--- format: "aidr-story-markdown/v1" id: "b8e2d01df43a03984d3af2635ee11eaee4073bd6695b87dbd3a0188102240120" canonical_url: "https://aidr.today/b8e2d01d?lang=en" title: "HeyGen x Google Cloud: Bringing Avatar IV to TPUs" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-09-15T19:46:17.000Z" category: "Infra" topics: [] source_urls: ["https://developers.googleblog.com/heygen-x-google-cloud-bringing-avatar-iv-to-tpus/"] summary: "HeyGen ported their 18B+ parameter Avatar IV video generation model to Google Cloud's Trillium (v6e) TPUs via torchax and XLA, utilizing FSDP and Ulysses sequence parallelism across an eight-chip mesh. To achieve a 1.86x speedup for real-time streaming, the engineering team pipelined exposed all-to-all collectives, aligned sparse attention block sizes to eliminate mask padding, and bypassed softmax serial dependencies using a precomputed Cauchy-Schwarz upper bound. These custom Pallas kernel and compiler optimizations were deployed only after passing rigorous two-tier quality gates to guarantee byte-identical or mathematically equivalent pixel outputs." --- # HeyGen x Google Cloud: Bringing Avatar IV to TPUs > [Open the canonical story]() **Published:** 2026-09-15T19:46:17.000Z **Category:** Infra ## Summary HeyGen ported their 18B\+ parameter Avatar IV video generation model to Google Cloud's Trillium \(v6e\) TPUs via torchax and XLA, utilizing FSDP and Ulysses sequence parallelism across an eight\-chip mesh\. To achieve a 1\.86x speedup for real\-time streaming, the engineering team pipelined exposed all\-to\-all collectives, aligned sparse attention block sizes to eliminate mask padding, and bypassed softmax serial dependencies using a precomputed Cauchy\-Schwarz upper bound\. These custom Pallas kernel and compiler optimizations were deployed only after passing rigorous two\-tier quality gates to guarantee byte\-identical or mathematically equivalent pixel outputs\. ## Sources - [Story source]()