optimize: Implement adaptive GPU cache clearing for BlockSwap efficiency

BlockSwap now uses minimal, pressure-based cache clearing instead of forced deep cleans:
- Clear only when VRAM < 5% free (was 15%, more aggressive threshold)
- Use minimal GPU cache clear (deep=False) for reduced overhead
- Applied after each block/IO swap when pressure detected

Also fixes peft dependency: Add explicit peft>=0.17.0 in requirements for diffusers>=0.33.1 compatibility in ComfyUI ecosystem where other plugins may install outdated peft versions.

pyproject.toml

@@ -21,6 +21,7 @@ dependencies = [
     "einops",
     "omegaconf>=2.3.0",
     "diffusers>=0.33.1",
+    "peft>=0.17.0",
     "rotary_embedding_torch>=0.5.3",
     "opencv-python",
     "gguf",

requirements.txt

@@ -7,6 +7,7 @@ psutil
 einops
 omegaconf>=2.3.0
 diffusers>=0.33.1
+peft>=0.17.0
 rotary_embedding_torch>=0.5.3
 opencv-python
 gguf

src/optimization/blockswap.py

@@ -444,7 +444,7 @@ def _wrap_block_forward(
             _log_swap_timing(debug, t_start, self._block_idx, "block")
 
             # Only clear cache under memory pressure
-            clear_memory(debug=debug, deep=True, force=False, timer_name="wrap_block_forward")
+            clear_memory(debug=debug, deep=False, force=False, timer_name="wrap_block_forward")
         else:
             output = original_forward(*args, **kwargs)
 
@@ -523,7 +523,7 @@ def _wrap_io_forward(
         _log_swap_timing(debug, t_start, self._module_name, "I/O")
 
         # Only clear cache under memory pressure
-        clear_memory(debug=debug, deep=True, force=False, timer_name="wrap_block_forward")
+        clear_memory(debug=debug, deep=False, force=False, timer_name="wrap_block_forward")
 
         return output
     

src/optimization/memory_manager.py

@@ -191,7 +191,7 @@ def clear_memory(debug: Optional['Debug'] = None, deep: bool = False, force: boo
     
     Args:
         debug: Debug instance for logging (optional)
-        force: If True, always clear. If False, only clear when <15% free
+        force: If True, always clear. If False, only clear when <5% free
         deep: If True, perform deep cleanup including GC and OS operations.
               If False (default), only perform minimal GPU cache clearing.
         timer_name: Optional suffix for timer names to make them unique per invocation
@@ -230,9 +230,9 @@ def clear_memory(debug: Optional['Debug'] = None, deep: bool = False, force: boo
         mem_info = get_basic_vram_info(device=None)
         
         if "error" not in mem_info:
-            # Check VRAM/MPS memory pressure (15% free threshold)
+            # Check VRAM/MPS memory pressure (5% free threshold)
             free_ratio = mem_info["free_gb"] / mem_info["total_gb"]
-            if free_ratio < 0.15:
+            if free_ratio < 0.05:
                 should_clear = True
                 if debug:
                     backend = "MPS" if torch.mps.is_available() else "VRAM"
@@ -241,7 +241,7 @@ def clear_memory(debug: Optional['Debug'] = None, deep: bool = False, force: boo
         # For non-MPS systems, also check system RAM separately
         if not should_clear and not torch.mps.is_available():
             mem = psutil.virtual_memory()
-            if mem.available < mem.total * 0.15:
+            if mem.available < mem.total * 0.05:
                 should_clear = True
                 if debug:
                     debug.log(f"RAM pressure: {mem.available/(1024**3):.2f}GB free of {mem.total/(1024**3):.2f}GB", category="memory")