GDS offloader

README.md

@@ -365,6 +365,14 @@ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app w
 > Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above.
 <br/><br/>If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal.
 
+## How to run heavy workflow on mid range GPU (NVIDIA-Linux)?
+
+Use the `--enable-gds` flag to activate NVIDIA [GPUDirect Storage](https://docs.nvidia.com/gpudirect-storage/) (GDS), which allows data to be transferred directly between SSDs and GPUs. This eliminates traditional CPU-mediated data paths, significantly reducing I/O latency and CPU overhead. System RAM will still be utilized for caching to further optimize performance, along with SSD.
+
+This feature is tested on NVIDIA GPUs on Linux based system only.
+
+Requires: `cupy-cuda12x>=12.0.0`, `pynvml>=11.4.1`, `cudf>=23.0.0`, `numba>=0.57.0`, `nvidia-ml-py>=12.0.0`.
+
 ## Support and dev channel
 
 [Discord](https://comfy.org/discord): Try the #help or #feedback channels.

comfy/cli_args.py

@@ -139,6 +139,17 @@ class LatentPreviewMethod(enum.Enum):
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
 parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
 
+# GPUDirect Storage (GDS) arguments
+gds_group = parser.add_argument_group('gds', 'GPUDirect Storage options for direct SSD-to-GPU model loading')
+gds_group.add_argument("--enable-gds", action="store_true", help="Enable GPUDirect Storage for direct SSD-to-GPU model loading (requires CUDA 11.4+, cuFile).")
+gds_group.add_argument("--disable-gds", action="store_true", help="Explicitly disable GPUDirect Storage.")
+gds_group.add_argument("--gds-min-file-size", type=int, default=100, help="Minimum file size in MB to use GDS (default: 100MB).")
+gds_group.add_argument("--gds-chunk-size", type=int, default=64, help="GDS transfer chunk size in MB (default: 64MB).")
+gds_group.add_argument("--gds-streams", type=int, default=4, help="Number of CUDA streams for GDS operations (default: 4).")
+gds_group.add_argument("--gds-prefetch", action="store_true", help="Enable GDS prefetching for better performance.")
+gds_group.add_argument("--gds-no-fallback", action="store_true", help="Disable fallback to CPU loading if GDS fails.")
+gds_group.add_argument("--gds-stats", action="store_true", help="Print GDS statistics on exit.")
+
 class PerformanceFeature(enum.Enum):
     Fp16Accumulation = "fp16_accumulation"
     Fp8MatrixMultiplication = "fp8_matrix_mult"