#5899 [Engine] Offer: XeMaya Intel SYCL model paths and optimizations for upstream integration
open · @Jumbomuffin777 · 0 comments · View on GitHub
BenchmarksMulti-GPUAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindowsLinux
Description
Hi Niko — I developed [XeMaya](https://github.com/Jumbomuffin777/XeMaya) to run models natively on Intel Arc where the available SYCL paths were insufficient. I'd rather contribute the useful implementations and R&D directly to Strata than maintain a competing engine indefinitely. Would selective integration or a technical handoff be useful? The offer covers native SYCL/Level Zero DeepSeek-V4.1-Flash/Coder and GLM-5.3-Flash paths; multi-GPU execution; expert residency/caching across VRAM, RAM and SSD; chunked prefill; and Intel XMX attention and draft optimizations. Strata already has its own Intel work; this is an offer to identify useful differences, not replace it. This is separate from #5544's CUDA GLM proposal. Recorded evidence (Linux, oneAPI icpx 2026.2; Arc Pro B70/B65, plus B60 where specified; **no benchmarks rerun for this issue**): - **Qwen3.8-Flash-Next IQ3_S, B70+B65:** [matched 26.8K decode comparison](https://github.com/Jumbomuffin777/XeMaya/blob/8fa4141666e10d81bf932ab4625aa52a37eed4bd/xemaya/benchmarks/2026-11-11-qwen-ab/README.md), same model/prompts, greedy, `--spec 4`, 256-token requests: XeMaya **79.4 code / 59.9 prose** versus Strata **v0.2.12's 78.7 / 59.2 tok/s**. Small differences, about 0.5 tok/s recorded variability; not a comparison against current Strata. Newer XeMaya prefill was **1,864–1,940 tok/s**; Strata's corresponding report records 2,276–2,293. In the separate [earlier window-4 comparison](https://github.com/Jumbomuffin777/XeMaya/blob/8fa4141666e10d81bf932ab4625aa52a37eed4bd/xemaya/benchmarks/2026-11-09-window4-qwen/README.md), Strata led prefill **2,211 versus 1,842–1,884** with XeMaya's default MMQ. These are distinct runs/settings. - **DeepSeek regular Flash IQ2_XXS (384 experts), B70+B65, Gen-5 SSD:** [prefill before/after](https://github.com/Jumbomuffin777/XeMaya/blob/8fa4141666e10d81bf932ab4625aa52a37eed4bd/xemaya/benchmarks/2026-11-11-ds41-flash-prefill/README.md) **20.0→36.3 / 36.5→71.6 / 35.2→91.8 tok/s** at 576/2,048/8,192 tokens across the two days' changes. Expert retention alone reduced 8K drive reads **804→218 GB**. Regular and Coder each matched **16/16** forced reference tokens in the latest checks; this is limited reference coverage. - **GLM Maya-S v2, three GPUs:** [XMX comparison](https://github.com/Jumbomuffin777/XeMaya/blob/8fa4141666e10d81bf932ab4625aa52a37eed4bd/xemaya/STATUS.md) at 2,377 tokens improved prefill **94.8→152.1–161.3**, decode **13.97→17.36 tok/s** (**20.03 with MTP**, a separate setting). Latest beta reference agreement is **125/128**, versus earlier **126/128**; CPU/GPU expert rounding changes near-tie decisions. No exact-output or production-readiness claim. Multi-GPU scaling depends on topology: the older regular-Flash run decoded at 14.56 tok/s on two GPUs versus 11.58 on three. Qwen's development results do not qualify the published beta binary; Windows and broad cross-backend coverage remain unverified. These are workload-specific records, not general speed claims. Credit belongs to you and Strata's contributors, MistVVK and the XeStrata contributors for the Intel foundation, mw00/Project Maya for model work and quantizations, and ggml/llama.cpp for references. XeMaya as a whole is LGPL-3.0-or-later; Strata is MIT. Any extraction needs file-level provenance/license review: inherited LGPL code cannot simply be relabeled MIT. For my own contributions, we can discuss suitable licensing where ownership permits, or transfer findings for an independent implementation. If interested, which small component would you prefer to evaluate first? I'm offering focused collaboration and knowledge transfer directly into Strata, not asking you to adopt or maintain the fork.
Related on this site
Quick links to related pages.