feat: 支持 GLM-4.7-Flash 推理 - #25

Merged

kuma-loong merged 48 commits into

CURRENTF:mainfrom

kuma-loong:codex/glm-4.7-flash

Aug 13, 2026

Commits on Aug 12, 2026

feat: vendor lightllm mla kernels

codex
authored and
kuma-loong
committed
e7a681d
refactor: tag attention payload views

codex
authored and
kuma-loong
committed
eb37251
feat: add mla cache storage

codex
authored and
kuma-loong
committed
7eea70c
feat: add mla attention provider

codex
authored and
kuma-loong
committed
a27a81f
feat: add mla prefill and decode

codex
authored and
kuma-loong
committed
1ff41eb
refactor: share packed moe experts

codex
authored and
kuma-loong
committed
4551cd3
feat: add glm4 moe lite model

codex
authored and
kuma-loong
committed
552a9d3
feat: add glm response template

codex
authored and
kuma-loong
committed
c947de4
perf: cache mla prefill validation

codex
authored and
kuma-loong
committed
ba43997
refactor: inject glm mla runtime

codex
authored and
kuma-loong
committed
1209d15
fix: correct mla memory accounting

codex
authored and
kuma-loong
committed
0fbb579
fix: strip terminal eos from parser

codex
authored and
kuma-loong
committed
e51f88f
fix: align parser stop boundaries

codex
authored and
kuma-loong
committed
f5e885a
docs: document glm4 support

codex
authored and
kuma-loong
committed
5436b37
docs: record glm tp validation

codex
authored and
kuma-loong
committed
34639aa
feat: expand glm runtime support

codex
authored and
kuma-loong
committed
a2213c8
perf: accelerate glm flash runtime

CURRENTF
authored and
kuma-loong
committed
5cb8595
perf: checkpoint glm flash runtime

CURRENTF
authored and
kuma-loong
committed
cae8e2b
perf: optimize glm tp2 ep2 runtime

CURRENTF
authored and
kuma-loong
committed
58dad01
fix: validate only real mla graph rows

CURRENTF
authored and
kuma-loong
committed
aa6e6a1
perf: optimize glm tp2 ep2 decode

CURRENTF
authored and
kuma-loong
committed
8cfadba
perf: fuse glm moe routing

CURRENTF
authored and
kuma-loong
committed
4ba9354
perf: accelerate minimax long context

CURRENTF
authored and
kuma-loong
committed
5639c65
perf: bind minimax runtime operators

CURRENTF
authored and
kuma-loong
committed
9b70000
refactor: align glm provider contracts
kuma-loong
committed
73f4189
test: remove glm validation artifacts
kuma-loong
committed
5645405
docs: generalize glm setup guidance
kuma-loong
committed
e253062
test: inject glm cpu router provider
kuma-loong
committed
cbb70ff

Commits on Aug 13, 2026

fix: allow glm graph without eager mode
kuma-loong
committed
c0103fc
perf: unify all-reduce providers
kuma-loong
committed
dbae852
refactor: delegate model runtime binding
kuma-loong
committed
c30d35f
refactor: remove obsolete all-reduce wrappers
kuma-loong
committed
89e9468
refactor: remove model-specific cache fallbacks
kuma-loong
committed
695a084
refactor: organize kernel integrations
kuma-loong
committed
fda46e5
refactor: simplify cuda graph runner
kuma-loong
committed
664bc69
refactor: share packed moe experts
kuma-loong
committed
6798c5b
refactor: simplify glm runtime
kuma-loong
committed
3906ec5
docs: simplify dependency installation
kuma-loong
committed
dffa920
feat: add selectable CUDA extras
kuma-loong
committed
a47e481
fix: constrain cutlass dsl version
kuma-loong
committed
f47925f
fix: make cuda extras mutually exclusive
kuma-loong
committed
dfec1d1
fix: validate sgl kernel binaries
kuma-loong
committed
593b3ef
feat: support sglang kernel 0.4.5
kuma-loong
committed
3839783
fix: bound cuda graph context buckets
kuma-loong
committed
01ed2aa
fix: correct sgl moe expert alignment
kuma-loong
committed
183ae5f
feat: enable glm providers on h20
kuma-loong
committed
ef45304
test: initialize graph runner counter
kuma-loong
committed
4d229fb
fix: restore cpu test collection
kuma-loong
committed
97a3051