SolitaryThinker 6096b00aeb [feat] v2 LTX-2.3 T2VS GPU-verified: audio VAE/vocoder wiring + dual-connector audio fix ...

The full joint text->video+audio LTX-2.3 path now generates on the real 18.99B model:
- GPU audio components: build_torch_audio_vae (AudioDecoderLoader -> LTX2AudioDecoder, chains the
  vocoder) + build_torch_vocoder (VocoderLoader -> LTX2Vocoder); registered the 'audio_vae'/'vocoder'
  cuda component kinds; stamped their checkpoint subfolders (_WAN21_SUBFOLDERS).
- Fix: TorchGemma.encode_av must pass output_hidden_states=True — the 2.3 connector's SEPARATE audio
  projection lives in hidden_states[0] only then (gemma.py:703); without it the audio text fell back to
  the video embedding (4096 vs 2048 -> audio cross-attn shape mismatch).
- GPU-verified T2VS: video (3,33,256,384, std 0.68) + audio (stereo 2x61920 @24kHz, nonzero, std 0.059).
- test_torch_backend: cuda component kinds now include audio_vae + vocoder. CPU suite green (202+2).

2026-07-04 17:15:34 +00:00

..

test_adapters.py

[feat] Adapter plane, non-linear workflows, RL→distill flywheel

2026-07-04 17:15:34 +00:00

test_adaptive_compute.py

[feat] Five more stress tests: LTX-2 A/V, weight-sync, served reward, cache-dit, nested workflows

2026-07-04 17:15:34 +00:00

test_admission.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_cache.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_contracts.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_cudagraph.py

[feat] static-buffer capture form for the cudagraph step body (Path A)

2026-07-04 17:15:34 +00:00

test_extensions.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_flywheel.py

[feat] Adapter plane, non-linear workflows, RL→distill flywheel

2026-07-04 17:15:34 +00:00

test_inference.py

[fix] v2: name LTX-2 cards by architecture (2-stage vs single-stage) + Wan2.1 is T2V-only

2026-07-04 17:15:34 +00:00

test_interleave_gate.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_joint_multi_rl.py

[feat] Cross-model T2I→I2V workflow + N-way joint RL over arbitrary experts

2026-07-04 17:15:34 +00:00

test_ltx2_av.py

[fix] v2: name LTX-2 cards by architecture (2-stage vs single-stage) + Wan2.1 is T2V-only

2026-07-04 17:15:34 +00:00

test_nested_workflow.py

[feat] Five more stress tests: LTX-2 A/V, weight-sync, served reward, cache-dit, nested workflows

2026-07-04 17:15:34 +00:00

test_omni.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_parity.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_platform.py

[fix] v2 tests: make no-torch-import guards GPU-aware (skipif torch installed)

2026-07-04 17:15:34 +00:00

test_policies.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_review_fixes.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_served_reward.py

[feat] Five more stress tests: LTX-2 A/V, weight-sync, served reward, cache-dit, nested workflows

2026-07-04 17:15:34 +00:00

test_serving_fixes.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_serving.py

[fix] v2: name LTX-2 cards by architecture (2-stage vs single-stage) + Wan2.1 is T2V-only

2026-07-04 17:15:34 +00:00

test_speculative.py

[feat] #8b speculative (draft-verify) decoding — exact + lower-latency AR

2026-07-04 17:15:34 +00:00

test_t2i_i2v_workflow.py

[feat] Register cross-model workflows as first-class named servables

2026-07-04 17:15:34 +00:00

test_thinker_talker.py

[feat] Add Qwen-Omni thinker→talker→vocoder model (3 experts, 3 loops)

2026-07-04 17:15:34 +00:00

test_tiled_scheduling.py

[feat] Three stress tests: interactive sessions, workflow RL, heterogeneous co-scheduling

2026-07-04 17:15:34 +00:00

test_torch_backend.py

[feat] v2 LTX-2.3 T2VS GPU-verified: audio VAE/vocoder wiring + dual-connector audio fix

2026-07-04 17:15:34 +00:00

test_training.py

[refactor] rename package mini_fastvideo → v2

2026-07-04 17:15:34 +00:00

test_unified_rl.py

[feat] UniRL/PromptRL joint LM+generator RL stress test + designv4

2026-07-04 17:15:34 +00:00

test_weight_sync_live.py

[feat] Five more stress tests: LTX-2 A/V, weight-sync, served reward, cache-dit, nested workflows

2026-07-04 17:15:34 +00:00

test_workflow_rl.py

[feat] Three stress tests: interactive sessions, workflow RL, heterogeneous co-scheduling

2026-07-04 17:15:34 +00:00

test_workflow_shapes.py

[fix] v2: name LTX-2 cards by architecture (2-stage vs single-stage) + Wan2.1 is T2V-only

2026-07-04 17:15:34 +00:00

test_world_session.py

[feat] Three stress tests: interactive sessions, workflow RL, heterogeneous co-scheduling

2026-07-04 17:15:34 +00:00