fix: add triton kernels for MXFP4, fix GGUF KV cache quantization

- Add 'kernels' package to Dockerfile for native MXFP4 execution
  (fixes gpt-oss-20b OOM: 15.2GB→13.5GB)
- Reduce GGUF n_ctx from 8192 to 4096 and quantize KV cache to Q8_0
  to reduce VRAM usage
- Use GGML_TYPE_Q8_0 constant instead of string for type_k/type_v

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

kischdle/llmux/Dockerfile

@@ -27,10 +27,11 @@ RUN pip install --no-cache-dir --break-system-packages \
     "sentencepiece>=0.2.0" \
     "protobuf>=5.0.0"
 
-# Install transformers + accelerate (needed for device_map)
+# Install transformers + accelerate + kernels (MXFP4/FP8 triton kernels)
 RUN pip install --no-cache-dir --break-system-packages --no-build-isolation \
     "transformers>=5.4.0" \
-    "accelerate>=1.0.0"
+    "accelerate>=1.0.0" \
+    "kernels"
 
 # Install chatterbox-tts WITHOUT its dependencies (it would downgrade
 # torch from 2.11 to 2.6 and pull gradio, librosa, etc.)

kischdle/llmux/llmux/backends/llamacpp.py

@@ -38,10 +38,13 @@ class LlamaCppBackend(BaseBackend):
         logger.info(f"Loading GGUF model {model_path} with n_gpu_layers={n_gpu_layers}")
 
         def _load():
+            from llama_cpp import GGML_TYPE_Q8_0
             kwargs = {
                 "model_path": model_path,
                 "n_gpu_layers": n_gpu_layers,
-                "n_ctx": 8192,
+                "n_ctx": 4096,
+                "type_k": GGML_TYPE_Q8_0,
+                "type_v": GGML_TYPE_Q8_0,
                 "verbose": False,
             }
             if physical.mmproj_file: