Repository navigation

GLM-5.2/GLM-5.3 PR 1: engine: MLA attention, fp8 fused MoE, symmetric all-reduce - #386

Merged

vasilevklart merged 36 commits into

mainfrom

glm/01-engine

Oct 20, 2026

Merged

GLM-5.2/GLM-5.3 PR 1: engine: MLA attention, fp8 fused MoE, symmetric all-reduce#386
vasilevklart merged 36 commits into
mainfrom
glm/01-engine

Commits on Oct 15, 2026
- engine: MLA, fp8 MoE, step budget, symm all-reduce
  vasilevklart
  committed
  a74c56a
- serving_chat: decode text across chunk boundaries
  vasilevklart
  committed
  fe290bf
- attn: MLA wrappers per slot, graph indices per row
  vasilevklart
  committed
  74eb6a5
- fused_moe: prefill experts as one custom op
  vasilevklart
  committed
  7978911
- micro_scheduler: step budget takes the oldest
  vasilevklart
  committed
  2db5679
- fused_moe: decode router takes at most 64 tokens
  vasilevklart
  committed
  49423a4
- attn: FlashMLA only for eager MLA plans
  vasilevklart
  committed
  421c683
- fused_moe: an empty step skips the align op
  vasilevklart
  committed
  94251af
- distributed: symm all-reduce on every rank or none
  vasilevklart
  committed
  42bd26a
- cuda_graph_runner: refuse empty/cudagraph modes
  vasilevklart
  committed
  8cf5b51
- fused_moe: check block scales and inputs
  vasilevklart
  committed
  3d0589f
- kv: SHM transfer reads an MLA cache by token
  vasilevklart
  committed
  930abb1
- conductor: pre-build the MoE op only when asked
  vasilevklart
  committed
  b608622
- attn: refuse what only the MLA backend reads
  vasilevklart
  committed
  4f60268
- test: CI runs the new CPU tests
  vasilevklart
  committed
  c3b3848
- engine: select TP async scheduling from config
  vasilevklart
  committed
  299c824
- api: warmup_requests can repeat their text
  vasilevklart
  committed
  15b10a5
- worker: tp_async_sched reads boolean words
  vasilevklart
  committed
  a707775
- cuda_graph_runner: empty compile mode env is unset
  vasilevklart
  committed
  392f2e8
- distributed: load timeout ends after warmup
  vasilevklart
  committed
  c12e701
Commits on Oct 17, 2026
- micro_scheduler: drop the step token budget
  vasilevklart
  committed
  126b097
- attn: refuse context_only on the backends that don't read it
  vasilevklart
  committed
  454e606
- attn: drop FlashMLA
  vasilevklart
  committed
  603ea45
Commits on Oct 18, 2026
- cuda_graph_runner: drop MSTAR_GRAPH_COMPILE_MODE
  vasilevklart
  committed
  a4a72f3
- distributed: tp_allreduce and dist_timeout_s from the config only
  vasilevklart
  committed
  7331b80
- distributed: all-reduce backends as classes
  vasilevklart
  committed
  bc5c827
- distributed: one multimem kernel on every rank
  vasilevklart
  committed
  d08c0c0
- docs: the MLA attention resource in adding_models
  vasilevklart
  committed
  e681819
- distributed: a class per symmetric-memory kernel
  vasilevklart
  committed
  7bb3b48
- Merge main into glm/01-engine
  vasilevklart
  committed
  71544d0
Commits on Oct 19, 2026
- Merge main into glm/01-engine
  vasilevklart
  committed
  402f3ff
- Merge main into glm/01-engine
  vasilevklart
  committed
  3fc40d8
Commits on Oct 20, 2026
- Merge main into glm/01-engine
  vasilevklart
  committed
  07d96e3
- distributed: compiled one-shot all-reduce stays on NCCL
  vasilevklart
  committed
  9e84091
- Merge main into glm/01-engine
  vasilevklart
  committed
  14c9677
- test: #745's run-loop worker stubs the timeout restore #386's warmup ends with
  vasilevklart
  committed
  957ceee

Commits

Commits on Oct 15, 2026

Commits on Oct 17, 2026

Commits on Oct 18, 2026

Commits on Oct 19, 2026

Commits on Oct 20, 2026