vLLM v0.27.0 landed on August 10, and it's a big one: 561 commits from 242 contributors, 64 of them new. The headline for me is that hybrid-attention models — architectures that mix sliding-window attention (SWA) with full attention layers — can now run through vLLM's offline inference path, something that was previously restricted.

Hybrid Attention Finally Reaches Offline Inference

The change is PR #47699: hybrid (SWA + full) attention models in offline inference. Until now, if you were running a model with modular attention — say, a Qwen3-style architecture with SWA layers for long context plus full-attention layers — you were effectively locked into the online serving path. Offline batch inference, the path you use for evaluation, dataset generation, and bulk scoring, didn't support the hybrid attention backends.

That gap mattered more than it sounds. Offline inference is where you validate a model before you commit to serving it, and where you run the long, non-interactive workloads. Having hybrid models work in one path but not the other meant you couldn't trust offline results to predict online behavior, and you couldn't use the offline path for the exact architectures that are becoming the default for long-context work.

The release notes also call out that "attention backends stay eligible for text-only serving of prefix-LM models" (#48796), which is the same theme: the attention machinery is being unified so modular-attention architectures don't fall off a cliff depending on which path you pick.

What Else Landed in v0.27.0

Kimi K3 gets a full-stack landing in one release: core model files and kernels (#50089, #50000), Python (#50093) and Rust (#50104) frontends, AttnRes kernels (#50090), DeepGEMM support (#50458), compressed-tensors quantized checkpoints (#50500), and an option to shard the shared expert instead of replicating it (#50656). Qwen3.5 text-only dense and MoE models are in (#50210), plus K-EXAONE-2.0-750B-A37B (#50524) and VaultGemma via the Transformers modeling backend (#49803).

The DeepSeek-V4 performance push has real numbers attached: sequence parallelism (#46789), a ~2x kernel improvement by skipping empty c128 launches (#48957), 3.4% E2E TTFT from skipping an unneeded topk/router (#49486), 3.9% E2E TTFT from workspace reuse (#49236), a 1.88x kernel from removing a redundant full kernel (#50298), and 448 MiB of GPU memory saved in the PP buffer (#50312).

Model Runner V2 expands beyond generative workloads: encoder-only attention (#49331), sequence pooling for embedding and classification (#48791), encoder token classification (#50293) and token embedding (#50574), BGE-M3 pooling (#50661), and multimodal on CPU (#50073).

The Breaking Change to Plan Around

PyTorch 2.13.0 is now the baseline, along with torchvision 0.28.0 and Triton 3.7.1 (#48155). The release notes are explicit that this is a breaking environment change — XPU (#48677) and CPU (#50412) followed to torch 2.13 as well. If you're pinning vLLM in a container or a requirements file, this is the thing that will bite you on upgrade, not the new models.

There's also early next-gen hardware enablement: an sm_107 target for NVIDIA Rubin (#49387) with NVLink all-reduce paths on SM107 (#49647), and ROCm gfx1250 (#46516). And a new JIT warmup infrastructure (#47451) with runner-owned Triton kernel warmup (#49903) that removes first-request compilation stalls — that's the fix for the "first token takes forever" complaint that shows up every time you restart a vLLM server.

What I'd Watch

The hybrid-attention offline support is the one I'd actually go verify. If you're evaluating a long-context model before putting it in production, the fact that offline and online now share the same attention backends means your offline benchmarks finally reflect what you'll get in serving. That's worth a re-run of your eval suite on v0.27.0 even if you're not touching the new models.

The catch: it's a breaking environment change, so budget for the torch 2.13 migration before you adopt. And as always with vLLM, the release notes are the source of truth — the PR numbers above are all in there.

Sources: