Merge pull request #109 from newfla/feat_auto_detect_max_vram

feat: update max_vram option to auto-detect free VRAM minus 1 GiB

Author: newfla

cli/README.md

@@ -40,7 +40,7 @@ Options:
       --low-vram
           Enable optimization to use less VRAM: clip_on_cpu, vae tiling, flash_attention, offload_params_to_cpu
       --max-vram <MAX_VRAM>
-          Maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting [default: 0]
+          Maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; -1 auto-detects free VRAM minus 1 GiB [default: -1]
   -s, --seed <SEED>
           RNG seed (-1 --> random) [default: -1]
   -h, --help

cli/src/main.rs

@@ -119,7 +119,7 @@ struct Args {
     low_vram: bool,
 
     /// Maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting
-    #[arg(long, default_value_t = 0.0)]
+    #[arg(long, default_value_t = -1.0)]
     max_vram: f32,
 
     /// RNG seed (-1 --> random)

src/api.rs

@@ -281,8 +281,8 @@ pub struct ModelConfig {
     #[builder(default = "false")]
     offload_params_to_cpu: bool,
 
-    /// Maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting
-    #[builder(default = "0.0")]
+    /// Maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; -1 auto-detects free VRAM minus 1 GiB
+    #[builder(default = "-1.0")]
     max_vram: f32,
 
     /// Path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now