azp=None)

    # result
    return at2[1]


def rms_replacement_static(result: torch.Tensor, result_rms: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
                           scale: torch.Tensor):
    at = auto_functionalized(torch.ops.vllm.fused_rms_norm_quant_static.default, result=result, input=input, weight=weight,
                             epsilon=1e-6, scale=scale, azp=None)

    # result
    return at[1]


def empty_bf16(*args, **kwargs):
    return torch.empty(*args, **kwargs, dtype=torch.bfloat16)


def empty_int8(*args, **kwargs):
    return torch.empty(*args, **kwargs, dtype=torch.int8)


my_patterns = PatternMatcherPass()
inputs = [empty_int8(5, 4), empty_bf16(5, 4), empty_bf16(5, 4), empty_bf16(5, 1), torch.empty(1, 1)]
register_replacement(rms_pattern_static, rms_replacement_static, inputs, fwd_only, my_patterns)

def custom_pass(graph: torch.fx.Graph) -> torch.fx.Graph:
    #print("Pre-pass:")
    #print(graph)
    count = my_patterns.apply(graph)
    print(f"Count: {count}")
    #print(graph)
    return graph

def custom_backend(graph: torch.fx.GraphModule, example_inputs: List[torch.Tensor]) -> Callable:
    from torch._inductor import config
    current_config = config.shallow_copy_dict()
    from torch._inductor.compile_fx import compile_fx
    current_config['post_grad_custom_post_pass'] = custom_pass
    return compile_fx(graph, example_inputs, config_patches=current_config)

@torch.compile(backend=custom_backend)
def my_func_static(x, w, epsilon):
    result = torch.empty_like(x)
    torch.ops.vllm.rms_norm(result, x, w, epsilon)
    scale = torch.ones((1, 1))
    quant_result = torch.empty_like(x, dtype=torch.int8)
    torch.ops.vllm.static_scaled_int8_quant(quant_result, result, scale, None)
    return quant_result


print("Run my_func_static")
inputs = [torch.empty((5, 4)), torch.empty((5, 1)), 1e-6]
my_func_static(*inputs)

cc @ezyang @chauhang @penguinwu @zou3519 @eellison @laithsakka @masnesral

Versions

Note: the pytorch git hash is 6b14e6c

Collecting environment information...
PyTorch version: 2.5.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.30.1
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.0-35-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.5.82
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA H100 80GB HBM3
GPU 1: NVIDIA H100 80GB HBM3
GPU 2: NVIDIA H100 80GB HBM3
GPU 3: NVIDIA H100 80GB HBM3
GPU 4: NVIDIA H100 80GB HBM3
GPU 5: NVIDIA H100 80GB HBM3
GPU 6: NVIDIA H100 80GB HBM3
GPU 7: NVIDIA H100 80GB HBM3

Nvidia driver version: 555.42.02
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8462Y+
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
Stepping: 8
CPU max MHz: 4100.0000
CPU min MHz: 800.0000
BogoMIPS: 5600.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 3 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 128 MiB (64 instances)
L3 cache: 120 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-31,64-95
NUMA node1 CPU(s): 32-63,96-127
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] flashinfer==0.0.9+cu121torch2.3
[pip3] mypy==1.9.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] onnx==1.14.1
[pip3] onnxruntime==1.18.1
[pip3] pytorch-triton==3.1.0+5fe38ffd73
[pip3] torch==2.5.0+cu124
[pip3] torchaudio==2.5.0.dev20240919+cu121
[pip3] torchvision==0.20.0.dev20240919+cu121
[pip3] triton==3.0.0
[conda] Could not collect

[eellison]

cc @zou3519 this is because of the (outdated) pattern match mutation check here. is there a new canonical way to check for mutation ?

[zou3519]

Do we have FakeTensors available? If so the way to check is to run a FakeTensor prop and check if the version counters got mutated (assuming we're not in inference mode, that's a bug to be fixed)

[eellison]

Ideally we would not have to rerun fake tensor prop (which can be expensive) to check if mutating. and the version counter this is a little imprecise because if you had

gm(x):
    y = x[0:2]
    y.add_(1)

the slice would also register as being mutating.

Didn't you a write a function for checking mutation of op in python as one point, or am i misremembering ?

[zou3519]

Didn't you a write a function for checking mutation of op in python as one point, or am i misremembering ?

The way I did that was run a function with real tensor inputs and then check the data for mutation