[TRTLLM-10616][feat] Add FLUX.1 and FLUX.2 text-to-image pipeline support - #11556

Merged

karljang merged 37 commits into

NVIDIA:mainfrom

karljang:user/kanghwan/modeling_flux

Feb 23, 2026

Commits on Feb 17, 2026

Add FLUX.1 and FLUX.2 pipeline supports

karljang
and
claude
committed
31e47a5
feat: FLUX pipeline improvements - torch.compile, FLUX.2 text encoder fix, shared FP8 handling

karljang
and
claude
committed
a97821d
fix: FLUX.1 PSNR from ~17dB to ~30dB - VAE shift_factor, T5 encoding, RoPE precision

karljang
and
claude
committed
4439d9f
fix: disable torch.compile for GELU(tanh) to prevent precision drift in FLUX.1

karljang
and
claude
committed
8fac758
feat: unify FLUX example scripts and fix FLUX.2 infer bug

karljang
and
claude
committed
6eccd65
feat: add FLUX.1 HF baseline script and fix executable permissions

karljang
and
claude
committed
fd1a815
feat: add FLUX.1 to example scripts and update FLUX.2 to unified script

karljang
and
claude
committed
b32799d
fix: extract transformer weights from nested dict after upstream WeightLoader change

karljang
and
claude
committed
eaf0c7d
refactor: consolidate FLUX.1/FLUX.2 into single flux/ package with shared modules

karljang
and
claude
committed
fdfc369
fix: adapt FLUX example to upstream VisualGen API change (prompt → inputs)

karljang
and
claude
committed
da3655d
chore: add SPDX copyright headers to all new FLUX files

karljang
and
claude
committed
a34c286
fix: remove unrelated LTX2 baseline test from hf_examples.sh

karljang
and
claude
committed
caf421a
fix: address CodeRabbit review findings

karljang
and
claude
committed
39da9c4

Commits on Feb 18, 2026

Commits on Feb 19, 2026

fix: address PR #11556 review comments for FLUX modeling

karljang
and
claude
committed
bd88f28

Commits on Feb 20, 2026

fix: unroll _make_ffn factory functions into inline constructors

karljang
and
claude
committed
b61066e
fix: use CI scratch checkpoint paths for FLUX pipeline tests

karljang
and
claude
committed
1ac5ebb
Merge branch 'main' into user/kanghwan/modeling_flux
karljang
authored
d36c238
feat: add FLUX warmup support and CI test coverage

karljang
and
claude
committed
dfc0469
refactor: remove redundant transformer_info fallback in pipeline registry

karljang
and
claude
committed
3100134
feat: add Ulysses sequence parallelism for FLUX transformers

karljang
and
claude
committed
a517968

Commits on Feb 21, 2026

feat: enable torch.compile and add combined optimizations test for FLUX

karljang
and
claude
committed
b86615e
feat: add torch.compile and warmup CLI args to FLUX example script

karljang
and
claude
committed
eeb9e3a
refactor: remove redundant out_bias parameter from Attention module

karljang
and
claude
committed
a6a0970
fix: remove redundant docstring details from Flux2ParallelSelfAttention

karljang
and
claude
committed
578036f
refactor: replace unflatten with view for QKV reshape consistency

karljang
and
claude
committed
2db8051
fix: remove redundant swiglu comment in Flux2ParallelSelfAttention

karljang
and
claude
committed
e363adf
refactor: use torch.inference_mode instead of no_grad in FLUX pipelines

karljang
and
claude
committed
9330e39
fix: remove redundant __call__ wrappers from FLUX pipelines

karljang
and
claude
committed
71a401b
refactor: move inline imports to top-level in FLUX.2 pipeline and transformer

karljang
and
claude
committed
4ef10f2
fix: remove unused position ID helpers from pos_embed_flux.py

karljang
and
claude
committed
b26a629
fix: use register_buffer for theta in FluxPosEmbed to enable torch.compile

karljang
and
claude
committed
6aa35b5
refactor: remove _per_head_norm and auto-detect 4D input in apply_qk_norm

karljang
and
claude
committed
bbcd2a0
fix: remove unused gradient_checkpointing from FLUX transformers

karljang
and
claude
committed
5c8794a