[megatron] GLM-5.3-Flash (glm5_next) support + 4-layer GPU CI entry - #2156

Closed

erictang000 wants to merge 8 commits into

NovaSky-AI:mainfrom

erictang000:glm5.3-flash-megatron

Commits on Sep 4, 2026

[megatron] GLM-5.3-Flash (glm5_next) support: KDA + DSA(kpool) + mHC layers, bridge, 4-layer GPU CI entry

erictang000
and
claude
committed
290fafd
[deps] align transformers 5.16.1 / tokenizers pins with #2155; note vLLM dev flashinfer pin

erictang000
and
claude
committed
dd0c771
[test] cap vLLM max_num_seqs for GLM-5.3-Flash (KDA triton grid-dim overflow at 1024 seqs)

erictang000
and
claude
committed
70240f6
[inference] follow vLLM 0.28.1 move of CLI args to entrypoints.launchers (with fallback)

erictang000
and
claude
committed
a7116bc
[deps] DeepGEMM (torch 2.11 / CUDA 13 build) from a GitHub release for vLLM's GLM-5.3 DSA indexer

erictang000
and
claude
committed
1e12b3b
[weight_sync] accept vLLM's native sharded_rdt engine registration (vLLM >= 0.28.1)

erictang000
and
claude
committed
4a58f5e
[inference] enable vLLM scale-out endpoints (/inference/v1/generate is gated in vLLM >= 0.28.1)

erictang000
and
claude
committed
39143b9
[test] GLM-5.3-Flash-4layer megatron threshold 1e-1 (measured bf16 noise floor of the truncated slice)

erictang000
and
claude
committed
98d33be