Add Cosmos3-Super VideoPhy-2 Reasoner SFT recipe#101
Merged
Conversation
Add the Super-tier counterpart to the Nano VideoPhy-2 recipe. It fully fine-tunes Qwen3-VL-32B
using FSDP full sharding and initializes from the Cosmos3-Super language model merged with the 32B
visual tower.
The new experiment mirrors the Nano configuration but uses the 32B backbone and sets
data_parallel_shard_degree=-1. It also adds a launcher and TOML recipe.
Use a learning rate of 1e-6 for both tiers. In a 150-iteration Super run, 1e-5 caused a temporary
loss spike to about 6.6 at iteration 23. Both learning rates reached the same last-20 average loss of
0.437, so 1e-6 provides the same result without the instability.
Set lr_multipliers={"model.visual": 1.0} to override the reasoner default of 0.1 and train the
visual projector at the base learning rate. The previous mm_projector and merger keys matched no Qwen3-
VL parameters because the projector is named model.visual.merger, so the intended 20x multiplier was
never applied.
Document the Super recipe in docs/training.md and examples/README.md.
Xuanmeng-Zhang
requested review from
Dinghow,
foreverlms,
lfengad and
yy-code-nv
lfengad
approved these changes
Jul 10, 2026
Dinghow
approved these changes
Jul 10, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Add Cosmos3-Super VideoPhy-2 Reasoner SFT recipe. It fully fine-tunes
Qwen3-VL-32Busing FSDP full sharding and initializes from the Cosmos3-Super language model merged with the 32B visual tower.The new experiment adds a launcher and TOML recipe.
Use a learning rate of
1e-6for both Nano and Super.Set
lr_multipliers={"model.visual": 1.0}to override the reasoner default of0.1and train the visual projector at the base learning rate. The previousmm_projectorandmergerkeys matched noQwen3-VLparameters because the projector is namedmodel.visual.merger, so the intended20xmultiplier was never applied.Document the Super recipe in
docs/training.mdandexamples/README.md.