All pull requests

Pull requests

• fix(download): eliminate continuous HF polling for unrequested models and add auth backoff

  #2481·presentyourlove opened Oct 24, 2026·

  ·
  1

• fix(download): prevent download screen dismissal and fix failed download retry lockout

  #2480·presentyourlove opened Oct 24, 2026·

  ·
  1

• fix(master): validate per-device memory in tensor parallel placement to prevent OOM crash loops

  #2479·presentyourlove opened Oct 24, 2026·

  ·
  1

• fix(worker): prevent placement deadlock during runner init by flushing channels and bounding timeouts

  #2478·presentyourlove opened Oct 24, 2026·

  ·
  1

• feat(logging): per-module log filters and default to DEBUG log levels

  #2477·presentyourlove opened Oct 24, 2026·

  ·

• feat(info_gatherer): validate RDMA connections with ibv_devinfo

  #2476·presentyourlove opened Oct 24, 2026·

  ·

• feat(download): fast model loads by streaming model from peer cluster devices

  #2475·presentyourlove opened Oct 24, 2026·

  ·

• feat(routing): dynamically switch to higher priority connection with InstanceReplacedAtomically

  #2474·presentyourlove opened Oct 23, 2026·

  ·

• feat(runner): implement chunked prefill to eliminate decode blocking during continuous batching

  #2473·presentyourlove opened Oct 23, 2026·

  ·

• feat(placement): memory pressure awareness instead of memory used

  #2472·presentyourlove opened Oct 23, 2026·

  ·

• feat(telemetry): profiled network latency, capacity, and real-time bandwidth usage in UI

  #2471·presentyourlove opened Oct 23, 2026·

  ·

• feat(viz): show hardware connection types (TB5, Ethernet, WiFi, RDMA) in UI

  #2470·presentyourlove opened Oct 23, 2026·

  ·

• feat(download): allow direct offline loading of complete model folders without internet

  #2469·presentyourlove opened Oct 23, 2026·

  ·

• fix(networking): restore and normalize EXO_BOOTSTRAP_PEERS connection resolution

  #2468·presentyourlove opened Oct 23, 2026·

  ·

• feat(cpu_jit): add CPUID-aware vectorized JIT and cache topology tiling engine

  #2467·presentyourlove opened Oct 23, 2026·

  ·

• feat(quantization): add builtin one-click model quantize CLI pipeline

  #2466·presentyourlove opened Oct 23, 2026·

  ·

• feat(npu): add unified NPU HAL and dynamic subgraph auto-partitioner

  #2465·presentyourlove opened Oct 23, 2026·

  ·

• feat(moe): add hierarchical expert offloading and early-layer speculative prefetcher

  #2463·presentyourlove opened Oct 23, 2026·

  ·

• feat(kv-cache): add distributed content-addressed RadixAttention tree for prompt deduplication

  #2462·presentyourlove opened Oct 23, 2026·

  ·

• feat(kv-cache): add distributed PagedAttention engine and speculative double-buffering prefetcher

  #2461·presentyourlove opened Oct 23, 2026·

  ·

• feat(quantization): add 1.0-bit BNN codec and vectorized XNOR-Popcount GEMM engine

  #2460·presentyourlove opened Oct 23, 2026·

  ·

• feat(npu): add on-chip SRAM KV cache manager and native INT2/ternary LUT GEMM engine

  #2459·presentyourlove opened Oct 23, 2026·

  ·

• feat(streaming): add AirLLM-style sliding-window layer streamer with double-buffered ring

  #2458·presentyourlove opened Oct 23, 2026·

  ·

• feat(moe): add elastic expert fallback executor with weight redistribution and residual bypass

  #2457·presentyourlove opened Oct 23, 2026·

  ·

• feat(routing): add device-aware adaptive MoE router with queue and thermal compensation

  #2456·presentyourlove opened Oct 23, 2026·

  ·