[torch.compile] rework compile control with piecewise cudagraph

[youkaichao]

rework the compilation control.

the user-facing flags are:

Usage	CompilationLevel	how vLLM uses Dynamo	how vLLM uses Inductor	use vLLM's custom ops (*)	how to customize the compilation
export VLLM_TORCH_COMPILE_LEVEL=0 (default)	NO_COMPILATION (0)	N/A	N/A	✅	N/A
export VLLM_TORCH_COMPILE_LEVEL=1	DYNAMO_AS_IS (1)	use as-is	N/A	✅	vllm.plugins.set_torch_compile_backend (default to "eager")
export VLLM_TORCH_COMPILE_LEVEL=2	DYNAMO_ONCE (2)	use only once, make sure computation graph does not change	N/A	✅	vllm.plugins.set_torch_compile_backend (default to "eager")
export VLLM_TORCH_COMPILE_LEVEL=3	PIECEWISE (3)	same as 2	compilation behavior determined by the config	❌	write a json config file and specify it with VLLM_TORCH_COMPILE_CONFIG, or call vllm.plugins.set_compilation_config to set the config directly

* : users can also use VLLM_CUSTOM_OPS env var to have fine-grained control over custom ops.

For the detailed compilation config, please check the code doc.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[youkaichao]

example code:

import torch
from vllm.compilation.decorators import support_torch_compile
from vllm.compilation.compile_context import set_compile_context
from vllm.plugins import set_attention_ops
set_attention_ops(["silly.attention"])

import torch
from torch import nn

@torch.library.custom_op("silly::attention", mutates_args=["out"])
def silly_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, out: torch.Tensor) -> None:
    print("silly")
    out.copy_(q)
    print(q)
    out[0] += 1

@silly_attention.register_fake
def _(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, out: torch.Tensor) -> None:
    return

@support_torch_compile
class SillyModel(nn.Module):
    def __init__(self) -> None:
        super().__init__()

    def forward(
        self,
        x: torch.Tensor
    ) -> torch.Tensor:
        x = x + 1
        x = x + 2
        out = torch.empty_like(x)
        torch.ops.silly.attention(x, x, x, out)
        x = out
        x = x - 2
        x = x - 1
        out = torch.empty_like(x)
        torch.ops.silly.attention(x, x, x, out)
        x = out
        x = x + 1
        return x

model = SillyModel()

input_buffer = torch.randn(100).cuda()

with set_compile_context([1, 2]):
    model(input_buffer)

    model(input_buffer[:2])
    model(input_buffer[:1])

input_buffer[:2].zero_()
output = model(input_buffer[:2])
print(output.__class__)
print(output[:2])

run with:

VLLM_LOGGING_LEVEL=DEBUG VLLM_TORCH_COMPILE_LEVEL=3 python test.py

requirements:

• attention ops will be the boundary of piecewise graphs
• if the output of attention ops is used in the subsequent graph, then it needs to be allocated in the previous graph, and passed to attention ops for mutation.

# FIXME: it seems pytorch changes the output to a tuple
it can be fixed by pytorch/pytorch#138980

[youkaichao]

I added some tests based on counters, following pytorch's test principle. @ProExpertProg

[youkaichao]

test failure is unrelated and also appear in main branch, merging