Add option to use unbacked, and backed size obl dynamic shapes for more sounds compilation.

[laithsakka]

Purpose

Dynamic shapes guards are dropped unsoundly in vLLM, those can be added during both dynamo and inductor compilations. The proper way to compile a sounds graph is to use unbacked dynamic shapes or ""backed size oblivious
maybe"" ! (see the note in the diff content). This PR adds a config that allow choosing between backed, unbacked and backed_size_oblivious as explained in the comment in the content of the PR.

When using unbacked, users might want to to provide invariants about the shapes of the Model, a lambda argument
is added to support_torch_compile decorators where users can provide a lambda that assert on invariants
on the model inputs shapes. Those are needed to avoid DDE and to be able to trace the model with unbacked.
example:

def llama_model_invariants(input_ids,
                           positions,
                           intermediate_tensors=None,
                           inputs_embeds=None):
    """Shape invariants for Llama model compilation"""
    if input_ids is not None:
        torch._check(positions.size()[0] == input_ids.size()[0])
        
@support_torch_compile(shape_invariants=llama_model_invariants)
..

see the qwen example in the code also.

Test Plan

Added a unit test .
only works for torch. 2.10+

Perf testing

command

Qwen/Qwen2-1.5B-Instruct

will look into some of those perf issues in the future, but the long term vision with pre_compile is that this will be
a fallback mode. results anyway still better than eager.

CUDA_VISIBLE_DEVICES=1 vllm bench throughput --model Qwen/Qwen2-1.5B-Instruct --input-len 512 --output-len 128 --num-prompts 100 --gpu-memory-utilization 0.8

backed

Throughput: 88.94 requests/s, 56738.83 total tokens/s, 11383.87 output tokens/s

backed size oblivious

88.78

unbacked

Throughput: 82.98  51860.18 total tokens/s, 10405.04 output tokens/s

eager

Throughput: 63.29 requests/s, 40376.23 total tokens/s, 8100.94 output tokens/s

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @laithsakka.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gemini-code-assist]

Code Review

This pull request introduces a significant and valuable refactoring of the torch.compile integration, particularly around handling dynamic shapes. The new TorchCompileGuardsStripWrapper and the introduction of shape invariants provide a much cleaner and more robust approach to compilation. The changes are well-structured and the new tests are a great addition.

I've found a few issues that need to be addressed: a critical bug in the new DynamicShapesConfig that would cause a runtime error, a minor bug in hash computation, and an opportunity to strengthen the new dynamic shapes test.

[hmellor]

The conflicts are caused by our migration to ruff. Please see https://vllm-dev.slack.com/archives/C07R5Q1Q2BB/p1759663228844749 which contains detailed instructions to make updating your branch as painless as possible.

[laithsakka]

@ProExpertProg
why removed ready i addressed all comments this morning?
I kind of need the ready also to know if things break.

[ProExpertProg]

Sorry I thought you were still pushing, but also the CI ran already anyway

[laithsakka]

just addressed richard comment on the debug section. @zou3519 looking good?

[zou3519]

this is a bit of a repeat of the previous section. Let's resolve it in a future PR, figuring out what to write is a bit annoying

[laithsakka]

ah i see

## Debugging Dynamic Shape full graph capture

we probably should merge the two i agree

[laithsakka]

@ProExpertProg the failing job is a timeout in building docs is there a way to retry?

[ProExpertProg]

@hmellor

[laithsakka]

dummy update to force running all tests/

[hmellor]

Tiny nit to get the help text to render nicely. Should bhe done in a follow up to save CI

Suggested change

	- BACKED: Default PyTorch behavior with potential guards ignored.
	- UNBACKED: No guards guaranteed (most sound) but may throw
	data dependent errors.
	- BACKED_SIZE_OBLIVIOUS: Experimental safer alternative to
	backed/unbacked.
	- BACKED: Default PyTorch behavior with potential guards ignored.\n
	- UNBACKED: No guards guaranteed (most sound) but may throw
	data dependent errors.\n
	- BACKED_SIZE_OBLIVIOUS: Experimental safer alternative to
	backed/unbacked.

[zou3519]

@laithsakka

[laithsakka]

on it.

[hmellor]

Docs failure/timeout was likely related to the Python docs (which our docs references) being down. It appears to have resolved itself