[optimization] Added Wan 2.2 VAE decode optimizations (#7056)

Signed-off-by: Maciej Bala <mbala@nvidia.com>
Signed-off-by: MaciejBalaNV <mbala@nvidia.com>

recipes/cosmos3/Cosmos3-Nano.md

@@ -129,6 +129,10 @@ To run **without** guardrails (you are responsible for license compliance),
 add `--no-guardrails` (no token/`cosmos-guardrail` needed). For extra GPUs use
 `--ulysses-degree N` (context parallel) or `--tensor-parallel-size N`;
 `--enable-layerwise-offload` reduces VRAM on smaller GPUs;
+`--vae-fast-path channels_last` speeds up the Wan VAE video decode by switching the
+decoder convolutions to channels-last kernels (output no longer bit-identical to
+diffusers; the default `lossless` fast path is bit-exact, see
+[Wan VAE Decoder Fast Path](../../docs/user_guide/diffusion/vae_fast_path.md));
 `--quantization fp8` (online, no calibration) cuts peak VRAM for 720p video
 generation from ~50 GB to ~36 GB with BF16-level quality (T2V composition can
 shift at the same seed). The pipeline

recipes/cosmos3/Cosmos3-Super.md

@@ -58,6 +58,10 @@ vllm serve nvidia/Cosmos3-Super \
 Guardrails are on by default (gated `nvidia/Cosmos-1.0-Guardrail` — `pip install
 cosmos-guardrail`, accept the license, set `HF_TOKEN`); add `--no-guardrails` to
 disable. `--enable-layerwise-offload` reduces VRAM on smaller GPUs;
+`--vae-fast-path channels_last` speeds up the Wan VAE video decode by switching the
+decoder convolutions to channels-last kernels (output no longer bit-identical to
+diffusers; the default `lossless` fast path is bit-exact, see
+[Wan VAE Decoder Fast Path](../../docs/user_guide/diffusion/vae_fast_path.md));
 `--quantization fp8` (online, no calibration) cuts peak VRAM for 720p video
 generation from ~83 GB to ~55 GB per GPU (2-GPU) with BF16-level quality (T2V
 composition can shift at the same seed).