feat: NORMAL_VRAM + AIMDO + copy=False patch + kernel caching

Major unified memory optimization changes:

1. model_management.py: HIGH_VRAM → NORMAL_VRAM
   - GB10 unified memory: offloading to CPU doesn't save physical RAM
     (same pool), but NORMAL_VRAM allows per-layer partial loading when
     memory is tight instead of all-or-nothing OOM
   - text_encoder_offload_device() and vae_offload_device() now return
     CPU (allows ComfyUI to offload unused models)
   - intermediate_device() still returns GPU (VAE outputs must stay in
     CUDA allocator for honest memory tracking)
   - User can force HIGH_VRAM with --highvram if models fit

2. utils.py: copy=True → copy=False for tensor.to(device)
   - On GB10 unified memory, copy=True creates a full duplicate in both
     CPU and CUDA allocators simultaneously (ComfyUI issue #10896)
   - copy=False makes .to(device) a zero-copy device label change since
     both allocators draw from the same physical LPDDR5X
   - Halves model loading memory usage when --disable-mmap is set

3. Removed --disable-dynamic-vram from ComfyUI flags
   - Was preventing AIMDO (comfy_aimdo) from initializing
   - AIMDO now activates: VBAR-based page-level VRAM management at 32MB
     granularity instead of blunt .to(cpu) copies
   - Falls back to NORMAL_VRAM per-layer loading if AIMDO has issues

4. Added CUDA_CACHE_MAXSIZE=4294967296 (4GB kernel cache)
   - PTX→SASS kernel caching for sm_121 (GB10 Blackwell)
   - 3x speedup on subsequent runs reported by DGX Spark community

5. System: vm.swappiness reduced from 60 to 1
   - Swap thrashing on unified memory causes silent system freezes
   - Near-zero swappiness ensures clean OOM kills instead

docker-compose.yml

@@ -37,18 +37,18 @@ services:
       TORCH_COMPILE_DISABLE: "1"
       TORCHDYNAMO_DISABLE: "1"
 
-      # Grace-Blackwell unified memory optimizations
-      CUDA_CACHE_DISABLE: "1"
-      # PYTORCH_NO_CUDA_MEMORY_CACHING removed — our model_management patch handles
-      # caching properly by skipping empty_cache() on unified memory instead of disabling
-      # PyTorch's allocator entirely. Keeping caching ON reduces allocation overhead.
-      CUDA_DEVICE_MAX_CONNECTIONS: "1"
-      CUDA_DEVICE_MAX_COPY_CONNECTIONS: "4"
-      CUDA_MODULE_LOADING: "EAGER"
-      CUDA_MANAGED_FORCE_DEVICE_ALLOC: "1"
-      OMP_NUM_THREADS: "20"
+      # Grace-Blackwell unified memory — removed aggressive CUDA tuning (5/21):
+      # CUDA_CACHE_DISABLE, CUDA_DEVICE_MAX_CONNECTIONS, CUDA_DEVICE_MAX_COPY_CONNECTIONS,
+      # CUDA_MODULE_LOADING=EAGER, CUDA_MANAGED_FORCE_DEVICE_ALLOC, OMP_NUM_THREADS
+      # These were over-tuning. The ComfyUI flags + Sparky patch handle the architecture.
+      # Keeping only CUBLAS_WORKSPACE_CONFIG for determinism.
       CUBLAS_WORKSPACE_CONFIG: ":0:0"
 
+      # CUDA kernel caching — PTX→SASS compilation cache for GB10 (sm_121)
+      # First run compiles kernels, subsequent runs reuse from disk. 3x speedup reported.
+      # 4GB cache covers all typical ComfyUI kernel variants.
+      CUDA_CACHE_MAXSIZE: "4294967296"
+
     volumes:
       # Models from existing ComfyUI install (read-only)
       - ${COMFYUI_HOST_PATH}/models:/opt/ComfyUI/models:ro
@@ -66,8 +66,12 @@ services:
       - ${SPARKYUI_DATA_PATH}/wheels:/opt/wheels
 
       # Sparky patches - Grace-Blackwell unified memory optimizations
-      # This overrides ComfyUI's model_management.py with our patched version
+      # model_management.py: HIGH_VRAM→NORMAL_VRAM, intermediate_device()→cuda, soft_empty_cache skip,
+      #   95% vram_for_weights, UNIFIED_MEMORY detection, offload devices → cuda
+      # utils.py: copy=False on tensor.to(device) — avoids double-allocation on unified memory
+      #   where CPU and GPU share the same physical RAM (ComfyUI issue #10896)
       - ./patches/model_management.py:/opt/ComfyUI/comfy/model_management.py:ro
+      - ./patches/utils.py:/opt/ComfyUI/comfy/utils.py:ro
 
     networks:
       - sparky_net