[FLUX.2] Add e2e text-to-image pipeline: nightly + benchmark

[ashokkumarkannan1]

Ticket

• [FLUX.2- dev] Wire e2e pipeline into nightly and benchmark CI #5234

Problem description

• FLUX.2-dev passed component bring-up (text encoder, transformer, VAE decoder — added in FLUX.2: Add component tests (text encoder, transformer, VAE decoder) #5317) and the full e2e text-to-image pipeline runs on Tenstorrent, but there was no e2e pipeline wired into benchmark CI, and the nightly e2e test carried no machine marker, so nothing in CI actually collected/ran it.
• This mirrors the image-gen pattern introduced in [Playground v2.5] Wire e2e pipeline into nightly and benchmark CI #5044 (Playground v2.5) and extended by SDXL-Lightning ([SDXL Lightning] Add e2e pipeline in nightly and benchmark CI #5244) and Janus-Pro ([Janus-Pro] Add e2e text-to-image pipeline (1B + 7B): nightly + benchmark #5291): a model whose components pass bring-up should get an e2e pipeline in nightly + a perf benchmark.
• FLUX.2 is the first multichip (tensor-parallel) image-gen model to be wired in: the ~24B Mistral3 text encoder and ~32B Flux2 transformer are SPMD-sharded across the mesh's model axis (contraction-parallel degree 4); the VAE is replicated. It targets the 4-chip Blackhole QuietBox (qb2-blackhole).

What's changed

Wires the full FLUX.2-dev pipeline (the standard diffusers Flux2Pipeline orchestrates; every compute module runs on TT via torch.compile(backend="tt"), tensor-parallel sharded with the same SPMD shard specs as the component tests) into nightly + benchmark CI, following the playground_v2_5 / sdxl_lightning e2e conventions.

• Nightly (tests/torch/models/flux2/test_flux2_pipeline.py): self-contained — pipeline class + test inline, imports shard specs from third_party.tt_forge_models (no import from examples/). Added @pytest.mark.qb2_blackhole so the existing standalone-models multichip nightly job collects it; markers: tensor_parallel, nightly, model_test, large, qb2_blackhole, record_test_properties. Mesh is now selected from MESH_SHAPES[num_devices] (model-parallel axis fixed at degree 4; no-op at 4 chips). Memory strategy: the text encoder is placed → used → evicted before the transformer is placed, and the VAE is placed lazily at first decode, so peak DRAM ≈ max(component) rather than the sum.
• Benchmark: the shared diffusion harness tests/benchmark/benchmarks/imagegen_benchmark.py is made component-set-agnostic — a model reports only the components it runs (FLUX.2 has a single text encoder, so te2 is omitted from the breakdown + measurements), with an optional per-step label and mesh_shape reporting. SDXL/Playground output is unchanged (verified on hardware, see table). A self-contained multichip benchmarks/flux2_pipeline.py sets CONVERT_SHLO_TO_SHARDY=1 before use_spmd() (required so the StableHLO handed to tt-mlir carries shardy annotations) and builds the mesh from MESH_SHAPES[num_devices], with per-component _perf timing. Entry test_flux2 in test_imagegen.py. Two-pass warmup + steady-state. optimization_level=0.
• Matrix: flux2 entry added to .github/workflows/perf-bench-matrix.json with runs-on: qb2-blackhole (4-chip). A qb2-blackhole standalone-models nightly job (scoped to the flux2 dir, mark qb2_blackhole) added to model-test-lb-blackhole-nightly.json.
• Machine placement: the component correctness tests (FLUX.2: Add component tests (text encoder, transformer, VAE decoder) #5317) run on lb-blackhole (the functional-nightly convention); the e2e pipeline + perf benchmark run on qb2-blackhole — matching the perf-matrix convention (all Blackhole multichip perf runs on qb2) and co-locating the e2e with the box it was validated on.

Validated on a 4-chip qb2 Blackhole:

Test	Result	Time
test_flux2_pipeline (nightly e2e)	✅ passed	20m59s (cold cache)
test_flux2 (benchmark)	✅ passed	19m54s (warmup+steady)

Steady-state benchmark numbers (qb2-blackhole, 4 chips, opt_level=0, 50 steps):

metric	value
e2e latency	835.8s
warmup (incl. compile)	336.2s
text encoder	29.8s
transformer step mean	13.55s
VAE	6.4s

Both the nightly and benchmark passes produce a coherent 1024×1024 image.

Checklist

• Validate nightly e2e + benchmark on 4-chip qb2 blackhole (logs - zip)

• Confirm single-chip imagegen (SDXL-Lightning) benchmark unaffected by the harness change

• Perf benchmark CI run (test_flux2, qb2-blackhole): https://github.com/tenstorrent/tt-xla/actions/runs/28118639108