Add Dockerfile and startup script for HF Space

Dockerfile

@@ -0,0 +1,16 @@
+# Start from a PyTorch + CUDA base image
+FROM pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel
+
+WORKDIR /code
+COPY . .
+
+# Install dependencies
+RUN pip install -r requirements.txt
+RUN pip install flash-attn --no-build-isolation
+
+# Make the startup script executable
+COPY startup.sh .
+RUN chmod +x startup.sh
+
+# This is the new command: it runs your script automatically when the Space starts.
+CMD ["./startup.sh"]

startup.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+# This script automates the setup and training process for the HRM fine-tuning.
+# The 'set -e' command ensures that the script will exit immediately if any command fails.
+set -e
+
+echo "--- Starting Automated Finetuning Workflow ---"
+
+# Step 1: Login to WandB using the secret
+# Make sure you have your WANDB_API_KEY set in the Space secrets
+wandb login $WANDB_API_KEY
+echo "Step 1 complete: Logged into WandB."
+
+# Step 2: Run the data curation and processing scripts
+# (Assuming your curation and processing scripts are in the repo)
+# python dataset_curation.py
+# python dataset/build_abstract_dataset.py
+echo "Step 2 complete: Data processing finished."
+
+# Step 3: Launch the training job
+echo "--- LAUNCHING TRAINING ---"
+OMP_NUM_THREADS=8 python pretrain.py \
+  data_path=data/abstract_optimizer_processed \
+  arch=hrm_abstract_optimizer \
+  epochs=20000 \
+  global_batch_size=32 \
+  eval_interval=1000
+
+echo "--- TRAINING COMPLETE ---"
+
+# Step 4: Keep the space running after the script finishes (optional)
+# echo "Workflow finished. Idling..."
+# sleep infinity