Add vl support - #7

Merged

curnane-lab merged 128 commits into

vl-dflash-prfrom

Aug 27, 2026

Commits on Jul 30, 2026

dflash sliding window layer support, fix eager backend
dcw02
committed
8de026a
make anchor sampling batch invariant and normalize loss globally across optimizer steps
dcw02
committed
7757409

Commits on Aug 1, 2026

Add configurable learning-rate schedules

maocheng
authored and
maocheng
committed
9fc6514
Decouple online prompt ordering from run seed

maocheng
authored and
maocheng
committed
3ef0094
Add Kimi K3 DSpark capture support

maocheng
authored and
maocheng
committed
99431d9
Add Kimi K3 V1C disaggregated recipe

maocheng
authored and
maocheng
committed
967fd9e
Support Inkling managed capture
maocheng23
committed
9ed2462
Add Inkling two-node recipe
maocheng23
committed
623c07e
docs: harden Kimi K3 disaggregated launch settings

maocheng
authored and
maocheng
committed
2021359
feat(dspark): add Kimi-K3 MLA and KDA draft backbones

maocheng
authored and
maocheng
committed
b7d21db
fix(data): register Kimi-K3 thinking template

maocheng
authored and
maocheng
committed
15fd6c4
fix(runtime): force terminal Mooncake cleanup

maocheng
authored and
maocheng
committed
2ecd685
fix(runtime): validate online flow window before data build

maocheng
authored and
maocheng
committed
af428cc
fix(recipe): preserve Kimi K3 global batch window

maocheng
authored and
maocheng
committed
e1a4bd1
fix(runtime): fill optimizer window before byte throttle

maocheng
authored and
maocheng
committed
daa41a3
fix(tracking): log resolved config with recursive redaction

maocheng
authored and
maocheng
committed
6ea5ab1
fix(runtime): stream large online prompt datasets

maocheng
authored and
maocheng
committed
3e27ad2
fix(model): compile long Kimi K3 MLA attention

maocheng
authored and
maocheng
committed
bfb662b
fix(model): fuse expanded long Kimi MLA attention

maocheng
authored and
maocheng
committed
c76ca3c
fix(runtime): reclaim durable Mooncake samples per step

maocheng
authored and
maocheng
committed
16e4d74
fix(dspark): checkpoint config carries BOTH rope schemas
Boreas618
committed
108eda3
fix(dspark): shard production KDA launches

maocheng
authored and
maocheng
committed
dd8b00a
fix(tracking): retain owned W&B run handle

maocheng
authored and
maocheng
committed
3faa234
fix(tracking): commit W&B step metrics

maocheng
authored and
maocheng
committed
5301d1e

Commits on Aug 2, 2026

fix(kimi-k3): pipeline capture and trainer throughput

maocheng
authored and
maocheng
committed
4095c16
fix(kimi-k3): complete disaggregated training adoption

maocheng
authored and
maocheng
committed
e62fe83
style: apply repository formatters

maocheng
authored and
maocheng
committed
27502ff

Commits on Aug 3, 2026

fix lint and fix:generate rope_theta that matches old transformer version
jiapingW
committed
0c47c0a
Merge branch 'main' into agent/dual-rope-export
jiapingW
authored
5072675
fix test
dcw02
committed
40d88f8
Merge branch 'main' into agent/inkling-sglang-capture-compat
jiapingW
authored
745b092
fix lint and del overfit intro in examples/gates/README
jiapingW
committed
2dcfcdf
more fixes
dcw02
committed
edc450a
Merge pull request #734 from maocheng23/agent/inkling-sglang-capture-compat
jiapingW
authored
5c93516
Merge pull request #731 from modal-projects/dcw02/dflash-correctness
jiapingW
authored
f53db79
Merge branch 'main' into agent/kimi-k3-dspark-disaggregated-recipe
jiapingW
authored
ec6af00
fix lint and merge main
jiapingW
committed
5a8d926
restore code according to PR 731 and fix unittest
jiapingW
committed
ed06ab2
feat: Kimi-K3 DSpark support -- template, draft config, example run
Boreas618
committed
469da28
perf: make reference channel polling linear
ddlearn
committed
3e71683
Merge pull request #735 from maocheng23/agent/kimi-k3-dspark-disaggregated-recipe
jiapingW
authored
642f6ed
Merge branch 'main' into feat/kimi-k3-dspark-support
jiapingW
authored
e87ae6c
update news about specforge v0.3.0
jiapingW
committed
2768619
add inkling and kimi-k3
jiapingW
committed
0d505d6

