diff --git a/examples/inference/basic/V2_PORTING_STATUS.md b/examples/inference/basic/V2_PORTING_STATUS.md index 2af71c391..f5e6b1211 100644 --- a/examples/inference/basic/V2_PORTING_STATUS.md +++ b/examples/inference/basic/V2_PORTING_STATUS.md @@ -18,6 +18,15 @@ some are blocked by the environment (gated weights, local-only weights, non-vide The v2 bring-up runs **single-GPU, resident, on the `TORCH_SDPA` backend** (no fastvideo-kernel / VSA / FP4), so distilled models that depend on VSA/SLA/FP4 kernels or non-flow-match samplers won't match. +(Wan2.2-A14B MoE is the exception to "resident": its two 14B experts are CPU-offloaded and swapped onto +the GPU one at a time at the boundary-timestep transition, so it fits a single 80GB GPU.) + +**Environment status (current):** the box was rescheduled from an aarch64 host to an x86_64 host +mid-session (`ipp1a1…` → `ipp2-0493`; `uname -m` = x86_64). The `.venv` is aarch64 wheels +(numpy/torch/fastvideo/CUDA) which cannot execute on x86 (numpy `_core` is aarch64; `uv` hits a +missing-aarch64-loader binfmt error). GPU verification is blocked until the box returns to aarch64 (the +existing venv then works as-is) or the venv is rebuilt for x86. Every model marked *verified* above was +GPU-confirmed **before** the flip; the ‡ models are code-complete, pending GPU re-verify. ## Working today (verified, real video) — committed on `v2` | Official example(s) | Model | v2 card | v2 example | @@ -26,19 +35,22 @@ FP4), so distilled models that depend on VSA/SLA/FP4 kernels or non-flow-match s | `basic_self_forcing_causal.py` | `wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers` | wan_causal | `v2_basic_self_forcing_causal.py`, `v2_basic_new_api.py` | | `basic_ltx2_distilled.py`, `basic_ltx2_distilled_fast_profile.py` | `FastVideo/LTX2-Distilled-Diffusers` | ltx2 — real `LTX2LatentUpsampler` between base/refine (un_normalize→learned 2× upsample→normalize) | `v2_basic_new_api.py` | | `basic_wan2_2_ti2v.py` (T2V branch) | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | wan2.2-ti2v | `v2_basic_wan2_2_ti2v.py` | +| `basic_wan2_2.py` | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` (MoE) | wan2.2-a14b + **CPU expert offload** (one 14B expert resident at a time, swapped at the boundary → 60GB peak) | `v2_basic_wan2_2.py` | +| `basic_ltx2.py` ‡ | `Davids048/LTX2-Base-Diffusers` | ltx2 base (single-stage, request-driven steps) | `v2_basic_ltx2.py` | +| `basic_ltx2_3_distilled.py` ‡ | `FastVideo/LTX-2.3-Distilled-Diffusers` | ltx2 base card (single-stage; pass few steps) | `v2_basic_ltx2_3_distilled.py` | \* `basic_mps.py` (Apple MPS) and `basic_ray.py` (ray executor) are device/executor variants of the same Wan2.1 model — the model itself is covered by wan21; those runtime modes are out of v2 scope. +‡ Code-complete + CPU-verified (cards/programs build, dispatch routes, schedule correct); GPU re-verify +pending the env (see **Environment status** below). ## Needs per-model work (architecture/sampler matches partially) | Official example(s) | Model | What v2 needs | |---|---|---| | `basic_dmd.py`, `basic_dmd_new_api.py` | FastWan2.1-T2V-1.3B (WanDMDPipeline) | **Blocked (env):** loading needs a *non-strict* load to skip the VSA-only `to_gate_compress` (root cause: the checkpoint key mis-maps under the generic Wan path, and `TransformerLoader` is strict except Cosmos2.5 — `component_loader.py:1005`). Even loaded, the DMD model is trained for