GLM-5 DSA indexer fusion on ROCm (diff vs upstream main) - #1

Draft

Jacob0226 wants to merge 311 commits into

base/upstream-main-glm5-indexerfrom

jacob/glm5-dsa-indexer-fusion-unified

This pull request is big! We're only showing the most recent 250 commits

Commits on Aug 14, 2026

[Docs] Add GB300 cells and benchmarks for Qwen3.8-27B (#34863)

authored
70e291b
Fix mamba checkpoint depth under dcp (#34808)
ispobock
authored
c20acee
Retain SWA down to the last state checkpoint (#34729)
ispobock
authored
7562e74
fix(qwen3): support DeepEP-class backends and early EPLB state (#34810)
UNIDY2002
authored
d8399af
[Fix] Fix Qwen3.5 MTP startup with HiCache (#34560)

authored
41cd5a7
[SM12x] Default the fused MHC post+pre path on (#34019)

zhendonghua
and
claude
authored
1af761a
Fix startup weight load after TorchAO removal (#34869)
mmangkad
authored
d7207be
fix(muse-glimmer): parse required/named tool calls natively (#34781)

sid-rp
and
JustinTong0323
authored
42e8718
docs: sync LMSYS SGLang blog cards (#32982)
sglang-bot
authored
b676793
[Docs] Fill GLM-5.2 H200 FP8 speed cells (low-latency, balanced); fix MTP notation (#31554)

Smallfu666
and
claude
authored
22dde1d
perf: add H200 Triton MoE configs for E256 N512 (#34150)

authored
03c1d58
Add Reasoning-Aware Compression (RAC) pruning recipe for reasoning models (#32414)

authored
bfb224f
fix(dsa): use FlashInfer fused top-k for packed PAGED rows (#33006)
zianglih
authored
a9654ea
Remove 11 dead CODEOWNERS rules (#34875)
DanielSwift1992
authored
17efe37
[Docs] Update Kimi-K3 installation options (#34886)
Fridge003
authored
8b4faa3
[Fix: RL] Snapshot async state-capture outputs before overlap (#34319)
jvmncs
authored
90b3db6
[RL] Expose top-p-only sampling masks (#33593)
nanjiangwill
authored
be804c1
[Fix] Work around xgrammar 0.2.1 negative integer minimum in Kimi-K3 structural tags (#34778)

gilfordting
and
chromecast56
authored
3f64f14
[CI] Trim Qwen3.5 FP8 GB300 performance batches (#34882)
Fridge003
authored
a184470
[Fix] Require JSON booleans for response_format json_schema.strict (#34777)

gilfordting
and
chromecast56
authored
6f005e4

Commits on Aug 15, 2026

Commits on Aug 16, 2026

Commits on Aug 17, 2026

Commits on Aug 18, 2026

Commits on Aug 19, 2026

[Spec] DFlash2: local convolution + candidate selector (#35371)

authored
c14312a
[Doc] Fix TP and attention-TP group layout in initialize_model_parallel docstring (#34862)

QNan0513
and
BBuf
authored
64e4042
[perf] overlap page preprocessing, pack the vit, enable prefill CUDA graph for paddle-ocr (#35318)

mickqian
and
claude
authored
77fc5c1
[diffusion] feat: support cache-dit, cfg gating, attention backend override as per-request param (#35339)

mickqian
and
claude
authored
e73201e
quant: extract shared checkpoint quant metadata resolver (#35172)
mickqian
authored
ef49085
[diffusion] chore: make --vae-tiling honest, fix the decode oom advice, gate nvfp4 on blackwell (#35353)
mickqian
authored
baa2251
Add fmha_v2 attention backend for SM90/120 (#23112)

akhilg-nv
and
YAMY1234
authored
5d12280
Fix DP attention on CPU (#12961)
chunyuan-w
authored
58c5bee
[diffusion] refactor: reuse srt qwen vision and text modules (#35006)
mickqian
authored
4cef72f
[HiCache] Buffer-only mode for HiCache host memory layer (#34798)
xiezhq-hermann
authored
977412a
[diffusion] fix: make MiniMax-H3 AdaLN cache rebuild transactional (#34993)

Dayuxiaoshui
and
mickqian
authored
eb08552
[diffusion][Minimax H3]support subblock sparse attention on SM90 (#34680)
IPostYellow
authored
ee1f2e8
[Perf] Restore the 16-token router GEMM threshold on SM10X (#34953)

mmangkad
and
Mohammad Angkad
authored
1beb805
[HiCache] Batch PP write and load completion sync (#33473)
luoroger37
authored
3b065a5
feat(api): add sglext_spec (#33518)

Muqi1029
and
codex
authored
88f6074
[Fix] DCP: advertise the logical KV-event block size (#35298)

Jiminator
and
claude
authored
c863760
Fix HiCache PP sync test fixture (#35446)
alphabetc1
authored
a72d8d2
[Constrained] Support MistralCommon tokenizers in the XGrammar backend (#35215)

alisonshao
and
JustinTong0323
authored
3391ab3
[diffusion] optimization: reduce minimax h3 mps memory pressure (#33880)
mickqian
authored
593b1a9
[diffusion] chore: reuse shared checkpoint quant metadata resolver (#35174)
mickqian
authored
a5c9636
VLM: feed the packed qkv projection output to vision backends uncopied (#35336)

mickqian
and
claude
authored
c0c87e0
[AMD] Let the diffusion AITer backend take grouped-query K/V (fix Cosmos3-Nano startup) (#34485)

michaelzhang-ai
and
cursoragent
authored
f415871
[Fix] Scale the req_to_token row headroom by attn_dcp_size (#35424)
hnyls2002
authored
e614121
[CI] Trim the base-c 4-gpu-h100 stage from 5 shards to 4 (#35407)
hnyls2002
authored
ccbe380
[diffusion] refactor: gate native encoder quantized checkpoints (#35183)

mickqian
and
decajoin
authored
73e5fa4
[AMD][DI][CI] Run MI355X disagg nightly at 7AM UTC (#35467)

kangwangamd
and
bingxche
authored
ebec85f
Qwen3.8-27B Model Support (#34859)

authored
8a1e6e4
[docs] Add a fused-kernels page for SGLang Diffusion (#35436)

BBuf
and
claude
authored
9113fc6
[HiCache] Support DCP with DSpark (#35221)

kpham-sgl
and
cursoragent
authored
0e4a094
[PD] Overlap prefill DP-rank bootstrap queries (#35071)
YAMY1234
authored
aa215e5
[PD] Deferred decode-side KV release for the NIXL backend (#35360)
ShangmingCai
authored
adca19c
Add three new test cases (#35502)

HeYao0714
and
HeYao
authored
f22442d
[AMD] DeepSeek-V4 MI355X: eliminate bpreshuffle fp8-scale relayout copy in dense w8a8 linear (#33165)
karverma-amd
authored
ce1830c
[AMD] DeepSeek-V4: route decode wo_a bf16 batched matmul to aiter batched_gemm_bf16 (#33313)

authored
f446e85
fix: fix transcription & audio-understanding for ASR/audio/speech models (#32611)

SKRohit
and
Singh
authored
3e5ce26
[AMD] cookbook: serve Qwen3.5 MXFP4 on MI355X with an fp8_e4m3 KV cache (#35445)
yichiche
authored
5742746

Commits on Aug 20, 2026

[AMD] [GLM5] Fuse the ROCm DSA indexer q/k prep into aiter's single kernel

Jacob0226
and
cursoragent
committed
82bd89f