fix lfm 2.5

llamacpp/deployment-lfm25-26b.yaml

@@ -1,5 +1,5 @@
-# LiquidAI LFM2.5-2.6B with speculative decoding served by llama.cpp.
-# The primary and draft GGUFs are downloaded into the shared model PVC before
+# LiquidAI LFM2.5-2.6B served by llama.cpp (Vulkan, no speculative decoding).
+# The primary GGUF is downloaded into the shared model PVC before
 # llama-server starts.
 ---
 apiVersion: apps/v1
@@ -36,8 +36,7 @@ spec:
           set -eu
           apk add --no-cache curl
           for entry in \
-            "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF/resolve/main/LFM2.5-2.6B-Q4_K_M.gguf|LFM2.5-2.6B-Q4_K_M.gguf" \
-            "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF/resolve/main/LFM2.5-2.6B-Q4_0.gguf|LFM2.5-2.6B-Q4_0.gguf"; do
+            "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF/resolve/main/LFM2.5-2.6B-Q4_K_M.gguf|LFM2.5-2.6B-Q4_K_M.gguf"; do
             url=${entry%%|*}
             file=${entry##*|}
             if [ -s "/models/$file" ]; then
@@ -60,13 +59,13 @@ spec:
         args:
         - -m
         - /models/LFM2.5-2.6B-Q4_K_M.gguf
-        - --model-draft
-        - /models/LFM2.5-2.6B-Q4_0.gguf
-        - --spec-default
-        - --spec-type
-        - draft-simple
+        # Draft model + speculative decoding removed: the spec path crashes the
+        # server (GGML_ASSERT slot.spec_i_batch...) when the KV cache is under
+        # pressure, taking all in-flight requests (incl. home-manager crons) down.
         - --ctx-size
-        - "121000"
+        - "80000"
+        - --parallel
+        - "2"
         - --temp
         - "0.1"
         - --cache-type-k