feat: add qwopus 3.6 35b a3b v1 configuration

Author: cnsiva

scripts-local/qwopus3.6-35b-a3b-v1.conf

@@ -0,0 +1,63 @@
+# ==============================================================================
+# Model: Qwopus3.6-35B-A3B-v1 (APEX-I-Balanced)
+# Optimized for: 16GB RTX 4070 Ti Super | AMD 7950X3D
+# Architecture: Hybrid MoE (256 experts, 8 active) + Gated DeltaNet
+# ==============================================================================
+
+SERVICE_NAME="llama-qwopus.service"
+MODEL_PATH="/home/siva/models/Qwen3.6-35B-A3B/mudler/Qwopus3.6-35B-A3B-v1-APEX-I-Balanced.gguf"
+MMPRJ_PATH=""
+MODEL_ALIAS="qwopus-3.6-35b-v1"
+
+# --- COMPUTE (Mixed Affinity per GEMINI.md) ---
+# Cross-CCD string for maximum single-model burst performance
+CPU_AFFINITY="0-7,24-31"
+THREADS=16
+THREADS_BATCH=16
+N_GPU_LAYERS=999
+# N_CPU_MOE=32: Final calibration for stable dual-model execution.
+N_CPU_MOE=36
+BATCH_SIZE=2048
+UBATCH_SIZE=1024
+MLOCK=true
+FLASH_ATTN=on
+NUMA=isolate
+MMPRJ_OFFLOAD=false
+LOG_DISABLE=false
+
+# --- MEMORY (Hybrid Precision Mandate) ---
+# Q8_0 Keys (precision for DeltaNet) and Q4_0 Values (VRAM savings)
+CACHE_TYPE_K="q8_0"
+CACHE_TYPE_V="q4_0"
+CTX_SIZE=262144 # 256k context limit
+PARALLEL=2      # Focused single-stream reasoning
+PORT=8082
+
+# --- INFRASTRUCTURE ---
+JINJA=true
+JINJA_KWARGS='{"preserve_thinking": true}'
+
+# --- SAMPLING (APEX-I Logic/Coder Chain - Optimized per Research) ---
+# Temperature: 0.75 prevents "Expert Collapse" by ensuring router diversity.
+#TEMP=0.75
+# Min-P: 0.10 is the recommended production threshold for PRISM-Balanced.
+#MIN_P=0.10
+#TOP_P=0.95
+#TOP_K=20
+#REPEAT_PENALTY=1.05
+#REPEAT_LAST_N=256.0
+# --- REFINED DRY SAMPLER (Coding/MoE Focus) ---
+#DRY_MULTIPLIER=0.8
+#DRY_BASE=1.75
+#DRY_ALLOWED_LENGTH=2
+#DRY_PENALTY_LAST_N=4096
+#SAMPLERS="dry;top_k;min_p"
+
+# --- REASONING ---
+REASONING="on"
+REASONING_FORMAT="auto"
+REASONING_BUDGET=-1
+REASONING_BUDGET_MESSAGE=" [Logic Finalized] "
+
+# --- UI/UX STANDARDS ---
+EXTRA_ARGS=""