Add Grace-Blackwell unified memory optimizations

Key changes based on HurbaLurba's DGX Spark research:

- Remove --gpu-only flag (fights unified memory fabric)
- Add --disable-pinned-memory, --force-fp16, --dont-upcast-attention
- Add CUDA env vars for unified memory: CUDA_MANAGED_FORCE_DEVICE_ALLOC,
  PYTORCH_NO_CUDA_MEMORY_CACHING, OMP_NUM_THREADS=20
- Document unified memory architecture best practices
- Add host-level GPU optimization instructions (clock locking, vboost)
- Document SageAttention PR #297 status (merged then reverted)
- Add credits section

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

docker-compose.yml

@@ -27,13 +27,26 @@ services:
 
     environment:
       COMFYUI_PORT: "${COMFYUI_PORT:-8188}"
-      COMFYUI_FLAGS: "${COMFYUI_FLAGS:---listen 0.0.0.0 --port 8188 --gpu-only}"
+      # Optimized for Grace-Blackwell unified memory architecture
+      # Key insight: DON'T use --gpu-only - let the unified memory fabric work naturally
+      COMFYUI_FLAGS: "${COMFYUI_FLAGS:---listen 0.0.0.0 --port 8188 --disable-pinned-memory --force-fp16 --fp16-unet --fp16-vae --fp16-text-enc --dont-upcast-attention}"
       NVIDIA_VISIBLE_DEVICES: "all"
       NVIDIA_DRIVER_CAPABILITIES: "compute,utility"
+
       # Disable torch.compile/inductor - Triton doesn't support Blackwell sm_121a yet
       TORCH_COMPILE_DISABLE: "1"
       TORCHDYNAMO_DISABLE: "1"
 
+      # Grace-Blackwell unified memory optimizations
+      CUDA_CACHE_DISABLE: "1"
+      PYTORCH_NO_CUDA_MEMORY_CACHING: "1"
+      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"
+      CUBLAS_WORKSPACE_CONFIG: ":0:0"
+
     volumes:
       # Models from existing ComfyUI install (read-only)
       - ${COMFYUI_HOST_PATH}/models:/opt/ComfyUI/models:ro