feat: add KV caching support for Wan models

Author: Perseus14

src/maxdiffusion/configs/base_wan_14b.yml

@@ -348,6 +348,7 @@ flow_shift: 3.0
 # Skips the unconditional forward pass on ~35% of steps via residual compensation.
 # See: FasterCache (Lv et al. 2024), WAN 2.1 paper §4.4.2
 use_cfg_cache: False
+use_kv_cache: False
 use_magcache: False
 magcache_thresh: 0.12
 magcache_K: 2

src/maxdiffusion/configs/base_wan_1_3b.yml

@@ -301,6 +301,7 @@ flow_shift: 3.0
 
 # Diffusion CFG cache (FasterCache-style, WAN 2.1 T2V only)
 use_cfg_cache: False
+use_kv_cache: False
 
 # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
 guidance_rescale: 0.0

src/maxdiffusion/configs/base_wan_27b.yml

@@ -323,6 +323,7 @@ boundary_ratio: 0.875
 
 # Diffusion CFG cache (FasterCache-style)
 use_cfg_cache: False
+use_kv_cache: False
 # SenCache: Sensitivity-Aware Caching (arXiv:2602.24208) — skip forward pass
 # when predicted output change (based on accumulated latent/timestep drift) is small
 use_sen_cache: False

src/maxdiffusion/configs/base_wan_i2v_14b.yml

@@ -306,6 +306,7 @@ flow_shift: 5.0
 
 # Diffusion CFG cache (FasterCache-style)
 use_cfg_cache: False
+use_kv_cache: False
 # SenCache: Sensitivity-Aware Caching (arXiv:2602.24208)
 use_sen_cache: False
 use_magcache: False

src/maxdiffusion/configs/base_wan_i2v_27b.yml

@@ -318,6 +318,7 @@ boundary_ratio: 0.875
 
 # Diffusion CFG cache (FasterCache-style)
 use_cfg_cache: False
+use_kv_cache: False
 # SenCache: Sensitivity-Aware Caching (arXiv:2602.24208)
 use_sen_cache: False