[None][feat] Add GLM-5.3-Flash (glm5_next) support - #19136

Merged

juney-nvidia merged 35 commits into

NVIDIA:mainfrom

ruocheng-nv:feat/glm5.3-flash-full-support

Sep 28, 2026

Commits on Sep 27, 2026

[None][feat] Add GLM-5.3-Flash support
ruocheng-nv
committed
6c2bd06
[None][fix] Validate GLM FP8 KV cache and sparse forward contracts
ruocheng-nv
committed
439f801
[None][feat] Enable GLM-5.3-Flash attention DP with MTP and FP8 KV cache
ruocheng-nv
committed
9033427
[None][docs] Clarify GLM KV cache auto dtype behavior
ruocheng-nv
committed
bab6f06
[None][fix] Honor all stop tokens in greedy sampling
ruocheng-nv
committed
55d9550
[None][test] Include greedy stop-token regression in GPU CI
ruocheng-nv
committed
86fcd93
[None][refactor] Simplify GLM-5.3-Flash runtime and sparse backend
ruocheng-nv
committed
bc4f9d7
[None][refactor] Trim GLM-5.3-Flash model helpers and comments
ruocheng-nv
committed
6e44815
[None][docs] Pin GLM-5.3-Flash deployment to Transformers 5.17.0
ruocheng-nv
committed
80a2ef9
[None][refactor] Consolidate GLM sparse attention and KDA runtime setup
ruocheng-nv
committed
6795727
[None][perf] Enable compact KDA decode for 64 heads at small batches
ruocheng-nv
committed
34a7706
[None][docs] Simplify GLM-5.3-Flash supported-model note
ruocheng-nv
committed
971d447
[None][fix] Honor GLM indexer head-weight strides
ruocheng-nv
committed
e18a345
[None][docs] Trim GLM-5.3-Flash benchmark commentary
ruocheng-nv
committed
d765474
[TRTLLM-16480][perf] Extend mHC pre-mapping tuning coverage
ruocheng-nv
committed
a02fbb6
[TRTLLM-16480][perf] Optimize GLM-5.3-Flash small-batch decoding
ruocheng-nv
committed
f048260
[TRTLLM-16018][fix] Fix GLM k-pool offsets and TopK fallback
ruocheng-nv
committed
5dea8f3
[None][refactor] Reuse FP32 linear for GLM router
ruocheng-nv
committed
901fc65
[None][refactor] Align GLM sparse attention with shared interfaces
ruocheng-nv
committed
2a02428
[None][refactor] Query GLM raw cache slots through the public layer API
ruocheng-nv
committed
76d5471
[None][fix] Preserve GLM 64-head dispatch in the updated KDA backend
ruocheng-nv
committed
5892029
[None][fix] Declare GLM transceiver preference and simplify attribute access
ruocheng-nv
committed
eaf9e32
[None][refactor] Validate GLM sparse metadata and forward argument types
ruocheng-nv
committed
41b5db8
[None][test] Add GLM-5.3-Flash coverage to GB300 CI
ruocheng-nv
committed
71a6860
[None][fix] Separate GLM image and video preprocessing settings
ruocheng-nv
committed
24fe42a
[None][test] Reduce GLM-5.3-Flash end-to-end CI coverage
ruocheng-nv
committed
2ce8304
[None][doc] Document GLM-5.3-Flash long-input benchmark results
ruocheng-nv
committed
6bbd17c
[None][doc] Restore GLM-5.3-Flash performance guide
ruocheng-nv
committed
55ce509
[None][fix] Account for GLM indexer cache memory
ruocheng-nv
committed
c1140b4
[None][refactor] Address GLM-5.3-Flash review follow-ups
ruocheng-nv
committed
de9f05c
[None][chore] Trim GLM KV cache manager routing comments
ruocheng-nv
committed
f6a796e
[None][test] Skip MoE A2A CFT regression under CUDA forward compatibility
ruocheng-nv
committed
70017b8
[None][test] Gate MoE A2A CFT skip on a newer user-mode driver
ruocheng-nv
committed
4fa804d
[None][fix] Gate MoE A2A CFT on the NVIDIA kernel driver version

zhangcl
authored and
ruocheng-nv
committed
2d17aae
[None][doc] Update GLM-5.3-Flash 1K/1K performance chart
ruocheng-nv
committed
1ec4cfe

Navigation Menu

[None][feat] Add GLM-5.3-Flash (glm5_next) support - #19136

[None][feat] Add GLM-5.3-Flash (glm5_next) support#19136
juney-nvidia merged 35 commits into
NVIDIA:mainfrom
ruocheng-nv:feat/glm5.3-flash-full-support

Commits on Sep 27, 2026

[None][feat] Add GLM-5.3-Flash support

[None][fix] Validate GLM FP8 KV cache and sparse forward contracts

[None][feat] Enable GLM-5.3-Flash attention DP with MTP and FP8 KV cache

[None][docs] Clarify GLM KV cache auto dtype behavior

[None][fix] Honor all stop tokens in greedy sampling

[None][test] Include greedy stop-token regression in GPU CI

[None][refactor] Simplify GLM-5.3-Flash runtime and sparse backend

[None][refactor] Trim GLM-5.3-Flash model helpers and comments

[None][docs] Pin GLM-5.3-Flash deployment to Transformers 5.17.0

[None][refactor] Consolidate GLM sparse attention and KDA runtime setup

[None][perf] Enable compact KDA decode for 64 heads at small batches

[None][docs] Simplify GLM-5.3-Flash supported-model note

[None][fix] Honor GLM indexer head-weight strides

[None][docs] Trim GLM-5.3-Flash benchmark commentary

[TRTLLM-16480][perf] Extend mHC pre-mapping tuning coverage

[TRTLLM-16480][perf] Optimize GLM-5.3-Flash small-batch decoding

[TRTLLM-16018][fix] Fix GLM k-pool offsets and TopK fallback

[None][refactor] Reuse FP32 linear for GLM router

[None][refactor] Align GLM sparse attention with shared interfaces

[None][refactor] Query GLM raw cache slots through the public layer API

[None][fix] Preserve GLM 64-head dispatch in the updated KDA backend

[None][fix] Declare GLM transceiver preference and simplify attribute access

[None][refactor] Validate GLM sparse metadata and forward argument types

[None][test] Add GLM-5.3-Flash coverage to GB300 CI

[None][fix] Separate GLM image and video preprocessing settings

[None][test] Reduce GLM-5.3-Flash end-to-end CI coverage

[None][doc] Document GLM-5.3-Flash long-input benchmark results

[None][doc] Restore GLM-5.3-Flash performance guide

[None][fix] Account for GLM indexer cache memory

[None][refactor] Address GLM-5.3-Flash review follow-ups

[None][chore] Trim GLM KV cache manager routing comments

[None][test] Skip MoE A2A CFT regression under CUDA forward compatibility

[None][test] Gate MoE A2A CFT skip on a newer user-mode driver

[None][fix] Gate MoE A2A CFT on the NVIDIA kernel driver version

[None][doc] Update GLM-5.3-Flash 1K/1K performance chart