--- title: IsoExec: Unified Execution to Eliminate... description: IsoExec is a cross-framework execution abstraction built for SkyRL that eliminates numerical mismatch between rollout (vLLM) and training (Megatron) engines in... canonical: https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL | daily.dev og:description: IsoExec is a cross-framework execution abstraction built for SkyRL that eliminates numerical mismatch between rollout (vLLM) and training (Megatron) engines in... og:url: https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn og:image: https://api.daily.dev/og/posts/4scDY1Ovn.png og:image:alt: IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL 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. # IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL **[vLLM](https://daily.dev/sources/vllm)** · 9 min read · 0 upvotes · 0 comments ## Summary IsoExec is a cross-framework execution abstraction built for SkyRL that eliminates numerical mismatch between rollout (vLLM) and training (Megatron) engines in RL workloads. It combines an execution contract enforcing identical rounding-sensitive kernel choices with a unified, batch-invariant, bitwise-consistent model, plus a new chunkwise-parallel recurrent (CPR) algorithm for Gated DeltaNet architectures. On an 8×H100 node training Qwen3.5-35B-A3B with DAPO, it reduced average rollout-versus-training logprob difference below 1e-6 at a 25.3% end-to-end step-time overhead versus the native SkyRL stack, with no clear reward gain observed over a short 50-step run. The implementation is open-sourced. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is IsoExec and how does it fix trainer-inference mismatch in RL training with vLLM and Megatron? IsoExec is a cross-framework unified execution abstraction implemented in SkyRL that pairs an execution contract specifying rounding-sensitive kernel details with a unified model using bitwise-consistent kernels across vLLM (rollout) and Megatron (training). On an 8xH100 node training Qwen3.5-35B-A3B with DAPO, it cut the average rollout-versus-training logprob difference below 1e-6, at a 25% end-to-end throughput overhead versus the native SkyRL stack. _Teams debugging RL training instability from engine mismatch can track infrastructure fixes like this via daily.dev._ ### How does chunkwise-parallel recurrent (CPR) GDN avoid the slowdown of using recurrent-only computation for Gated DeltaNet training-inference parity? CPR keeps the recurrence as the main computation but evaluates it in parallel across chunks, computing recurrent state at chunk boundaries then scanning outputs within chunks, and resynchronizing hidden state periodically during decode. This achieves bitwise-exact parity like a fully recurrent approach while running at roughly 1.4-1.7x overhead per layer, versus the 2-5x slowdowns reported when using pure recurrent computation for training and prefill. _Developers optimizing linear-attention RL pipelines can follow performance techniques like this on daily.dev._ ### What train-inference mismatch effects has Fireworks reported for GLM-5.2 RL training runs? Fireworks reported a GLM-5.2 run with train-inference KL divergence around 0.013 where clipping discarded roughly 45% of tokens and reward collapsed around step 20, whereas a bitwise-aligned run had zero clipped tokens and remained stable. This illustrates how unaddressed numerical mismatch between rollout and training engines can destabilize reinforcement learning training. _Engineers tuning RL training stability can follow real-world incident data like this on daily.dev._ ## Similar posts on daily.dev - [Building an RL Theorem-Proving Workflow on Modal](https://daily.dev/posts/building-an-rl-theorem-proving-workflow-on-modal-qareohy4r) · Modal · 0 upvotes · 0 comments - [Systems Engineering Playbook: Optimizing Qwen 3.5-397B MoE on Ironwood \(TPU7x\)](https://daily.dev/posts/systems-engineering-playbook-optimizing-qwen-3-5-397b-moe-on-ironwood-tpu7x--quc6gr0fd) · Google Developers · 1 upvotes · 0 comments --- Tags: [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#vllm](https://daily.dev/tags/vllm) [View this post on daily.dev](https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn) ```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":"IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL","url":"https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn"},"datePublished":"2026-08-21T13:54:06.350Z","dateModified":"2026-09-14T08:07:14.129Z","description":"IsoExec is a cross-framework execution abstraction built for SkyRL that eliminates numerical mismatch between rollout (vLLM) and training (Megatron) engines in...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4a6d5c34de771dbadd96e35854e6ee97?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4a6d5c34de771dbadd96e35854e6ee97?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"vLLM","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":"vLLM","logo":"https://media.daily.dev/image/upload/s--hTxEuls9--/f_auto/v1744613054/logos/vllm","url":"https://daily.dev/sources/vllm"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"reinforcement-learning,vllm","timeRequired":"PT9M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"vLLM","item":"https://daily.dev/sources/vllm"},{"@type":"ListItem","position":3,"name":"IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/isoexec-unified-execution-to-eliminate-trainer-inference-mismatch-in-skyrl-4scdy1ovn#faq","mainEntity":[{"@type":"Question","name":"What is IsoExec and how does it fix trainer-inference mismatch in RL training with vLLM and Megatron?","acceptedAnswer":{"@type":"Answer","text":"IsoExec is a cross-framework unified execution abstraction implemented in SkyRL that pairs an execution contract specifying rounding-sensitive kernel details with a unified model using bitwise-consistent kernels across vLLM (rollout) and Megatron (training). On an 8xH100 node training Qwen3.5-35B-A3B with DAPO, it cut the average rollout-versus-training logprob difference below 1e-6, at a 25% end-to-end throughput overhead versus the native SkyRL stack. Teams debugging RL training instability from engine mismatch can track infrastructure fixes like this via daily.dev."}},{"@type":"Question","name":"How does chunkwise-parallel recurrent (CPR) GDN avoid the slowdown of using recurrent-only computation for Gated DeltaNet training-inference parity?","acceptedAnswer":{"@type":"Answer","text":"CPR keeps the recurrence as the main computation but evaluates it in parallel across chunks, computing recurrent state at chunk boundaries then scanning outputs within chunks, and resynchronizing hidden state periodically during decode. This achieves bitwise-exact parity like a fully recurrent approach while running at roughly 1.4-1.7x overhead per layer, versus the 2-5x slowdowns reported when using pure recurrent computation for training and prefill. Developers optimizing linear-attention RL pipelines can follow performance techniques like this on daily.dev."}},{"@type":"Question","name":"What train-inference mismatch effects has Fireworks reported for GLM-5.2 RL training runs?","acceptedAnswer":{"@type":"Answer","text":"Fireworks reported a GLM-5.2 run with train-inference KL divergence around 0.013 where clipping discarded roughly 45% of tokens and reward collapsed around step 20, whereas a bitwise-aligned run had zero clipped tokens and remained stable. This illustrates how unaddressed numerical mismatch between rollout and training engines can destabilize reinforcement learning training. Engineers tuning RL training stability can follow real-world incident data like this on daily.dev."}}]} ```