All pull requests

Pull requests

• perf: fuse scale reduction into per-tensor FP8 quantization

  #1501·baifengyang1-cloud opened Oct 12, 2026·

  ·
  1

• feat(cache): align remote cache with KVCM and PACE

  #1499·GrayXu opened Oct 12, 2026·

  ·

• build(cuda13): resolve CUDA native deps from bazel pip repos and drop flashinfer-jit-cache

  #1498·LLLLKKKK opened Oct 12, 2026Collaborator·3 tasks·

  ·

• feat: add cache-grid benchmark pipeline

  #1497·HaiyuZhou opened Oct 12, 2026·

  ·
  17

• 解决兼容问题

  #1496·sunmiaozju opened Oct 11, 2026Collaborator·

  ·

• Revert "feat(flexlb): add KVCM and multi-engine runtime support"

  #1495·sunmiaozju opened Oct 11, 2026Collaborator·

  ·

• feat: replace native NVFP4 mm and CUDA XQA with rtp-kernel

  #1493·moui0 opened Oct 10, 2026Collaborator·1 of 2 tasks·

  ·
  11

• feat: enable external ATOM models in RTP-LLM and add Qwen3.8 ROCm smoke coverage

  #1492·junna2016 opened Oct 10, 2026Collaborator·

  ·
  14

• feat: add PromptGenerator and input embeddings support

  #1491·parkerpang opened Oct 10, 2026Collaborator·

  ·
  149

• feat(cache): support RemoteCache GDR on block tree cache

  #1490·fengfeisha opened Oct 10, 2026Collaborator·

  ·
  39

• refactor(flexlb): remove residual scheduler APIs and redundant state

  #1489·jianglan89 opened Oct 10, 2026Collaborator·

  ·
  207

• feat(deepseek): add V4.1 Flash CED and DSpARK support on main

  #1488·netaddi opened Oct 10, 2026Collaborator·

  ·
  184

• FlexLB: add Mock evaluation, ZK HA and cache retirement regression

  #1483·wzy-99 opened Oct 9, 2026Collaborator·

  ·
  168

• feat(ascend): add Huawei Ascend NPU support

  #1481·ningweikang opened Oct 9, 2026·

  ·
  35

• optimize bert inference

  #1480·parkerpang opened Oct 9, 2026Collaborator·

  ·
  143

• fix(loader): match 're:' ignore patterns against the weight template, not layer 0

  #1477·MincongZhou opened Oct 8, 2026·

  ·
  36

• feat(cache): add CRC32C protection to host and disk KV cache

  #1476·alibaba-miji opened Oct 8, 2026Collaborator·

  ·
  42

• feat: support dynamic FP8 KV cache for Qwen3 Eagle3

  #1471·Zn0425 opened Sep 29, 2026·2 of 5 tasks·

  ·

• feat(qwen35): support community Qwen3.5 checkpoints and decision models

  #1470·LLLLKKKK opened Sep 28, 2026Collaborator·

  ·
  101

• feat: add DFlash2 draft support on CUDA and ROCm

  #1469·yzyDavid opened Sep 28, 2026Collaborator·

  ·

• feat: add opt-in chunked TP MoE prefill overlap

  #1466·yzyDavid opened Sep 28, 2026Collaborator·

  ·

• refactor: share engine lifecycle for sleep/wake and bounded graceful shutdown

  #1463·MMadhatter opened Sep 25, 2026Collaborator·

  ·
  30

• fix: enforce FP8 KV startup metadata contracts

  #1461·Zn0425 opened Sep 24, 2026·2 of 3 tasks·

  ·
  6

• feat: add chunked prefill for standard, MLA, and hybrid attention

  #1460·Jonathan-hwx opened Sep 23, 2026Collaborator·

  ·
  60

• fix: avoid ATREX buffer limit for large ROCm KV caches

  #1456·zhangjianning-zjn opened Sep 23, 2026Collaborator·

  ·
  3