add --split-output-tensor / -sot CLI parameter (port from custom_neo)

Adds a CLI flag to force-split the output tensor (output.weight,
output_extra.weight) across GPUs when using split mode graph. When
enabled, the output tensor is allocated on buft_matrix (split-capable
buffer type) instead of buft_output, allowing it to participate in
tensor parallelism. The output norm tensors remain on buft_output
(except gemma4/gemma4_mtp, which also split output_norm).

- common/common.h: split_output_tensor field in gpt_params
- common/common.cpp: -sot / --split-output-tensor CLI arg, help, yaml
- include/llama.h: split_output_tensor field in llama_model_params
- src/llama-model.h: split_output_tensor flag in llama_model
- src/llama.cpp: pass split_output_tensor to llm_load_tensors
- src/llama-load-tensors.cpp:
  * LOADING_PRELUDE sets ctx_output_split conditionally
  * get_context_for_tensor skips -ot overrides for output tensor names
  * create_embd_output / create_default_embd_output use ctx_output_split
  * All architecture tensor functions use ctx_output_split for output
  * Single centralized log when split_output_tensor = ON

Author: Nexesenex

common/common.cpp

@@ -2177,6 +2177,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.split_mode_graph_scheduling = true;
         return true;
     }
+    if (arg == "-sot" || arg == "--split-output-tensor") {
+        params.split_output_tensor = true;
+        return true;
+    }
     if (arg == "-sas" || arg == "--scheduler-async") {
         params.scheduler_async = true;
         return true;
@@ -3029,6 +3033,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "*",         "-grt, --graph-reduce-type",       "Type for data exchange between GPUs (default: %s)", "f32"});
     options.push_back({ "*",         "-gap, --graph-attn-precision",    "Flash-attn precision under -sm graph (default: %s)", "f16"});
     options.push_back({ "*",         "-smgs, --split-mode-graph-scheduling,", "Force Split Mode Graph Scheduling (default: %d)", params.split_mode_graph_scheduling});
+    options.push_back({ "*",         "-sot, --split-output-tensor,",    "Force split of the Output Tensor in Split Mode Graph (default: %d)", params.split_output_tensor});
     options.push_back({ "*",         "-sas,  --scheduler_async,",       "Async evaluation of compute graphs: %d)", params.scheduler_async});
     options.push_back({ "*",         "-vq, --validate-quants",          "validate quantized data while loading the model (default: %d)", params.validate_quants});
     options.push_back({ "*",           "-p,    --prompt PROMPT",        "prompt to start generation with\n"
@@ -4171,6 +4176,7 @@ struct llama_model_params common_model_params_to_llama(const gpt_params & params
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.repack_tensors  = params.repack_tensors;
+    mparams.split_output_tensor = params.split_output_tensor;
     mparams.use_thp         = params.use_thp;
     mparams.validate_quants = params.validate_quants;
     mparams.merge_qkv       = params.merge_qkv;
@@ -5301,6 +5307,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "k_cache_hadamard: %s # default: false\n", params.k_cache_hadamard ? "true" : "false");
     fprintf(stream, "v_cache_hadamard: %s # default: false\n", params.v_cache_hadamard ? "true" : "false");
     fprintf(stream, "split_mode_graph_scheduling: %s # default: false\n", params.split_mode_graph_scheduling ? "true" : "false");
+    fprintf(stream, "split_output_tensor: %s # default: false\n", params.split_output_tensor ? "true" : "false");
     //fprintf(stream, "split_mode_f16: %s # default: true\n", params.split_mode_f16 ? "true" : "false");
     fprintf(stream, "reduce_type: %s # default f16\n", params.reduce_type.c_str());
     fprintf(stream, "scheduler_async: %s # default: false\n", params.scheduler_async ? "true" : "false");

common/common.h

@@ -444,6 +444,7 @@ struct gpt_params {
     bool v_cache_hadamard  = false; // if true, use Hadamard transform for the V-cache (only makes sense with quantized cache, which requires FA)
     bool split_mode_graph_scheduling = false; // if true, force split mode graph scheduling
     //bool split_mode_f16    = true;  // if true, intermediate results will be cast to f16 before copying to other GPUs to perform reduce ops
+    bool split_output_tensor = false; // if true, force split the output tensor in split mode graph
     bool scheduler_async   = false; // if true, in split mode graph the scheduler will use multiple threads to evaluate the graph
     int  fused_delta_net   = 0;     // use fused delta-net if number of tokens in the batch is less than this value
     bool has_mtp           = false; // enable MTP if supported by the model

include/llama.h

@@ -429,6 +429,7 @@ extern "C" {
         bool merge_up_gate_exps;  // if true, merge ffn_up_exps and ffn_gate_exps tensors into a single, contiguous tensor
         bool mtp;           // if true, load MTP layers if present
         bool dry_run;       // skip loading tensors
+        bool split_output_tensor; // if true, force split the output tensor in split mode graph
         bool flash_attn;
         bool defer_experts;    // defer expert mmap residency to speed up model loading (Linux only)
     };