Commits on Aug 4, 2026

Remove in-house Kimi-K3 MLA/KDA experiments and publish a single renamed DSpark recipe.
Boreas618
committed
f19d389
Fix the leftover hardcoded recipe count in tests/test_config/test_unified_feature_reachability.py (67 → 64). That test passes now.
Boreas618
committed
a0ae7f4
Merge pull request #752 from Boreas618/main
jiapingW
authored
f110be3
fix
jiapingW
committed
211d83a
DFlash/Domino: Add support for the FlexAttention FLASH backend
zou3519
committed
3107f3c
support sliding-window

jiapingW
authored and
zou3519
committed
1e014d1
update lmsys blog link
jiapingW
committed
19836f8

Commits on Aug 5, 2026

Merge pull request #751 from sgl-project/news_update
jiapingW
authored
7d5a693
fix(data): attach record images in regenerate_train_data for VLM prompts
mingliangfu
committed
a8393c7
docs: add ROCm known-issues notes to installation guide

Chen, Todd
authored and
zhentaocc
committed
eaaf37f
fix: update Python version requirement in pyproject.toml

Chen, Todd
authored and
zhentaocc
committed
ac075b0
fix: update exception handling in Trainer class

Chen, Todd
authored and
zhentaocc
committed
b788657
feat: enhance spec_capture_sink with dynamic pinning configuration

Chen, Todd
authored and
zhentaocc
committed
a8c0993
docs: rewrite ROCm installation for container-based sglang v0.5.14 flow

Chen, Todd
authored and
zhentaocc
committed
2b43be9
rocm: map hard-pin onto with_soft_pin on older Mooncake builds

Chen, Todd
authored and
zhentaocc
committed
33b5baa
managed_local: expose Mooncake default_kv_lease_ttl to fix teardown drain

Chen, Todd
authored and
zhentaocc
committed
8a3bcdc
docs: add dedicated AMD ROCm tutorial and link from installation guide

Chen, Todd
authored and
zhentaocc
committed
0790275
docs: move and expand AMD ROCm guide into dedicated basic_usage tutorial

Chen, Todd
authored and
zhentaocc
committed
b29cff3
docs: add ROCm Qwen3.5-4B DFlash reference results for MI300X and MI355X

Chen, Todd
authored and
zhentaocc
committed
ff275f6
docs(configs): add AMD ROCm Qwen3.5-4B DFlash example configs and update tutorial

Chen, Todd
authored and
zhentaocc
committed
53a7f39
fix: disable SGLang radix cache by default and allow opt-in

Chen, Todd
authored and
zhentaocc
committed
4a2701c
perf: make server capture requests prefill-only
ddlearn
committed
f5fa579
server-capture: async capture streaming + DSpark method in the v0.5.14 patch

maocheng23
and
claude
committed
2a6c2e0

Commits on Aug 7, 2026

optimize train
jiapingW
committed
e8f0879
flex_attention use triton not auto
jiapingW
committed
a87aa7d
add test sh
jiapingW
committed
7df454f

Commits on Aug 8, 2026

Merge pull request #757 from maocheng23/feat/v0514-async-dspark-capture
jiapingW
authored
e09e3c3
Merge remote-tracking branch 'origin/main' into rocm-foundation
jiapingW
committed
24d0565
Merge branch 'main' into perf/zero-max-new-tokens
jiapingW
authored
35b94c3
Merge pull request #749 from ddlearn/perf/linear-ref-channel-poll
jiapingW
authored
d27bb69
Merge pull request #742 from Boreas618/agent/dual-rope-export
jiapingW
authored
a6b6b02
Merge remote-tracking branch 'origin/main' into optimize-train
jiapingW
committed
0ac2198
update dir and del fixed path
jiapingW
committed
7e1c767
fix unittest
jiapingW
committed
ea259cc

Commits on Aug 9, 2026

Merge branch 'main' into perf/zero-max-new-tokens
jiapingW
authored
12bd2be
fix CI:del unused network
jiapingW
committed
17fee18
Merge pull request #756 from ddlearn/perf/zero-max-new-tokens
jiapingW
authored
a9637c3

