# vLLM v0.7.2 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.7.2
- **Released:** Feb 06, 2025
- **Impact Score:** 5/10
- **Breaking Changes:** Not verified
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.7.2/

## Executive Summary (TL;DR)
vLLM v0.7.2. Source highlight: Qwen2.5-VL is now supported in vLLM. Please note that it requires a source installation from Hugging Face `transformers` library at the moment (#12604)

## New Features
- Add `transformers` backend support via `--model-impl=transformers`. This allows vLLM to be ran with arbitrary Hugging Face text models (#11330, #12785, #12727).
- Add SPDX-License-Identifier headers to python source files (#12628)
- [Hardware][Intel GPU] add XPU bf16 support
- [Misc] Add SPDX-License-Identifier headers to python source files
- [Model]: Add `transformers` backend support
- [Model] Add Deepseek V3 fp8_w8a8 configs for B200

## Improvements
- Performance enhancement to DeepSeek models.
- Align KV caches entries to start 256 byte boundaries, yielding 43% throughput enhancement (#12676)
- Apply `torch.compile` to fused_moe/grouped_topk, yielding 5% throughput enhancement (#12637)
- Improve hash collision avoidance in prefix caching (#12621)
- [Core] Improve hash collision avoidance in prefix caching
- [Perf] Mem align KV caches for CUDA devices (MLA perf improvement)

## Bug Fixes
- [Bugfix] fix moe_wna16 get_quant_method
- Fix for attention layers to remain unquantized during moe_wn16 quant
- [ci/build] fix gh200 test
- [Misc] Fix improper placement of SPDX header in scripts
- [Bugfix][Kernel] Fix per-token/per-channel quantization for Hopper scaled mm
- [Quant] Fix use_mla TypeError and support loading pure-sparsity Compressed Tensors configs

## Upgrade Command
```bash
python -m pip install 'vllm==0.7.2'
```