K3 model line: decode E2E at ep_size 4 with MegaMoE(situ) transport - #867

Merged

xiaguan merged 17 commits into

mainfrom

feat/k3-bringup

Aug 14, 2026

Commits on Aug 14, 2026

feat(kernels): bump DeepGEMM to the openinfer branch (situ activation)

xiaguan
and
claude
committed
05fb7df
feat(kernels): k3 feature — FP8xFP4 masked grouped GEMM + TileLang codegen

xiaguan
and
claude
committed
77730bb
feat(k3): model crate — config, weight loader, model line, scheduler skeleton

xiaguan
and
claude
committed
e3d228d
feat(kernels): k3 MoE decode chain (local dispatch, quant, situ, combine)

xiaguan
and
claude
committed
ad39994
feat(kernels): k3 batched TileLang surface (13 families, gemv retired)

xiaguan
and
claude
committed
cb9ff1f
feat(k3): device model build (loader output -> kernel-ready rank model)

xiaguan
and
claude
committed
a395006
feat(k3): batched decode executor — single-rank decode E2E

xiaguan
and
claude
committed
52810f3
feat(k3): free-running EP4 — fixed-chain MoE oracle, bitwise vs single-rank

xiaguan
and
claude
committed
a55cda8
chore(kernels): rustfmt drift in build.rs and k3_tilelang.rs

xiaguan
and
claude
committed
a420e6a
docs(k3): record full-depth EP4 decode E2E (free-running serve verified)

xiaguan
and
claude
committed
09cdc39
feat(k3): MegaMoE(situ) AOT integration at ep_size 1 — bit-parity gated

xiaguan
and
claude
committed
764e46b
feat(k3): MegaMoE at ep_size 4 — zero-collective EP, bitwise vs single-rank

xiaguan
and
claude
committed
5cec6f4
refactor(k3): retire the stepwise EP scaffolding — mega is the MoE transport

xiaguan
and
claude
committed
615bd6e
docs(benchmarks): K3 EP4 decode step profile — backbone GEMM is the bottleneck

xiaguan
and
claude
committed
082e85d
style(k3): fix clippy gates — let-else in build.rs gencode, then_some in admission
xiaguan
committed
5da1677
refactor(k3): prune the dead kernel surface and stale scaffolding comments
xiaguan
committed
3a35cc0
ci: replace actions-dco with a direct commit-list sign-off check
xiaguan
committed
71c88db

Navigation Menu

Uh oh!

K3 model line: decode E2E at ep_size 4 with MegaMoE(situ) transport - #867

K3 model line: decode E2E at ep_size 4 with MegaMoE(situ) transport#867
xiaguan merged 17 commits into
mainfrom
feat/k3-bringup

Commits on Aug 14, 2026

feat(kernels): bump DeepGEMM to the openinfer branch (situ activation)

feat(kernels): k3 feature — FP8xFP4 masked grouped GEMM + TileLang codegen

feat(k3): model crate — config, weight loader, model line, scheduler skeleton

feat(kernels): k3 MoE decode chain (local dispatch, quant, situ, combine)

feat(kernels): k3 batched TileLang surface (13 families, gemv retired)

feat(k3): device model build (loader output -> kernel-ready rank model)

feat(k3): batched decode executor — single-rank decode E2E

feat(k3): free-running EP4 — fixed-chain MoE oracle, bitwise vs single-rank

chore(kernels): rustfmt drift in build.rs and k3_tilelang.rs

docs(k3): record full-depth EP4 decode E2E (free-running serve verified)

feat(k3): MegaMoE(situ) AOT integration at ep_size 1 — bit-parity gated

feat(k3): MegaMoE at ep_size 4 — zero-collective EP, bitwise vs single-rank

refactor(k3): retire the stepwise EP scaffolding — mega is the MoE transport

docs(benchmarks): K3 EP4 decode step profile — backbone GEMM is the bottleneck

style(k3): fix clippy gates — let-else in build.rs gencode, then_some in admission

refactor(k3): prune the dead kernel surface and stale scaffolding comments

ci: replace actions-dco with a direct commit-list sign-off check