3/5 support kimi 2.5 full + lora: Kimi K2.5 VL full-parameter support - #1221
Merged
Conversation
nanjiangwill
requested review from
Zhichenzzz,
fzyzcjy,
guapisolo,
jybsuper,
maocheng23,
yueming-yuan and
yushengsu-thu
as code owners
May 26, 2026 23:42
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces support for Kimi VL and Kimi-K2.5 models, including weight conversion utilities, multimodal rollout data expansion, and training scripts. The review feedback highlights a critical runtime error where reversed() is called directly on a PyTorch tensor, which must be converted to a list first. Additionally, the environment variable PYTHONBUFFERED in the run script should be corrected to PYTHONUNBUFFERED. Other recommendations include using getattr for safer configuration retrieval, enforcing strict length checks in zip operations, and wrapping process group destruction in a try...finally block to guarantee cleanup.
nanjiangwill
changed the title
Add Megatron->HF weight conversion for Kimi K2.5 / VL (kimi_vl.py + dispatch in megatron_to_hf/__init__) and the language_model.-prefixed decoder layer matching in the weight iterator. Expand media placeholder tokens to grid-derived counts at train time (training_utils/data.py). General colocate weight-sync robustness used by the kimi path: IPC collect after MultiprocessingSerializer sync (memory_utils + actor) and a reloadable broadcast_object_list. Adds scripts/run-kimi-k25.sh. The original Kimi K2.5 full-parameter implementation is the work of JiLi (@GeLee-Q); this commit is the reviewable split of that work. Co-authored-by: JiLi <22428217+GeLee-Q@users.noreply.github.com>
nanjiangwill
force-pushed
the
kimi25-pr4-fullparam
branch
from
May 26, 2026 23:53
aa6712c to
dca1bd9
Compare
yueming-yuan
left a comment
yueming-yuan
left a comment
Collaborator
There was a problem hiding this comment.
overall LGTM with nits comments
data.py was growing with VL-only logic; move the 5 helpers and expand_multimodal_rollout_data_in_place into mm_data.py so generic batching stays out of multimodal preprocessing's way and future VL models can extend mm_data.py without inflating data.py. Pure mechanical refactor: function bodies are unchanged. Hardcoded media token id 163605 and 2x2 patch merger dims are promoted to KIMI_VL_MEDIA_TOKEN_ID / _KIMI_VL_MERGE_H / _KIMI_VL_MERGE_W constants. data.py keeps a re-export of expand_multimodal_rollout_data_in_place via 'from .mm_data import ...', so the two call sites in get_batch / get_data_iterator are unchanged.
nanjiangwill
force-pushed
the
kimi25-pr4-fullparam
branch
from
May 28, 2026 04:27
8f03759 to
c1ae7d3
Compare
Zhichenzzz
approved these changes
May 28, 2026
Zhichenzzz
left a comment
Zhichenzzz
left a comment
Contributor
There was a problem hiding this comment.
lgtm! Thanks @nanjiangwill
Frankstein73
pushed a commit
to sii-avalanche/miles
that referenced
this pull request
Jun 1, 2026
…radixark#1221) Co-authored-by: JiLi <22428217+GeLee-Q@users.noreply.github.com> Co-authored-by: Zhichenzzz <zczeng@uw.edu>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 3/5 of splitting #1057 (Kimi K2.5 full-param + LoRA support) into reviewable PRs, rebased on latest main.
Megatron->HF weight conversion for Kimi K2.5 / VL (
megatron_to_hf/kimi_vl.py+ dispatch inmegatron_to_hf/__init__.py) andlanguage_model.-prefixed decoder layer matching in the weight iterator (update_weight/common.py). Media placeholder tokens are expanded to grid-derived counts at train time (training_utils/data.py). General colocate weight-sync robustness used by this path: IPC collect after MultiprocessingSerializer sync (memory_utils+actor) and a reloadablebroadcast_object_list. Addsscripts/run-kimi-k25.sh.Uses the processor unification from #1219 at runtime; file sets are disjoint so it merges independently.