upgrade(ascend): support vLLM 0.28 on Ascend 910C - #487

Open

physics31415926 wants to merge 33 commits into

flagos-ai:0.4.0-devfrom

physics31415926:codex/ascend-vllm-0.28.0

Commits on Sep 21, 2026

fix(ascend): declare vLLM 0.28 kernel block sizes
physics31415926
committed
3a011ba
fix(ascend): adapt empty vLLM 0.28 runtime and attention contracts
physics31415926
committed
180cf27
fix(ascend): preserve hybrid cache writes and GDN prefill layout
physics31415926
committed
dc2a12c
fix(ascend): keep strided vision indexing on torch_npu
physics31415926
committed
fbb0d14
fix(ascend): run modular MoE with native grouped matmul
physics31415926
committed
7d708a4
build(ascend): pin empty vLLM 0.28 and document validated runtime
physics31415926
committed
cba7012
fix(ascend): complete vLLM 0.28 runtime contracts
physics31415926
committed
cc3b4c7
ci(ascend): expose eager 910C validation
physics31415926
committed
ef2d3cb
fix(ascend): harden rotary embedding fallbacks
physics31415926
committed
f1a42d2
feat(ascend): support DeepSeek-V4-Flash W8A8
physics31415926
committed
d951611
ci(ascend): use packaged vLLM 0.28 image
physics31415926
committed
87fdc2d
ci(ascend): enable existing 910C pipeline
physics31415926
committed
372212b
style: satisfy upstream ruff checks
physics31415926
committed
5743142
ci(ascend): install test plugins and isolate NPU test
physics31415926
committed
53c849a
fix(security): resolve CodeQL review findings
physics31415926
committed
3719c3e
ci(ascend): cap offline profile batch on 910C
physics31415926
committed
2d1684b
feat(ascend): support GLM-5.2 ModelSlim W8A8
physics31415926
committed
6af9674
test(ascend): clarify GLM config restoration path
physics31415926
committed
6782959
fix(ascend): ignore GLM MTP rotation in target load
physics31415926
committed
5246330
docs(ascend): align published CI image guidance
physics31415926
committed
1267199
fix(ascend): use native grouped matmul for W8A8 MoE
physics31415926
committed
c93aaee
fix(ascend): flatten expert map lookup for NPU routing
physics31415926
committed
9baeb8f
ci(ascend): publish final vLLM 0.28 image
physics31415926
committed
17af3e4
ci(ascend): bake test dependencies into image
physics31415926
committed
0de238b

Commits on Sep 22, 2026

fix(ascend): complete vLLM 0.28 runtime adaptation
physics31415926
committed
1e77439
test: resolve CodeQL mixed-import reviews
physics31415926
committed
b716ee5
fix(ascend): resolve remaining CodeQL warning
physics31415926
committed
3ec6052
fix(ascend): eliminate CodeQL unused branch
physics31415926
committed
2f34ee6
test(ascend): allow GDN reduction variation
physics31415926
committed
a3559ba
fix(ascend): make causal conv loads explicit
physics31415926
committed
77e0b79

Commits on Sep 23, 2026

fix(ascend): scope compatibility fallbacks to NPU
physics31415926
committed
c94d6b1

Commits on Sep 24, 2026

fix(ascend): enable FlagCX tensor parallel communication
physics31415926
committed
ebbbef9
docs(ascend): add 910C deployment guide
physics31415926
committed
39dfc47