Perplexity has open sourced Lily, the local inference engine behind Hybrid Compute in Perplexity Computer. It is a single-process runtime: a Rust layer loads the checkpoint and drives the generation loop, an OpenAI-compatible chat-completions API streams tokens, and hand-written Metal kernels execute the model. Neither PyTorch nor MLX sits in the execution path. Lily is deliberately narrow with one model, Qwen3.6-35B-A3B, on one hardware family and that narrowness is the performance argument.
Is it deployable? Yes. A standalone demo is public in the pplx-garden repository. A Rust and Metal inference server offering greedy text generation through a minimal OpenAI-compatible HTTP API. The 4-bit checkpoint is 19.4 GB, so an Apple silicon Mac with 32 GB or more of unified memory is the realistic floor; Perplexity’s shipping Hybrid Compute product lists macOS 15+, 24 GB minimum and 32 GB for best results.
Why specialize at all?
The default Mac stack is MLX plus MLX-LM, which already ships a Qwen implementation with grouped expert work, a fused recurrent Metal kernel, and GQA-aware attention. But its operations must stay reusable across architectures. Lily gives that up and puts model structure, execution plans, and kernel selection in one runtime.
Three workload shapes
Qwen3.6-35B-A3B stores 35B parameters and activates roughly 3B per token. A router scores 256 experts and picks eight, alongside one shared expert that sees every token. It also mixes 10 full-attention layers using grouped-query attention (16 query heads, two KV heads) with 30 Gated DeltaNet layers. That yields three patterns: uneven expert groups, attention over a growing KV cache, and a fixed-size recurrence.
Prefill: keep weights packed, keep routing on the GPU
The checkpoint uses groupwise affine 4-bit quantization, every group of 64 weights sharing a bfloat16 scale and bias, about 70 GB of bfloat16 weights compressed to 19.4 GB. Metal 4 tensor operations consume bfloat16, so weights must be reconstructed first. Lily does that one tile at a time inside the grouped GEMM, holding results in threadgroup memory and accumulating in FP32, so the expanded array never reaches unified memory. In Perplexity’s ablation that fusion raised end-to-end prefill 77.4% at a 512-token prompt.
Keeping the routing histogram, prefix scan, scatter and block map inside a single GPU command buffer added 89% at 512 tokens by removing CPU synchronization inside each MoE layer. Moving from 16-row to 32-row tiles with four simdgroups added 13.2% at 2K; a register-resident Gated DeltaNet scan added 5.6%. Expert GEMMs are roughly 90% of prefill time. Long prompts run in bounded chunks so temporary activations do not compete with weights and cache for memory.
Decode: minimize bytes moved per token
Batch-1 decode has almost no weight reuse, so bandwidth sets the ceiling. One recorded step launched 795 kernels forming 555 sequential stages; Lily records real dependencies in a concurrent Metal pass so independent kernels overlap. The selected token is written straight into the next step’s GPU-resident input slot, removing a per-token CPU round trip, and four kernel chains are fused to keep intermediates in registers.
Coalesced cache reads lifted key bandwidth from 33.8 to 47.9 GB/s and value bandwidth from 42.0 to 61.8 GB/s. GQA packing, four query heads sharing one threadgroup so each KV row loads once, improved decode 23.8% at 32K. A fixed-block attention layout at 32K and above improved decode 7.7% at 32K, 27.4% at 64K, and 40.2% at 128K.
Results
On one 40-core, 128 GB M5 Max at batch 1, loading identical 4-bit checkpoint bytes against MLX-LM’s fastest direct-generation path across ten lengths from 256 to 128K tokens, Lily averaged 4,156 prefill tokens/s versus 3,388 (1.23x) and 170.0 decode tokens/s versus 126.4 (1.35x). At a 4K prompt and 4K context it reached 5,749.9 and 186.6 tokens/s against 4,737.5 and 140.9, and was faster at every recorded point: 1.12–1.42x prefill, 1.31–1.37x decode. A teacher-forced check across 192 positions put Lily’s perplexity 0.04% higher, with the same top-ranked token 96.35% of the time.
Key Takeaways
- Lily is a Rust + Metal engine for Qwen3.6-35B-A3B on Apple silicon, with no PyTorch or MLX in the path.
- Averages 1.23x MLX-LM prefill and 1.35x decode on a 40-core, 128 GB M5 Max.
- Biggest prefill wins: GPU-resident expert routing (+89%) and dequantization fused into the grouped GEMM (+77.4%).
- Biggest decode wins: GQA packing (+23.8% at 32K) and fixed-block attention (+40.2% at 128K).
Check out the Technical details here and the GitHub Repo here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
原始发布方：MarkTechPost（RSS）
原文时间：2026-09-03 14:57:07 +08:00
