Experimental jank multiGPU inference that's 2x faster than native somehow (#2100)

README.md

@@ -240,7 +240,7 @@ Optionally, you can use the following command-line flags:
 | `--wbits WBITS`           | Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported. |
 | `--model_type MODEL_TYPE` | Model type of pre-quantized model. Currently LLaMA, OPT, and GPT-J are supported. |
 | `--groupsize GROUPSIZE`   | Group size. |
-| `--pre_layer PRE_LAYER`   | The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. |
+| `--pre_layer PRE_LAYER [PRE_LAYER ...]`  | The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. For multi-gpu, write the numbers separated by spaces, eg `--pre_layer 30 60`. |
 | `--checkpoint CHECKPOINT` | The path to the quantized checkpoint file. If not specified, it will be automatically detected. |
 | `--monkey-patch`          | Apply the monkey patch for using LoRAs with quantized models.
 | `--quant_attn`         | (triton) Enable quant attention. |

docs/GPTQ-models-(4-bit-mode).md

@@ -107,6 +107,8 @@ This is the performance:
 Output generated in 123.79 seconds (1.61 tokens/s, 199 tokens)
 ```
 
+You can also use multiple GPUs with `pre_layer` if using the oobabooga fork of GPTQ, eg `--pre_layer 30 60` will load a LLaMA-30B model half onto your first GPU and half onto your second, or `--pre_layer 20 40` will load 20 layers onto GPU-0, 20 layers onto GPU-1, and 20 layers offloaded to CPU.
+
 ## Using LoRAs in 4-bit mode
 
 At the moment, this feature is not officially supported by the relevant libraries, but a patch exists and is supported by this web UI: https://github.com/johnsmith0031/alpaca_lora_4bit

modules/GPTQ_loader.py

@@ -172,7 +172,12 @@ def load_quantized(model_name):
 
     # qwopqwop200's offload
     if model_type == 'llama' and shared.args.pre_layer:
-        model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, shared.args.pre_layer)
+        if len(shared.args.pre_layer) == 1:
+            pre_layer = shared.args.pre_layer[0]
+        else:
+            pre_layer = shared.args.pre_layer
+
+        model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, pre_layer)
     else:
         threshold = False if model_type == 'gptj' else 128
         model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, kernel_switch_threshold=threshold)

modules/shared.py

@@ -130,7 +130,7 @@ parser.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layer
 parser.add_argument('--wbits', type=int, default=0, help='Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.')
 parser.add_argument('--model_type', type=str, help='Model type of pre-quantized model. Currently LLaMA, OPT, and GPT-J are supported.')
 parser.add_argument('--groupsize', type=int, default=-1, help='Group size.')
-parser.add_argument('--pre_layer', type=int, default=0, help='The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models.')
+parser.add_argument('--pre_layer', type=int, nargs="+", help='The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. For multi-gpu, write the numbers separated by spaces, eg --pre_layer 30 60.')
 parser.add_argument('--checkpoint', type=str, help='The path to the quantized checkpoint file. If not specified, it will be automatically detected.')
 parser.add_argument('--monkey-patch', action='store_true', help='Apply the monkey patch for using LoRAs with quantized models.')
 parser.add_argument('--quant_attn', action='store_true', help='(triton) Enable quant attention.')