Commits on Aug 10, 2026

feat(npu): managed-local disaggregated DFlash online training on Ascend NPU
jiapingW
committed
4bd6371
Merge pull request #722 from curnane-lab/npu_disaggregated
jiapingW
authored
87e8cf4
Merge remote-tracking branch 'origin/main' into rocm-foundation
jiapingW
committed
c0b7644
--sglang-disable-radix-cache is false default.
jiapingW
committed
2cde75c
Merge pull request #760 from sgl-project/fix_ci
jiapingW
authored
368c077
fix small bugs
jiapingW
committed
ccce246
release: prepare PyPI 0.2.0 publishing
jiapingW
committed
20e952e
ci: update PyPI workflow actions
jiapingW
committed
72ddf63
Merge pull request #698 from zhentaocc/rocm-foundation
jiapingW
authored
368217e
docs: add Ascend NPU tutorial
mingliangfu
committed
2233a65
use the fixed parameters:max_attempts and retry_interval_s
jiapingW
committed
5b63574
Merge pull request #714 from zou3519/rzou/flex-attention-flash-backend
jiapingW
authored
e6440f0
Merge branch 'main' into add_npu_doc
curnane-lab
authored
c809bd5

Commits on Aug 12, 2026

Commits on Aug 14, 2026

Merge pull request #759 from sgl-project/optimize-train
jiapingW
authored
22bfa42

Commits on Aug 17, 2026

add trim_loss_positions to eagle3 strategy
jiapingW
committed
2c1ecd2
Merge origin/main into pr-trim-a
jiapingW
committed
7a699f3
fix unittest
jiapingW
committed
30e5db5
Merge pull request #762 from curnane-lab/add_npu_doc
jiapingW
authored
4c8e0c3
Merge branch 'main' into pypi
jiapingW
authored
946aed8

Commits on Aug 18, 2026

Merge pull request #705 from julyanghar/pr-trim-a
jiapingW
authored
bdeb7d8
Merge branch 'main' into feat/kimi-k3-dspark-support
jiapingW
authored
4fcc231
fix lint
jiapingW
committed
ca90c99
fix unittest
jiapingW
committed
8b7b0bd
fix unittest
jiapingW
committed
e9b221e
Merge pull request #748 from Boreas618/feat/kimi-k3-dspark-support
jiapingW
authored
2590f48
refactor configs dir
jiapingW
committed
64ad8f7

Commits on Aug 19, 2026

polish
jiapingW
committed
e5fa49d

Commits on Aug 20, 2026

update content
jiapingW
committed
586b48b
fix unittest
jiapingW
committed
c5cac93
del hardcode config tests
jiapingW
committed
3926ae0
Fuse Domino cross entropy with Triton
zou3519
committed
93adc52

Commits on Aug 21, 2026

Merge pull request #761 from sgl-project/pypi
jiapingW
authored
10a61a1
Merge branch 'main' into update_config_desc
jiapingW
authored
3a6db47
Merge pull request #770 from sgl-project/update_config_desc
jiapingW
authored
e54e0f8
Merge pull request #716 from zou3519/rzou/domino-fused-triton-ce
jiapingW
authored
5dc34ae
Merge remote-tracking branch 'origin/main' into del_hardcode_tests
jiapingW
committed
f7b1179
perf(rocm): enable Liger fused RMSNorm/SwiGLU for AMD DFlash configs
Chen, Todd
committed
17cee84
Merge pull request #778 from sgl-project/del_hardcode_tests
jiapingW
authored
de0ea2f
feat(mtp): add Qwen3.5 MTP training support
jiapingW
committed
bc4be0b
Merge pull request #667 from curnane-lab/add_mtp_support
jiapingW
authored
cb0ebfa

Commits on Aug 25, 2026

feat(dflash): allow plain-rope drafts on multimodal captures
mingliangfu
committed
6f9fbbc

Commits on Aug 26, 2026

Commits on Aug 27, 2026

Merge upstream main into add_vl_support
mingliangfu
committed
223019b
refactor(dflash): remove the mRoPE draft path, plain-rope only
mingliangfu
committed
9a7ffbd