gemma-4 config update

Author: cnsiva

scripts-local/gemma-4-26b-a4b-qat.conf

@@ -6,20 +6,18 @@
 # ==============================================================================
 
 SERVICE_NAME="llama-gemma-4-26b-qat.service"
-# MODEL_PATH="/home/siva/models/gemma-4-26b-a4b-qat/unsloth/gemma-4-26b-a4b-qat/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf"
-# MMPRJ_PATH="/home/siva/models/gemma-4-26b-a4b-qat/unsloth/gemma-4-26b-a4b-qat/mmproj-F16.gguf"
- MODEL_PATH="/home/siva/models/HauhauCS/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf"
- MMPRJ_PATH="/home/siva/models/HauhauCS/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP/mmproj-Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf"
+MODEL_PATH="/home/siva/models/gemma-4-26b-a4b-qat/unsloth/gemma-4-26b-a4b-qat/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf"
+MMPRJ_PATH="/home/siva/models/gemma-4-26b-a4b-qat/unsloth/gemma-4-26b-a4b-qat/mmproj-F16.gguf"
 MODEL_ALIAS="Gemma4-26B"
 
 # --- COMPUTE ---
 # Cross-CCD string for maximum single-model burst performance
-CPU_AFFINITY="0-7"  # 16 threads on CCD0 and CCD2 (AMD 7950X3D)
-THREADS=8
-THREADS_BATCH=8
+CPU_AFFINITY="0-7, 24-31"  # 16 threads on CCD0 and CCD2 (AMD 7950X3D)
+THREADS=16
+THREADS_BATCH=16
 N_GPU_LAYERS=999
 # 14 MOE layers on CPU to free VRAM for q4_0 KV cache, 16 layers on GPU
-N_CPU_MOE=16
+N_CPU_MOE=8
 BATCH_SIZE=2048
 UBATCH_SIZE=512
 MLOCK=false
@@ -31,9 +29,9 @@ LOG_DISABLE=true
 
 # --- MEMORY (KV in VRAM, q4_0 Precision, 256K Context) ---
 # KV cache allocated on GPU (default kv-offload), q4_0 for VRAM efficiency
-CACHE_TYPE_K="q4_0"
+CACHE_TYPE_K="q8_0"
 CACHE_TYPE_V="q4_0"
-CTX_SIZE=262144  # 256K total context
+CTX_SIZE=131072  # 256K total context
 PARALLEL=4       # Four independent conversation slots
 PORT=8080
 
@@ -45,25 +43,22 @@ KV_UNIFIED=true
 CACHE_RAM=24576   # 24 GB prompt cache for multi-turn retention
 CACHE_IDLE_SLOTS=true # Auto-save idle slots to prompt cache
 CONTEXT_SHIFT=true    # Shift context on slot reuse
-SLOT_SAVE_PATH="/home/siva/.cache/llama-slots"  # SSD persistence via /slots/save API
+SLOT_SAVE_PATH="/home/siva/.cache/llama-slots/gemma-4-26b-a4b-qat"  # SSD persistence via /slots/save API
 CONT_BATCHING=true    # Continuous batching for throughput
 
 # --- INFRASTRUCTURE ---
 JINJA=true
-#JINJA_KWARGS='{"preserve_thinking": true}'
 
-TEMP=0.6
-TOP_P=0.9
+TEMP=1.0
+TOP_P=0.95
 TOP_K=64
-MIN_P=0.05
-REPEAT_PENALTY=1.1
 
 # --- REASONING ---
 REASONING="on"
 REASONING_FORMAT="auto"
 REASONING_BUDGET=-1
 REASONING_BUDGET_MESSAGE=" [Logic Finalized] "
 
-# --- UI/UX STANDARDS ---
 # Balanced MTP for 16GB: Reduced depth and quantized draft cache
-EXTRA_ARGS="--model-draft /home/siva/models/HauhauCS/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP/mtp-gemma-4-26B-A4B-it.gguf --spec-type draft-mtp --api-key omlx-Manage43NrtAdmin320"
+#EXTRA_ARGS="--api-key omlx-Manage43NrtAdmin320"
+EXTRA_ARGS="--model-draft /home/siva/models/gemma-4-26b-a4b-qat/unsloth/gemma-4-26b-a4b-qat/MTP/mtp-gemma-4-26B-A4B-it-Q8_0.gguf --spec-type draft-mtp --spec-draft-n-max 3 --api-key omlx-Manage43NrtAdmin320"