feat: add regione gamma related dit cache changes.

Author: chuanhaoyu

xllm/core/common/global_flags.h

@@ -362,6 +362,10 @@ DECLARE_int64(dit_cache_end_blocks);
 
 DECLARE_string(dit_regione_refresh_steps);
 DECLARE_double(dit_regione_region_threshold);
+DECLARE_double(dit_regione_cache_threshold);
+DECLARE_bool(dit_regione_use_avd_gamma);
+DECLARE_bool(dit_regione_erosion_dilation);
+DECLARE_bool(dit_regione_profile);
 
 DECLARE_bool(dit_sp_communication_overlap);
 

xllm/core/framework/config/dit_config.cpp

@@ -61,6 +61,28 @@ DEFINE_double(dit_regione_region_threshold,
               0.80,
               "RegionE: cosine threshold for adaptive region partition.");
 
+DEFINE_double(dit_regione_cache_threshold,
+              0.03,
+              "RegionE: AVDCache error threshold δ (paper Eq.8). "
+              "Reuse velocity while 1-accumulate <= threshold. "
+              "Qwen-Image-Edit default in RegionE inplace.py is 0.03.");
+
+DEFINE_bool(dit_regione_use_avd_gamma,
+            true,
+            "RegionE: use AVDCache with gamma (paper/inplace.py method). "
+            "Uses the original diffusers 28-step gamma curve, linearly "
+            "upsampled/downsampled to the actual inference step count. "
+            "Set false to use fixed skip_interval instead.");
+
+DEFINE_bool(dit_regione_erosion_dilation,
+            true,
+            "RegionE: enable erosion/dilation for region mask cleanup.");
+
+DEFINE_bool(dit_regione_profile,
+            false,
+            "RegionE: print per-step timing breakdown for partial/full DiT and "
+            "K/V CPU offload.");
+
 DEFINE_bool(dit_sp_communication_overlap,
             true,
             "Communication & Computation overlap for sequence parallel");
@@ -141,6 +163,10 @@ void DiTConfig::from_flags() {
   XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_cache_end_blocks);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_regione_refresh_steps);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_regione_region_threshold);
+  XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_regione_cache_threshold);
+  XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_regione_use_avd_gamma);
+  XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_regione_erosion_dilation);
+  XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_regione_profile);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_sp_communication_overlap);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_debug_print);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(dit_laser_attention_enabled);
@@ -169,6 +195,10 @@ void DiTConfig::from_json(const JsonReader& json) {
   XLLM_CONFIG_ASSIGN_FROM_JSON(dit_cache_end_blocks);
   XLLM_CONFIG_ASSIGN_FROM_JSON(dit_regione_refresh_steps);
   XLLM_CONFIG_ASSIGN_FROM_JSON(dit_regione_region_threshold);
+  XLLM_CONFIG_ASSIGN_FROM_JSON(dit_regione_cache_threshold);
+  XLLM_CONFIG_ASSIGN_FROM_JSON(dit_regione_use_avd_gamma);
+  XLLM_CONFIG_ASSIGN_FROM_JSON(dit_regione_erosion_dilation);
+  XLLM_CONFIG_ASSIGN_FROM_JSON(dit_regione_profile);
   XLLM_CONFIG_ASSIGN_FROM_JSON(dit_sp_communication_overlap);
   XLLM_CONFIG_ASSIGN_FROM_JSON(dit_debug_print);
   XLLM_CONFIG_ASSIGN_FROM_JSON(dit_laser_attention_enabled);
@@ -210,6 +240,14 @@ void DiTConfig::append_config_json(nlohmann::ordered_json& config_json) const {
       config_json, default_config, dit_regione_refresh_steps);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
       config_json, default_config, dit_regione_region_threshold);
+  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
+      config_json, default_config, dit_regione_cache_threshold);
+  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
+      config_json, default_config, dit_regione_use_avd_gamma);
+  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
+      config_json, default_config, dit_regione_erosion_dilation);
+  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
+      config_json, default_config, dit_regione_profile);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
       config_json, default_config, dit_sp_communication_overlap);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(

xllm/core/framework/config/dit_config.h

@@ -53,6 +53,10 @@ class DiTConfig final {
          "dit_cache_end_blocks",
          "dit_regione_refresh_steps",
          "dit_regione_region_threshold",
+         "dit_regione_cache_threshold",
+         "dit_regione_use_avd_gamma",
+         "dit_regione_erosion_dilation",
+         "dit_regione_profile",
          "dit_sp_communication_overlap",
          "dit_debug_print",
          "dit_laser_attention_enabled",
@@ -93,6 +97,14 @@ class DiTConfig final {
 
   PROPERTY(double, dit_regione_region_threshold) = 0.80;
 
+  PROPERTY(double, dit_regione_cache_threshold) = 0.03;
+
+  PROPERTY(bool, dit_regione_use_avd_gamma) = true;
+
+  PROPERTY(bool, dit_regione_erosion_dilation) = true;
+
+  PROPERTY(bool, dit_regione_profile) = false;
+
   PROPERTY(bool, dit_sp_communication_overlap) = true;
 
   PROPERTY(bool, dit_debug_print) = false;

xllm/core/framework/dit_cache/dit_cache_config.h

@@ -57,10 +57,20 @@ struct FBCacheTaylorSeerOptions : public DiTBaseCacheOptions {
 };
 
 struct RegionEOptions : public DiTBaseCacheOptions {
+  // Fallback fixed-interval AVD when gamma is disabled or step count
+  // mismatches.
   int64_t skip_interval_steps = 3;
   int64_t tail_steps = 1;
   std::vector<int64_t> refresh_steps = {16};
   float region_threshold = 0.80f;
+  // AVDCache δ in paper Eq.8/9 / inplace.py cache_threshold (Qwen default
+  // 0.03).
+  float cache_threshold = 0.03f;
+  // Use fitted γ_t AVDCache (paper) instead of fixed skip_interval.
+  bool use_avd_gamma = true;
+  // Enable erosion/dilation morphological cleanup after ARP mask selection.
+  bool erosion_dilation = true;
+  bool profile = false;
 };
 
 struct ResidualCacheOptions {