TECHNICAL NOTES

Notes from the project

Setup details, failures, workarounds, and measurements after their sources and wording have been reviewed.

4× Intel Arc Pro B70 server chassis running DFlash2 speculative decoding
ARCHITECTURE

DFlash & DFlash2: Block Diffusion LLM Speedup

Local LLM inference rarely bottlenecks on matrix compute—it chokes on memory bandwidth, capping 70B models below 16 tokens/s on 608 GB/s GDDR6. DFlash2 eliminates this sequential memory wall through parallel block diffusion, tripling decode throughput on Intel Arc and modern workstations with zero mathematical precision loss across vLLM and SGLang.

Schematic illustration of the B65 specification topics: GDDR6 memory, Xe2 compute and PCI Express.
ARCHITECTURE

Intel Arc Pro B65: specifications in context

Intel lists 32 GB of GDDR6 and 608 GB/s for the Arc Pro B65. This overview separates the GPU specification from board details and measured inference performance.

Schematic illustration of backend selection, function imports and model startup.
TROUBLESHOOTING

SGLang v0.5.15.post1: DSA launch and GLM-5.2 fixes

The SGLang v0.5.15.post1 patch covers DSA model startup, CUDA image dependencies and GLM-5.2 fixes. A closer look at the corrected CuTe DSL import.

Schematic illustration of decode graph replay, piecewise prefill and opt-in XPU graphs.
ARCHITECTURE

SGLang v0.5.15: XPU graphs and speculative decoding

SGLang v0.5.15 brings XPU graph and speculative-decoding work. The upstream changes also document a graph-capture limitation in the Intel test backend.

Schematic illustration of the release topics: SYCL mRoPE, INT8 checkpoints and active BF16 states.
ARCHITECTURE

SGLang v0.5.14: XPU and prefix-cache changes

A look at SGLang v0.5.14: the Intel XPU mRoPE path, PyTorch dependency update, and optional INT8 checkpoints for linear-attention prefix reuse.