[Bugfix] Fix structured output crash on CPU due to pin_memory=True

[wjhrdy]

Essential Checks

• PR title follows the pattern [Tag] Short description
• I have searched for related issues and checked existing PRs
• I have run linting/formatting locally

Purpose

Fix RuntimeError: pin_memory=True requires a CUDA or other accelerator backend crash when using structured output (guided decoding) on CPU-only deployments.

Fixes #37705

Problem

apply_grammar_bitmask() in vllm/v1/structured_output/utils.py crashes on CPU when handling mixed batches (concurrent structured + non-structured requests):

1. pin_memory=True is hardcoded — torch.tensor(out_indices, ..., pin_memory=True) requires CUDA; fails on CPU-only systems.
2. xgrammar CPU kernel expects Sequence[int], not torch.Tensor — apply_token_bitmask_inplace_cpu() only accepts a Python list for the indices argument.

Note: the existing CPU float32 workaround (added in #31901) was never reachable because the pin_memory=True crash occurs first.

Fix

On CPU, pass out_indices as a plain Python list directly instead of converting to a pinned tensor. The GPU path with pinned memory is preserved.

Test Plan

Tested by starting vLLM on CPU with ibm-granite/granite-3.2-2b-instruct, then sending concurrent plain + structured output (response_format: json_schema) requests. Without the fix, both requests return 500 and the EngineCore dies. With the fix, both succeed and the server stays healthy.

import concurrent.futures
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
MODEL = "ibm-granite/granite-3.2-2b-instruct"

def plain_request():
    return client.chat.completions.create(
        model=MODEL,
        messages=[{"role": "user", "content": "Tell me a story"}],
        max_tokens=200,
    )

def structured_request():
    return client.chat.completions.create(
        model=MODEL,
        messages=[{"role": "user", "content": "What is the capital of France?"}],
        max_tokens=50,
        response_format={
            "type": "json_schema",
            "json_schema": {
                "name": "resp", "strict": True,
                "schema": {
                    "type": "object",
                    "properties": {"capital": {"type": "string"}},
                    "required": ["capital"],
                    "additionalProperties": False,
                },
            },
        },
    )

with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
    f1 = executor.submit(plain_request)
    f2 = executor.submit(structured_request)
    print(f1.result())
    print(f2.result())

[gemini-code-assist]

Code Review

This pull request effectively resolves a critical RuntimeError that occurred on CPU-only deployments due to pin_memory=True being hardcoded for torch.tensor creation. The changes correctly introduce conditional logic to handle CPU and GPU devices separately, ensuring that pin_memory=True is only applied when a CUDA device is available. Furthermore, it addresses the xgrammar CPU kernel's expectation of a Python list for indices by passing out_indices directly on CPU, which is a significant improvement for correctness and stability in mixed-batch scenarios. The updated type hint for indices also enhances code clarity.

[gemini-code-assist]

This conditional logic is a critical fix. By checking logits.device.type, the code now correctly avoids setting pin_memory=True on CPU, which was causing a RuntimeError. Additionally, passing out_indices as a plain Python list for CPU devices directly addresses the xgrammar CPU kernel's expectation for a Sequence[int], preventing potential issues with type mismatches.

[dougbtv]

Looks excellent -- do we need any validation on the testing side?

[mgoin]

Seems reasonable to me, thanks!

[mergify]

Hi @wjhrdy, the pre-commit checks have failed. Please run:

uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-files

Then, commit the changes and push to your branch.

For future commits, pre-commit will run automatically on changed files before each commit.

Tip

Is mypy failing?

mypy is run differently in CI. If the failure is related to this check, please use the following command to run it locally:

# For mypy (substitute "3.10" with the failing version if needed)
pre-commit run --hook-stage manual mypy-3.10

[andy-neuma]

thanks