VSA, which is **not built here (no nvcc)**, so SDPA output would be degraded. Needs the fastvideo-kernel VSA build + a non-strict WanDMD load. | | `basic_turbodiffusion*.py` | TurboWan2.1/2.2 (RCM) | **New sampler:** Reparameterized Consistency Model — v2 has no consistency loop (flow-match Euler won't match). | -| `basic_wan2_2.py` | Wan2.2-T2V-A14B (MoE) | **Card + MoE routing DONE + GPU-verified** (`build_wan22_a14b_card`: two `WanTransformer3DModel` experts + `BoundaryTimestepRouting` @0.875; both experts denoised correctly) but **OOMs in VAE decode on one 80GB GPU** (~70GB resident: 2×14B + UMT5). Upstream itself runs A14B with `num_gpus=2` + `dit_cpu_offload=True` ("DiT need to be offloaded for MoE"). Needs the resident path to offload an expert / free UMT5. `v2_basic_wan2_2.py` added. | -| `basic_wan2_2_i2v.py`, `basic_wan2_2_Fun.py` | Wan2.2-I2V / Fun | I2V needs image conditioning (+ A14B offload); Fun needs a control input. | -| `basic_ltx2.py` | `Davids048/LTX2-Base-Diffusers` | Base (non-distilled): base sigma schedule + more steps (the v2 ltx2 card hardcodes the distilled 8+3). | -| `basic_ltx2_3_distilled_i2v*.py` | LTX-2.3-Distilled I2V | Image conditioning on the ltx2 program (encode + condition). | +| `basic_wan2_2_i2v.py`, `basic_wan2_2_Fun.py` | Wan2.2-I2V / Fun | I2V needs an image-encode node + conditioning (the Wan adapter already accepts `encoder_hidden_states_image`; reuse the A14B offload); Fun needs a control input. | +| `basic_ltx2_3_distilled_i2v*.py` | LTX-2.3-Distilled I2V | Image conditioning on the ltx2 single-stage program (encode the input image + condition the denoise). | | `basic_lucy_edit.py` | `decart-ai/Lucy-Edit-Dev` | Wan-family video-edit: input-video conditioning. | ## Out of current scope — new families (kept as an upstream reference map only) diff --git a/examples/inference/basic/v2_basic_ltx2_3_distilled.py b/examples/inference/basic/v2_basic_ltx2_3_distilled.py new file mode 100644 index 000000000..e10b6e5db --- /dev/null +++ b/examples/inference/basic/v2_basic_ltx2_3_distilled.py @@ -0,0 +1,27 @@ +"""v2 port of basic_ltx2_3_distilled.py — LTX-2.3 Distilled (single-stage) through the v2 VideoGenerator. + +LTX-2.3-Distilled is single-stage (no spatial upsampler, unlike LTX-2.0 distilled's two-stage), so the +architecture-driven dispatch (v2/video_generator.py:_select_builders, via has_spatial_upsampler=False) +routes it to the SAME single-stage card as LTX-2 base (build_ltx2_base_card). Being distilled, it wants +FEW steps — pass a small num_inference_steps for the few-step schedule. + +NOTE: the v2 single-stage loop currently builds a request-driven *linspace* flow-match schedule +(base_flow_sigmas); the distilled-tuned few-step schedule is a follow-up. GPU re-verify of this example +is pending the env (the box was rescheduled aarch64->x86 mid-session — see V2_PORTING_STATUS.md). +""" +from v2 import VideoGenerator + +PROMPT = "a hummingbird hovering beside a red flower, wings blurred, macro, cinematic, highly detailed" + + +def main() -> None: + generator = VideoGenerator.from_pretrained("FastVideo/LTX-2.3-Distilled-Diffusers", num_gpus=1) + video = generator.generate_video( + prompt=PROMPT, output_path="v2_video_samples_ltx2_3", output_video_name="ltx2_3_hummingbird", + save_video=True, num_frames=9, height=512, width=768, num_inference_steps=8) + print(f"Output: {video.video_path}") + generator.shutdown() + + +if __name__ == "__main__": + main()