--- title: vLLM v0.25.0: transformers backend now matches native... description: vLLM v0.25.0 ships with a major performance breakthrough: the Hugging Face transformers library backend now matches or exceeds native vLLM model throughput.... canonical: https://daily.dev/posts/vllm-v0-25-0-transformers-backend-now-matches-native-vllm-speed-z8kvnsk7c twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: vLLM v0.25.0: transformers backend now matches native vLLM speed | daily.dev og:description: vLLM v0.25.0 ships with a major performance breakthrough: the Hugging Face transformers library backend now matches or exceeds native vLLM model throughput.... og:url: https://daily.dev/posts/vllm-v0-25-0-transformers-backend-now-matches-native-vllm-speed-z8kvnsk7c og:image: https://api.daily.dev/og/posts/Z8kvNsk7C.png og:image:alt: vLLM v0.25.0: transformers backend now matches native vLLM speed og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # vLLM v0.25.0: transformers backend now matches native vLLM speed **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary vLLM v0.25.0 ships with a major performance breakthrough: the Hugging Face transformers library backend now matches or exceeds native vLLM model throughput. This is achieved via torch.fx static graph analysis and AST manipulation to apply inference-specific layer fusions (fused MoE expert parallelism kernels, merged parallel linear layers) at runtime. Benchmarks on Qwen3 4B, 32B, and 235B-A22B-FP8 MoE models confirm parity or better performance. The practical benefit is that the same model code can now be used across training, evals, RL rollouts, and production inference without rewriting for vLLM, and new models are available on release day. Other changes include Model Runner V2 becoming the default for dense models, a Unified Streaming Parser Engine, universal speculative decoding across heterogeneous vocabularies, and new model support. ## Content ## What changed vLLM v0.25.0 shipped with 558 commits from 232 contributors, and the headline feature is something that's been a quiet pain point for a while: the Hugging Face Transformers modeling backend now runs at native vLLM speed, often matching or beating hand-written model implementations. Until now, getting a new model architecture into production inference meant writing it twice. Once in Transformers for training and research, then again in vLLM with custom inference code. That duplication slowed down new model releases, added maintenance overhead, and created opportunities for the two implementations to drift apart. ## How it works The backend uses `torch.fx` for static graph analysis and AST manipulation to dynamically apply inference-specific optimizations at runtime. That includes fused MoE expert parallelism kernels and merged parallel linear layers - without requiring model authors to write any vLLM-specific code. Benchmarks across Qwen3 4B, 32B, and 235B-A22B-FP8 MoE models show parity or better throughput versus native vLLM implementations. It supports tensor parallelism, data parallelism, expert parallelism, `torch.compile`, and CUDA Graphs. You enable it with a single flag: `--model-impl transformers`. ## Other v0.25.0 highlights - **Model Runner V2** is now the default for all dense models; the legacy PagedAttention implementation is retired - **Unified Streaming Parser Engine** for more consistent output handling - **Universal speculative decoding** across heterogeneous vocabularies (TLI), plus new DSpark and DFlash drafters - New model support including Hunyuan Hy3 and Unlimited OCR ## Why this matters The practical upside is that the same Transformers model code can now cover training, fine-tuning, evaluation, RL rollouts, and production inference. New architectures are available in vLLM as soon as they land in Transformers, with no extra porting work. For the open-source ecosystem, that's a meaningful reduction in the lag between a model release and it being usable at production speed. ## Similar posts on daily.dev - [Hybrid Models as First-Class Citizens in vLLM – PyTorch](https://daily.dev/posts/hybrid-models-as-first-class-citizens-in-vllm-pytorch-9wfhak9r1) · PyTorch · 0 upvotes · 0 comments --- Tags: [#transformers](https://daily.dev/tags/transformers), [#ai-inference](https://daily.dev/tags/ai-inference), [#vllm](https://daily.dev/tags/vllm) [View this post on daily.dev](https://daily.dev/posts/vllm-v0-25-0-transformers-backend-now-matches-native-vllm-speed-z8kvnsk7c) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"vLLM v0.25.0: transformers backend now matches native vLLM speed","url":"https://daily.dev/posts/vllm-v0-25-0-transformers-backend-now-matches-native-vllm-speed-z8kvnsk7c","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/vllm-v0-25-0-transformers-backend-now-matches-native-vllm-speed-z8kvnsk7c"},"datePublished":"2026-07-12T11:33:16.901Z","dateModified":"2026-07-18T06:44:07.408Z","description":"vLLM v0.25.0 ships with a major performance breakthrough: the Hugging Face transformers library backend now matches or exceeds native vLLM model throughput....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/654e21f5a49cdde7e690802ace949d16?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/654e21f5a49cdde7e690802ace949d16?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/vllm-v0-25-0-transformers-backend-now-matches-native-vllm-speed-z8kvnsk7c","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"transformers,ai-inference,vllm","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"vLLM v0.25.0: transformers backend now matches native vLLM speed"}]} ```