feat(vlm): add Kimi-K3 pokemon ShareGPT SFT recipe with collate and dataset - #1
Open
xiaoliang0601 wants to merge 3 commits into
Open
xiaoliang0601 wants to merge 3 commits into
xiaoliang0601 wants to merge 3 commits into
Conversation
…ded save Parameters that never receive a gradient (e.g. Kimi-K3 layer-0 self_attention_res_* params, whose first residual mix is skipped while the block residual is still empty) never get lazily-created Adam optimizer state. get_optimizer_state_dict() then silently omits their state at save time, yet the optimizer expects it at resume, so DCP load fails with a missing "...weight.step" key. Mirror the native (non-sharded) path and materialize zero state for all params before building the sharded optimizer state dict. Co-Authored-By: Claude <noreply@anthropic.com> Signed-off-by: root <xuezunyao@xiaohongshu.com>
…ataset Enable Kimi-K3 VLM SFT on the pokemon-gpt4o Chinese caption set: - kimi_k3_vl_collate_fn: feeds the model a single <|media_pad|> per image (K3's forward expands it internally) and returns merged-length, pre-shifted labels for the recipe's no-shift MaskedCrossEntropy. Right-pads to batch_max+1 so left_padding is always False and the internal image-token merge layout is deterministic. - make_pokemon_sharegpt_dataset: loads the local ShareGPT parquet, decodes inline image bytes lazily, and maps human/gpt turns to the user/assistant conversation structure the collate expects. - k3_pokemon_zh.yaml: FSDP2 + EP8 recipe (K3 has no TP), 200-step run against the real 5L/128E MXFP4 checkpoint. Co-Authored-By: Claude <noreply@anthropic.com> Signed-off-by: root <xuezunyao@xiaohongshu.com>
xiaoliang0601
force-pushed
the
huiyingl/feat/k3-vlm-pokemon-sft
branch
from
September 15, 2026 18:09
0751bd9 to
4fef284
Compare
torch.optim.AdamW keeps no internal fp32 master, so the resident parameter is the master copy. bf16 storage rounds Adam's ~lr-sized updates below the bf16 ULP (swamping), stalling early descent; float32 storage captures them and matches Relax's precision-aware fp32-master optimizer. Also drop warmup to 0 to align with the Relax reference (Megatron default). FSDP2 still computes/all-gathers in bf16 via mp_policy. Co-Authored-By: Claude <noreply@anthropic.com>
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
1 participant
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.
No description provided.