--- title: How to Utilize OKF Efficiently to Enable Knowledge... description: A pipeline built around Google's Open Knowledge Format (Markdown + YAML frontmatter) is repurposed to let three Qwen2.5-Coder models (7B, 3B, 1.5B) hand off... canonical: https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How to Utilize OKF Efficiently to Enable Knowledge Exchange Among LLMs | daily.dev og:description: A pipeline built around Google's Open Knowledge Format (Markdown + YAML frontmatter) is repurposed to let three Qwen2.5-Coder models (7B, 3B, 1.5B) hand off... og:url: https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv og:image: https://api.daily.dev/og/posts/8iwLrnbgv.png og:image:alt: How to Utilize OKF Efficiently to Enable Knowledge Exchange Among LLMs 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. # How to Utilize OKF Efficiently to Enable Knowledge Exchange Among LLMs **[Towards Data Science](https://daily.dev/sources/tds)** · 21 min read · 1 upvotes · 0 comments ## Summary A pipeline built around Google's Open Knowledge Format (Markdown + YAML frontmatter) is repurposed to let three Qwen2.5-Coder models (7B, 3B, 1.5B) hand off pre-tokenized integer arrays instead of each re-running BPE tokenization independently. Since the whole Qwen2.5-Coder family shares one identical vocabulary, an upstream agent tokenizes once and writes the token array to a RAM-backed shared memory path, referenced via a custom token_pointer field in the frontmatter. Benchmarks show a 28% TTFT reduction on the 3B model and 37.8% on the 1.5B model. The key engineering contribution is a runtime guardrail that performs a full ~151,936-entry vocabulary dictionary equality check (not just vocab_size) plus a special_tokens_map comparison before trusting cross-agent token arrays, since a mismatch would silently produce fluent but wrong output rather than an error. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Can different sized models from the same family, like Qwen2.5-Coder 7B, 3B, and 1.5B, share a KV cache to speed up multi-agent pipelines? No, KV caches cannot be shared across differently sized models in the same family because the KV cache shape is derived from hidden_size, which differs per model (3584 for the 7B, 2048 for the 3B, 1536 for the 1.5B). What can be shared instead is the pre-computed token ID array, since the entire Qwen2.5-Coder family uses one identical BPE vocabulary. _Developers designing multi-agent LLM pipelines can track architecture trade-offs like this on daily.dev._ ### How much can skipping tokenization improve time-to-first-token when passing pre-tokenized arrays between Qwen2.5-Coder models? Passing pre-tokenized int64 arrays instead of re-tokenizing reduced mean TTFT by 28.0% on Qwen2.5-Coder-3B-Instruct (69.3ms to 49.9ms) and by 37.8% on Qwen2.5-Coder-1.5B-Instruct (49.6ms to 30.9ms), measured over a median of 7 trials with greedy decoding and 64 new tokens. The smaller model shows a bigger percentage gain because its GPU forward-pass floor is lower, making the fixed tokenizer cost a larger fraction of its total time. _Engineers optimizing multi-agent LLM latency can compare techniques like this on daily.dev._ ### What safety check is needed before feeding one model's tokenized integer array directly into a different model's generate function? A full dictionary equality check on get_vocab() covering the entire vocabulary (roughly 151,936 entries for Qwen2.5-Coder), not just a vocab_size comparison, plus a special_tokens_map comparison for EOS and other special token consistency. A mismatch does not crash generation; it silently produces fluent but semantically wrong output because the model embeds integers under the wrong subword mapping. _Developers wiring token IDs across models can weigh correctness safeguards like this on daily.dev._ ## Similar posts on daily.dev - [qMLX: Maximising my AI psychosis by minmaxing my Mac Studio](https://daily.dev/posts/qmlx-maximising-my-ai-psychosis-by-minmaxing-my-mac-studio-gigqbqete) · Hacker News · 1 upvotes · 0 comments - [The Infrastructure Behind Making Local LLM Agents Actually Useful](https://daily.dev/posts/the-infrastructure-behind-making-local-llm-agents-actually-useful-ff4qnxir1) · Towards Data Science · 1 upvotes · 0 comments - [How Semantic Code Navigation Cuts Agent Token Costs by up to 36%](https://daily.dev/posts/how-semantic-code-navigation-cuts-agent-token-costs-by-up-to-36--tlu1vjjqp) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#nlp](https://daily.dev/tags/nlp), [#pytorch](https://daily.dev/tags/pytorch), [#qwen](https://daily.dev/tags/qwen) [View this post on daily.dev](https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv) ```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":"How to Utilize OKF Efficiently to Enable Knowledge Exchange Among LLMs","url":"https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv"},"datePublished":"2026-08-13T16:44:43.677Z","dateModified":"2026-09-14T07:18:23.774Z","description":"A pipeline built around Google's Open Knowledge Format (Markdown + YAML frontmatter) is repurposed to let three Qwen2.5-Coder models (7B, 3B, 1.5B) hand off...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/86b998a5982b75f167fe822edfab136a?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/86b998a5982b75f167fe822edfab136a?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Towards Data Science","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":"Towards Data Science","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tds","url":"https://daily.dev/sources/tds"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,nlp,pytorch,qwen","timeRequired":"PT21M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Towards Data Science","item":"https://daily.dev/sources/tds"},{"@type":"ListItem","position":3,"name":"How to Utilize OKF Efficiently to Enable Knowledge Exchange Among LLMs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms-8iwlrnbgv#faq","mainEntity":[{"@type":"Question","name":"Can different sized models from the same family, like Qwen2.5-Coder 7B, 3B, and 1.5B, share a KV cache to speed up multi-agent pipelines?","acceptedAnswer":{"@type":"Answer","text":"No, KV caches cannot be shared across differently sized models in the same family because the KV cache shape is derived from hidden_size, which differs per model (3584 for the 7B, 2048 for the 3B, 1536 for the 1.5B). What can be shared instead is the pre-computed token ID array, since the entire Qwen2.5-Coder family uses one identical BPE vocabulary. Developers designing multi-agent LLM pipelines can track architecture trade-offs like this on daily.dev."}},{"@type":"Question","name":"How much can skipping tokenization improve time-to-first-token when passing pre-tokenized arrays between Qwen2.5-Coder models?","acceptedAnswer":{"@type":"Answer","text":"Passing pre-tokenized int64 arrays instead of re-tokenizing reduced mean TTFT by 28.0% on Qwen2.5-Coder-3B-Instruct (69.3ms to 49.9ms) and by 37.8% on Qwen2.5-Coder-1.5B-Instruct (49.6ms to 30.9ms), measured over a median of 7 trials with greedy decoding and 64 new tokens. The smaller model shows a bigger percentage gain because its GPU forward-pass floor is lower, making the fixed tokenizer cost a larger fraction of its total time. Engineers optimizing multi-agent LLM latency can compare techniques like this on daily.dev."}},{"@type":"Question","name":"What safety check is needed before feeding one model's tokenized integer array directly into a different model's generate function?","acceptedAnswer":{"@type":"Answer","text":"A full dictionary equality check on get_vocab() covering the entire vocabulary (roughly 151,936 entries for Qwen2.5-Coder), not just a vocab_size comparison, plus a special_tokens_map comparison for EOS and other special token consistency. A mismatch does not crash generation; it silently produces fluent but semantically wrong output because the model embeds integers under the wrong subword mapping. Developers wiring token IDs across models can weigh correctness safeguards like this on daily.dev."}}]} ```