add imatrix quants and k8s training job

scripts/convert_and_quantize.sh

@@ -1,36 +1,53 @@
+
 #!/bin/bash
 set -e
 
-LLAMA_CPP=../llama.cpp
 MODEL_NAME=$1
+OUT_TYPE=${2:-"f16"}
+MODELS_DIR=${3:-"./models"}
+LLAMA_CPP=${4:-"./llama.cpp"}
 
-if [[ ! -d "./models/$MODEL_NAME" ]]; then
+if [[ ! -d "$MODELS_DIR/$MODEL_NAME" ]]; then
     echo "Unknown model $MODEL_NAME"
     exit -1
 fi
 
-if [ -f "./models/$MODEL_NAME/gguf_overrides.json" ]; then
-    OVERRIDES="--metadata ./models/$MODEL_NAME/gguf_overrides.json"
-    echo "Using metadata from ./models/$MODEL_NAME/gguf_overrides.json"
+if [ -f "$MODELS_DIR/$MODEL_NAME/gguf_overrides.json" ]; then
+    OVERRIDES="--metadata $MODELS_DIR/$MODEL_NAME/gguf_overrides.json"
+    echo "Using metadata from $MODELS_DIR/$MODEL_NAME/gguf_overrides.json"
 else
     OVERRIDES=""
 fi
 
 echo "Converting to GGUF..."
-if [ ! -f "./models/$MODEL_NAME/$MODEL_NAME.f16.gguf" ]; then
-    $LLAMA_CPP/convert_hf_to_gguf.py --outfile ./models/$MODEL_NAME/$MODEL_NAME.f16.gguf --outtype f16 ./models/$MODEL_NAME/ $OVERRIDES
+if [ ! -f "$MODELS_DIR/$MODEL_NAME/$MODEL_NAME.$OUT_TYPE.gguf" ]; then
+    $LLAMA_CPP/convert_hf_to_gguf.py --outfile $MODELS_DIR/$MODEL_NAME/$MODEL_NAME.$OUT_TYPE.gguf --outtype $OUT_TYPE $MODELS_DIR/$MODEL_NAME/ $OVERRIDES
 else
     echo "Converted model for already exists. Skipping..."
 fi
 
+echo "Generate imatrix for model..."
+if [ ! -f "groups_merged.txt" ]; then
+    echo "Downloading groups_merged.txt..."
+    wget https://huggingface.co/datasets/froggeric/imatrix/resolve/main/groups_merged.txt
+fi
 
-DESIRED_QUANTS=("Q8_0" "Q5_K_M" "Q4_0" "Q4_1" "Q4_K_M")
+if [ ! -f "$MODELS_DIR/$MODEL_NAME/$MODEL_NAME.imatrix.gguf" ]; then
+    $LLAMA_CPP/build/bin/llama-imatrix -m $MODELS_DIR/$MODEL_NAME/$MODEL_NAME.$OUT_TYPE.gguf -ngl 999 -c 512 -f groups_merged.txt -o $MODELS_DIR/$MODEL_NAME/$MODEL_NAME.imatrix.gguf
+else
+    echo "Imatrix model already exists. Skipping..."
+fi
+
+DESIRED_QUANTS=("Q8_0" "Q6_K" "Q5_K_M" "Q4_0" "Q4_1" "Q3_K_M" "IQ4_NL" "IQ4_XS")
 for QUANT in "${DESIRED_QUANTS[@]}"
 do
+    echo "Quantizing to $QUANT..."
     QUANT_LOWER=$(echo "$QUANT" | awk '{print tolower($0)}')
-    if [ ! -f "./models/$MODEL_NAME/$MODEL_NAME.$QUANT_LOWER.gguf" ]; then
-        $LLAMA_CPP/build/bin/llama-quantize ./models/$MODEL_NAME/$MODEL_NAME.f16.gguf ./models/$MODEL_NAME/$MODEL_NAME.$QUANT_LOWER.gguf $QUANT
+    if [ ! -f "$MODELS_DIR/$MODEL_NAME/$MODEL_NAME.$QUANT_LOWER.gguf" ]; then
+        $LLAMA_CPP/build/bin/llama-quantize --imatrix $MODELS_DIR/$MODEL_NAME/$MODEL_NAME.imatrix.gguf $MODELS_DIR/$MODEL_NAME/$MODEL_NAME.$OUT_TYPE.gguf $MODELS_DIR/$MODEL_NAME/$MODEL_NAME.$QUANT_LOWER.gguf $QUANT
     else
         echo "Quantized model for '$QUANT' already exists. Skipping..."
     fi
-done
+done
+
+echo "All done!"

train/training-job.yml

@@ -0,0 +1,71 @@
+apiVersion: batch/v1
+kind: Job
+metadata:
+  generateName: training-job-
+  namespace: ai
+  labels:
+    app: training-job
+spec:
+  template:
+    metadata:
+      labels:
+        app: training-job
+    spec:
+      containers:
+        - name: axolotl
+          image: axolotlai/axolotl-cloud:main-py3.11-cu128-2.8.0
+          imagePullPolicy: IfNotPresent
+          command: 
+            - axolotl
+            - train
+            - /workspace/configs/gemma3-270m.yml
+          env:
+            - name: AXOLOTL_DO_NOT_TRACK
+              value: "1"
+          volumeMounts:
+            - name: training-runs
+              mountPath: /workspace/data/training-runs
+            - name: training-data
+              mountPath: /workspace/data/datasets
+            - name: training-configs
+              mountPath: /workspace/configs
+            - name: hf-cache
+              mountPath: /workspace/data/huggingface-cache
+          resources:
+            limits:
+              nvidia.com/gpu: 2
+        - name: tensorboard
+          image: python:3.11-slim
+          imagePullPolicy: IfNotPresent
+          command:
+            - bash
+            - -c 
+            - "pip3 install tensorboard && tensorboard --logdir=/workspace/data/training-runs --host=0.0.0.0 --port 8080"
+          ports:
+            - containerPort: 8080
+              name: tensorboard
+              protocol: TCP
+          volumeMounts:
+            - name: training-runs
+              mountPath: /workspace/data/training-runs
+      runtimeClassName: nvidia
+      nodeSelector:
+        nvidia.com/gpu: "true"
+      restartPolicy: OnFailure
+      volumes:
+        - name: training-runs
+          hostPath:
+            path: /mnt/data/training-runs
+            type: DirectoryOrCreate
+        - name: training-data
+          hostPath:
+            path: /mnt/data/training-data
+            type: DirectoryOrCreate
+        - name: training-configs
+          hostPath:
+            path: /mnt/data/training-configs
+            type: DirectoryOrCreate
+        - name: hf-cache
+          hostPath:
+            path: /mnt/data/hf-cache
+            type: DirectoryOrCreate