All pull requests

Pull requests

• fix(dynamicemb): keep each table's local_hbm_for_values budget in fused groups

  #534·tiankongdeguiji opened Nov 3, 2026Contributor·

  ·
  2

• feat(dynamicemb): hash_bucket dist_type with a fan-out-invariant global bucket layout

  #533·tiankongdeguiji opened Nov 3, 2026Contributor·

  ·
  13

• feat(hstu): FP8 DeepGEMM and FA4 attention for Transformer inference

  #531·z52527 opened Nov 2, 2026Collaborator·3 tasks done·

  ·
  5

• fix(dynamicemb): keep prefetched keys' ref counts through a rehash

  #529·shijieliu opened Oct 30, 2026Collaborator·

  ·
  1

• [FEA] DynamicEmb: hierarchical dist for row-wise sequence sharding

  #528·shijieliu opened Oct 30, 2026Collaborator·

  ·

• fix(dynamicemb): send frequency weights for an empty input under index dedup

  #527·shijieliu opened Oct 29, 2026Collaborator·3 tasks done·

  ·
  1

• fix(dynamicemb): wait for queued GPU work before freeing VMMTensor memory

  #526·shijieliu opened Oct 29, 2026Collaborator·3 tasks done·

  ·
  1

• fix(dynamicemb): deduplicate int32 keys under index dedup

  #525·shijieliu opened Oct 29, 2026Collaborator·3 tasks done·

  ·
  1

• fix(dynamicemb): bucketize int32 keys by the int64 key a table stores

  #524·shijieliu opened Oct 29, 2026Collaborator·3 tasks done·

  ·
  1

• perf: fuse qk-norm + RoPE and make it the default on SM10x (up to 1.7x)

  #517·bellanich opened Oct 28, 2026Contributor·3 tasks done·

  ·
  8

• perf: vectorize the beam-KV write and make it the default (up to 3.7x)

  #513·bellanich opened Oct 28, 2026Contributor·3 tasks done·

  ·
  3

• [WIP][FEA](DynamicEmb) Light Weight Cache

  #506·jiashuy opened Oct 23, 2026Collaborator·3 tasks·

  ·
  7

• feat(dynamicemb): fp8 table storage with stochastic rounding for row-wise Adagrad

  #499·jyuan-sc opened Oct 9, 2026Contributor·

  ·
  3

• Fea transformer hstu inference 456

  #491·Junyi-Zheng opened Sep 20, 2026·3 tasks done·

  ·
  11

  

• [FEA] Support configurable indexing and NVE layers for exportable embedding collections

  #478·geoffreyQiu opened Sep 4, 2026Collaborator·3 tasks·

  ·
  12

• [FEAT] add HDFS support for DynamicEmb checkpoint dump and load

  #423·PWZER opened Jun 16, 2026·1 of 3 tasks·

  ·
  7

• selection_kernel_fusion in issue357

  #413·ShaobinChen-AH opened Jun 1, 2026Contributor·3 tasks·

  ·
  1

  5

• [Draft]Fake mode or CPU support

  #410·jiashuy opened May 29, 2026Collaborator·3 tasks·

  ·
  3

• Port output distribution classes to DynamicEmb

  #297·z52527 opened Feb 6, 2026Collaborator·3 tasks·

  ·
  6

• Add grouped linear layer with strided BMM optimization

  #263·z52527 opened Jan 6, 2026Collaborator·

  ·
  6

  

• Update benchmark results of dynamicemb on H200

  #261·jiashuy opened Dec 30, 2025Collaborator·3 tasks·

  ·
  2

  

• FP8 Enablement for TE layers + HSTU attn

  #197·esoba opened Oct 21, 2025Contributor·3 tasks·

  ·
  11