moe-expert-streaming / transformers.patch

Commit History

5 days ago

  1. RFC: device_map form implemented, measured warm start; transformers.patch at PR 1a1652cb5b

    tarekziade HF Staff
    Verified signature
    65d3c91
  2. RFC rewritten for the current state (maintainer directions, hooks); transformers.patch at PR b7477fbb00

    tarekziade HF Staff
    Verified signature
    c376385
  3. Update every page to 2026-11-17: details (results, weights, kernels, streaming, cache policy, known issues), RFC, README, results index, transformers.patch at 18abece3ab

    tarekziade HF Staff
    Verified signature
    c5cc338

6 days ago

  1. Upload folder using huggingface_hub

    tarekziade HF Staff
    Verified signature
    dc81d25

8 days ago

  1. Decode 10.0 tok/s: shared-expert overlap, early MPS commits, fused hyper-connection projections; dense ablation

    tarekziade HF Staff
    Verified signature
    e7bfe80
  2. Decode on four different prompts (cold cache), cache-policy replay

    tarekziade HF Staff
    Verified signature
    c7ebb65
  3. Qwen3.8: 8.7 tok/s decode (fused Gated DeltaNet decode step), current prefill

    tarekziade HF Staff
    Verified signature
    0501224
  4. Qwen3.8 only: results, the two Macs, architecture and kernels

    tarekziade HF Staff
    Verified signature
    a0de066

9 days ago

  1. Final sweep: page cache bypassed vs on; prefill and copy optimizations; kernel bug root cause

    tarekziade HF Staff
    Verified signature
    6072e5e
  2. Add vanilla quantized baselines (MetalConfig, preloaded quantized experts) and BF16 with Hub kernels

    tarekziade HF Staff
    Verified signature
    28b20c0
  3. MoE expert streaming results: OLMoE on M5 (BF16, 8-bit, 4-bit) and earlier cloud runs

    tarekziade HF Staff
    Verified signature
    8fb0b86