Contrib: FLUX.1-lite-8B-alpha (native FLUX.1 compatibility)#147
Open
jimburtoft wants to merge 4 commits into
Open
Contrib: FLUX.1-lite-8B-alpha (native FLUX.1 compatibility)#147jimburtoft wants to merge 4 commits into
jimburtoft wants to merge 4 commits into
Conversation
FLUX.1-lite-8B-alpha (Freepik) is architecturally identical to FLUX.1-dev with 8 double-stream blocks instead of 19. NxDI's FLUX.1 implementation reads num_layers from config.json at runtime, so it works out of the box with FLUX.1-lite weights -- no custom modeling code needed. Validated on trn2.3xlarge (LNC=2, TP=4): - 5.91s per 1024x1024 image (25 steps) - 4.49 backbone fwd/sec - ~128s compilation time - SDK 2.29, NxD Inference 0.9
| @@ -0,0 +1,3 @@ | |||
| # NxDI FLUX.1-lite-8B-alpha Diffusion Model | |||
There was a problem hiding this comment.
this __init__.py is just 3 comment lines with no actual imports/exports. Since there's no custom modeling code just a generation script, this is technically fine, but for future, we should be exporting the generation function for programmatic use: from .generate_flux_lite import run_generate.
|
The first commit, |
tejasamx-aws
approved these changes
May 10, 2026
| print(f"Image saved, pixel std={img_array.std():.1f}") | ||
|
|
||
|
|
||
| def test_warm_generation_time(neuron_app): |
There was a problem hiding this comment.
The description reports 5.91s average generation time, but the test threshold is 15s. we should tighen to 10s to catch performance regressions earlier.
Add generate_flux_lite_highres.py script and documentation for generating images at 2048x2048 and 4096x4096 resolution on Neuron. The original FLUX.1-lite/dev/schnell models do not natively support these resolutions. Key implementation: - 2K (16,384 tokens): Backbone compiled natively at TP=4 on trn2.3xlarge. VAE uses tiled decode (4 tiles) since it exceeds instruction limit at 2K. - 4K (65,536 tokens): Context parallelism (TP=4, CP=4, world_size=16) on trn2.48xlarge splits sequence to 16,384 tokens/shard. VAE uses 25 tiles. Requires NEURON_RT_VISIBLE_CORES=0-15 to prevent 64-rank collective deadlock. Benchmark results: - 1K: 5.91s, 2K: 31.53s, 4K: 107.25s (25 steps, guidance_scale=3.5) - 4K backbone: 3.98s/step, tiled VAE: 0.29s/tile Validated on SDK 2.30 (DLAMI 20260522), trn2.48xlarge.
lutfanm-aws
approved these changes
May 27, 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.
Summary
num_layersfrom the model'sconfig.jsonat runtime, so FLUX.1-lite works out of the box with no custom modeling code.Validation Results (trn2.3xlarge, LNC=2, TP=4)
Checklist
Model Type
Contribution Contents
src/directory with generation scripttest/integration/test_model.pywith 3 passing testsTesting
SDK Compatibility