Split by rows instead of layers for llama.cpp multi-gpu (#5435)

modules/llamacpp_hf.py

@@ -216,7 +216,8 @@ class LlamacppHF(PreTrainedModel):
             'tensor_split': tensor_split_list,
             'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
             'logits_all': shared.args.logits_all,
-            'offload_kqv': not shared.args.no_offload_kqv
+            'offload_kqv': not shared.args.no_offload_kqv,
+            'split_mode': 1 if not shared.args.row_split else 2
         }
 
         Llama = llama_cpp_lib().Llama

modules/llamacpp_model.py

@@ -95,7 +95,8 @@ class LlamaCppModel:
             'rope_freq_base': RoPE.get_rope_freq_base(shared.args.alpha_value, shared.args.rope_freq_base),
             'tensor_split': tensor_split_list,
             'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
-            'offload_kqv': not shared.args.no_offload_kqv
+            'offload_kqv': not shared.args.no_offload_kqv,
+            'split_mode': 1 if not shared.args.row_split else 2
         }
 
         result.model = Llama(**params)

modules/loaders.py

@@ -44,6 +44,7 @@ loaders_and_params = OrderedDict({
         'cpu',
         'numa',
         'no_offload_kqv',
+        'row_split',
         'tensorcores',
     ],
     'llamacpp_HF': [
@@ -66,6 +67,7 @@ loaders_and_params = OrderedDict({
         'no_use_fast',
         'logits_all',
         'no_offload_kqv',
+        'row_split',
         'tensorcores',
         'llamacpp_HF_info',
     ],

modules/shared.py

@@ -129,6 +129,7 @@ group.add_argument('--numa', action='store_true', help='Activate NUMA task alloc
 group.add_argument('--logits_all', action='store_true', help='Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower.')
 group.add_argument('--no_offload_kqv', action='store_true', help='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')
 group.add_argument('--cache-capacity', type=str, help='Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.')
+group.add_argument('--row_split', action='store_true', help='Split multi-gpu by row instead of layer. Faster on some cards.')
 
 # ExLlamaV2
 group = parser.add_argument_group('ExLlamaV2')

modules/ui.py

@@ -93,6 +93,7 @@ def list_model_elements():
         'numa',
         'logits_all',
         'no_offload_kqv',
+        'row_split',
         'tensorcores',
         'hqq_backend',
     ]

modules/ui_model_menu.py

@@ -107,6 +107,7 @@ def create_ui():
                         with gr.Column():
                             shared.gradio['tensorcores'] = gr.Checkbox(label="tensorcores", value=shared.args.tensorcores, info='Use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards. NVIDIA only.')
                             shared.gradio['no_offload_kqv'] = gr.Checkbox(label="no_offload_kqv", value=shared.args.no_offload_kqv, info='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')
+                            shared.gradio['row_split'] = gr.Checkbox(label="row_split", value=shared.args.row_split, info='Split model by rows across GPUs. Improves performance on some cards.')
                             shared.gradio['triton'] = gr.Checkbox(label="triton", value=shared.args.triton)
                             shared.gradio['no_inject_fused_attention'] = gr.Checkbox(label="no_inject_fused_attention", value=shared.args.no_inject_fused_attention, info='Disable fused attention. Fused attention improves inference performance but uses more VRAM. Fuses layers for AutoAWQ. Disable if running low on VRAM.')
                             shared.gradio['no_inject_fused_mlp'] = gr.Checkbox(label="no_inject_fused_mlp", value=shared.args.no_inject_fused_mlp, info='Affects Triton only. Disable fused MLP. Fused MLP improves performance but uses more VRAM. Disable if running low on VRAM.')