Add sglang-inspired low-latency NCCL ring buffer and CUDA connection optimization flags to start-server.sh

scripts/start-server.sh

@@ -1,20 +1,19 @@
-#!/usr/bin/env bash
+#!/bin/bash
-# ==============================================================================
+# llama-server launcher for Qwen3.8-27B on 3x Turing GPUs (RTX 2060 12GB + 2x CMP 50HX 10GB)
-# Production startup script for llama-server with NCCL tensor parallelism
+# Built with NVIDIA NCCL for hardware-accelerated multi-GPU tensor parallelism.
-# Supports: 256K Context Window, Multimodal Vision Projector, Flash Attention,
+export LD_LIBRARY_PATH=/opt/llama.cpp/build-nccl/bin:/opt/minicpm-venv/lib/python3.13/site-packages/nvidia/nccl/lib:/lib/x86_64-linux-gnu
-#           q4_0 KV Cache quantization, and Jinja reasoning control.
-# ==============================================================================
-set -euo pipefail
 
-MODEL_PATH="${MODEL_PATH:-/opt/models/gguf/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf}"
+# Low-Latency NCCL & CUDA Driver Pipeline Optimizations
-MMPROJ_PATH="${MMPROJ_PATH:-/opt/models/gguf/mmproj-Qwen3.8-27B-Uncensored-f16.gguf}"
+export CUDA_DEVICE_MAX_CONNECTIONS=1
-BINARY="${LLAMA_BINARY:-/opt/llama.cpp/build-nccl/bin/llama-server}"
+export NCCL_BUFFSIZE=2097152
+export NCCL_NET_GDR_LEVEL=0
+export NCCL_P2P_DISABLE=0
+export NCCL_ALGO=RING
+export NCCL_PROTO=SIMPLE
 
-export LD_LIBRARY_PATH="/opt/llama.cpp/build-nccl/bin:/opt/minicpm-venv/lib/python3.13/site-packages/nvidia/nccl/lib:/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH:-}"
+exec /opt/llama.cpp/build-nccl/bin/llama-server \
-
+  -m /opt/models/gguf/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
-exec "$BINARY" \
+  --mmproj /opt/models/gguf/mmproj-Qwen3.8-27B-Uncensored-f16.gguf \
-  -m "$MODEL_PATH" \
-  --mmproj "$MMPROJ_PATH" \
   -ngl 99 \
   -c 262144 \
   --parallel 1 \
@@ -28,4 +27,4 @@ exec "$BINARY" \
   --cache-type-k q4_0 \
   --cache-type-v q4_0 \
   --host 0.0.0.0 \
-  --port 8080
+  --port 8080