Bump

flake.lock

@@ -64,16 +64,16 @@
         ]
       },
       "locked": {
-        "lastModified": 1745868199,
-        "narHash": "sha256-kTy1oNcN4bNHDV7yqIOSTo/JUN4enHeqyUeVE/q4iHg=",
+        "lastModified": 1745909149,
+        "narHash": "sha256-TDGa+/P14TAS1DC9MsTilNQ3sw/mL08s4KYIlTwW850=",
         "owner": "ggerganov",
         "repo": "llama.cpp",
-        "rev": "eaea3253244dc4bbe07f6cd81325847ccc6cf93e",
+        "rev": "b6ce7430b7eb51f032152316880204e0a9c0470e",
         "type": "github"
       },
       "original": {
         "owner": "ggerganov",
-        "ref": "b5214",
+        "ref": "b5216",
         "repo": "llama.cpp",
         "type": "github"
       }

flake.nix

@@ -19,7 +19,7 @@
     flake-parts.url = "github:hercules-ci/flake-parts";
     devshell.url = "github:numtide/devshell";
     llama-cpp = {
-      url = "github:ggerganov/llama.cpp/b5214";
+      url = "github:ggerganov/llama.cpp/b5216";
       inputs.nixpkgs.follows = "nixpkgs";
       inputs.flake-parts.follows = "flake-parts";
     };
@@ -89,6 +89,7 @@
           wayland
         ];
         rocmTargets = [
+          "gfx1030"
           "gfx1100"
           "gfx1102"
           "gfx1103"

llama_proxy_man/config.yaml

@@ -3,9 +3,9 @@ system_resources:
   vram: 30G
 model_specs:
   - name: "tabby-code"
-    port: 18080
+    port: 28080 # to make sure its not sued for now
     # internal_port: 28080 # Optional
-    autostart: true
+    autostart: false
     vram_usage: 27G     # Coder-32B + draft 0.5B
     ram_usage: 3G       # Coder-32B + draft 0.5B
     # vram_usage: 8.25G # Coder-7B
@@ -21,12 +21,12 @@ model_specs:
       gpu-layers: 9999
       model-draft: /media/SanDisk/ai/models_live/Qwen2.5-Coder-0.5B-Instruct-Q8_0.gguf
       gpu-layers-draft: 9999
-      cache-type-k: q8_0
+      # cache-type-k: q8_0
       cache-type-v: q8_0
       draft-max: 16
       draft-min: 5
   - name: "tabby-embeddings"
-    port: 18081
+    port: 28081
     vram_usage: 0.4G
     ram_usage: 2.5G
     env:
@@ -46,10 +46,19 @@ model_specs:
       CUDA_VISIBLE_DEVICES: 0
       HSA_OVERRIDE_GFX_VERSION: '11.0.0'
     args:
-      ctx-size: 16384
+      ctx-size: 65536
       flash-attn: true
-      model: /media/SanDisk/ai/models_live/Qwen_Qwen3-30B-A3B-Q5_K_S.gguf
+      model: /media/SanDisk/ai/models_live/Qwen_Qwen3-30B-A3B-Q4_K_L.gguf
+      cache-type-k: f16
+      cache-type-v: f16
       gpu-layers: 9999
+      rope-scaling: yarn 
+      rope-scale: 4 
+      yarn-orig-ctx: 32768
+      temp: 0.65
+      top-p: 0.9
+      top-k: 20
+      min-p: 0.0
       # model-draft: /media/SanDisk/ai/models_live/Qwen_Qwen3-0.6B-Q6_K.gguf
       # gpu-layers-draft: 9999
       # draft-max: 16