perf: cache MiniMax H3 text conditioning (#1966)

src/pipeline/diffusion_engine.cpp

@@ -1762,6 +1762,8 @@ bool StableDiffusionGGML::apply_loras(const sd_lora_t* loras, uint32_t lora_coun
         extension->collect_loras(all_loras);
     }
 
+    conditioning_cache_allowed_ = all_loras.empty();
+
     int64_t t0 = ggml_time_ms();
     end_runners();
     clear_lora_adapters();

src/pipeline/diffusion_engine.h

@@ -178,6 +178,7 @@ public:
     std::recursive_mutex execution_mutex;
     std::unique_ptr<ModelConfig> config_;
     RunnerState runner_state_;
+    bool conditioning_cache_allowed_ = false;
     bool executing_ = false;
 
     std::shared_ptr<Denoiser> denoiser;

src/pipeline/video.cpp

@@ -1156,6 +1156,10 @@ namespace sd::pipeline {
         condition_params.zero_out_masked       = true;
         condition_params.ref_images            = &latents.ref_images;
         condition_params.minimax_h3_references = &latents.minimax_presentation_refs;
+        condition_params.allow_cache =
+            sd_version_is_minimax_h3(sd->version) &&
+            sd->conditioning_cache_allowed_ &&
+            !request.use_uncond;
         if (sd_version_is_lingbot_video(sd->version) || sd_version_is_minimax_h3(sd->version)) {
             condition_params.ref_image_params.vlm_resize_mode = RefImageResizeMode::AREA;
         }