[feature]support vllm-ascend delta disk weight - #428

Closed

wangx700 wants to merge 63 commits into

vllm-project:mainfrom

wangx700:feat/pr413-disk-weight-sync

Commits on Jun 11, 2026

ascend: adapt vime for NPU
CalvinXKY
committed
766988c
[NPU][A3] Add patch workflow and Docker support for Atlas 800I A3 (#163)
CalvinXKY
committed
ce20158
fix(npu): adapt convert_hf_to_torch_dist.py for NPU with mindspeed and HCCL support
CalvinXKY
committed
765c383
fix(npu): merge PR #230 changes into ascend branch
CalvinXKY
committed
c441ca7
fix(npu): restore is_npu() conditionals in convert script and add patch robustness flags
CalvinXKY
committed
41f9ed3

Commits on Jun 12, 2026

Commits on Jun 13, 2026

fix(npu): bypass broken initialize_layerwise_reload and correct Ray NPU resource allocation
CalvinXKY
committed
7c3c67b

Commits on Jun 17, 2026

rebase dockerfile to main
CalvinXKY
committed
4ff601e

Commits on Jun 18, 2026

Commits on Jun 22, 2026

[NPU] fix qwen3-30b ep (#269)
floatlibai
authored
141a107

Commits on Jun 23, 2026

feat(npu): use vllm-ascend verified commit in Dockerfile.npu (#281)

Meihan-chen
and
yzeyu71
authored
75a2c80

Commits on Jun 25, 2026

fix(npu): default megatron-to-hf-mode to bridge on Ascend (#275)
CalvinXKY
authored
e78a315

Commits on Jun 30, 2026

Commits on Jul 1, 2026

Commits on Jul 2, 2026

feat(npu): use released vllm-ascend base image in Dockerfile.npu (#305)

Meihan-chen
and
claude
authored
7ee7aee

Commits on Jul 10, 2026

Commits on Jul 16, 2026

【NPU】vime NPU CI Integration (#316)
yzeyu71
authored
82c7876
[NPU] Add Retool example (#333)
Fulin-Gao
authored
376468e

Commits on Jul 17, 2026

[NPU] Enable Qwen3-VL GRPO (single- & multi-turn) on Ascend (#314)
Meihan-chen
authored
fde9f6f

Commits on Jul 20, 2026

refactor(vllm): use native IPC engine for colocated weight sync (#345)
miracle0517
authored
a92ee13

Commits on Jul 21, 2026

[NPU] refactor(ci): device-aware test harness for GPU/NPU + Qwen3 NPU smoke tests (#349)

Meihan-chen
and
claude
authored
235ab90

Commits on Jul 22, 2026

[NPU] fix default qwen3-4b script (#364)
floatlibai
authored
45921d4
[NPU] support multi-agent rollout pipeline (#298)

Windfeng8
and
wky
authored
ca1bacf

Commits on Jul 23, 2026

chore/add npu ci image build step (#361)
yzeyu71
authored
c67acc8

Commits on Jul 24, 2026

add npu 4b fully_async script (#368)
floatlibai
authored
09aeb60

Commits on Jul 25, 2026

[fix] save dist checkpoint for MoE models (#366)
yuxinshan
authored
821c933

Commits on Jul 28, 2026

chore/update pip repo (#373)
yzeyu71
authored
573b1bf

Commits on Aug 3, 2026

fix: restore MoE ALLTOALL expert IDs after NPU memory reuse (#376)
miracle0517
authored
e5aa5e7

Commits on Aug 21, 2026

[NPU][CI] Refactor patch reconciliation (#390)
Meihan-chen
authored
9afaf89

Commits on Sep 1, 2026

Apply-MegatronAdaptor-NPU-migration-to-clean-branch (#385)

momo609
and
wangxiaoxin-sherie
authored
35eac6b
refactor(npu): establish S0 platformized baseline
Meihan-chen
committed
f1483a0

Commits on Sep 2, 2026

sync(main): merge S1 through 69cb849b
Meihan-chen
committed
2543862
[NPU] Add Qwen3.5-35B-A3B support (#409)
Fulin-Gao
authored
f5b8491

Commits on Sep 3, 2026

[NPU] Fix load torch_dist weight error (#396)
floatlibai
authored
745d952
sync(main): merge S2 through e0ca8283
Meihan-chen
committed
31782f2
fix(npu): adapt S2 arguments and VL memory budget
Meihan-chen
committed
7d9f977

Commits on Sep 4, 2026

sync(main): merge S3-S4 through 8144096e
Meihan-chen
committed
4e90cfa
fix(npu): adapt S3-S4 to vLLM 0.25.1
Meihan-chen
committed
edea686

Commits on Sep 7, 2026

sync(main): merge S5 through 85ac4267
Meihan-chen
committed
b879eb1
fix(npu): adapt S5 to vLLM 0.27.1
Meihan-chen
committed
f930676

Commits on Sep 8, 2026

sync(main): merge S6 through ce92eff1
Meihan-chen
committed
e5f7e24
fix(npu): adapt S6 to vLLM e6bfe03a
Meihan-chen
committed
4c182d4

Commits on Sep 9, 2026

sync(ascend): merge S7 through 745d952e
Meihan-chen
committed
9d72be5
fix(npu): adapt S7 runtime and Qwen3.5 GDN
Meihan-chen
committed
7a27808
fix(npu): isolate serving FLA libraries and force spawn
Meihan-chen
committed
81ae79d
revert(npu): defer Qwen3.5 support beyond S7
Meihan-chen
committed
46f325c
test(npu): enable torch_dist ref-load by default for Qwen3-30B
Meihan-chen
committed
e0757f9

Commits on Sep 10, 2026

feat: support GLM-4.7-Flash (MLA + MoE) on Ascend NPU (#403)

is-not
and
louyujing
authored
083aea6
style(npu): satisfy pre-commit formatting checks
Meihan-chen
committed
95f6309
merge(npu): integrate GLM-4.7-Flash native G1 support
Meihan-chen
committed
1d55086
fix(npu): enable GLM MTP draft updates and graph execution
Meihan-chen
committed
edf744e

Commits on Sep 14, 2026

fix(ascend): harden disk delta checkpoint handling
wangx700
committed
8f3ffc9