Upload 5 files

Dockerfile

@@ -1,7 +1,7 @@
 FROM nvidia/cuda:12.5.1-cudnn-devel-ubuntu20.04
 
 ENV DEBIAN_FRONTEND=noninteractive \
-	TZ=Europe/Paris
+        TZ=Asia/Shanghai 
 
 # Remove any third-party apt sources to avoid issues with expiring keys.
 # Install some basic utilities
@@ -51,7 +51,7 @@ RUN mkdir $HOME/.cache $HOME/.config \
 # Set up the Conda environment
 ENV CONDA_AUTO_UPDATE_CONDA=false \
     PATH=$HOME/miniconda/bin:$PATH
-RUN curl -sLo ~/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-py39_4.10.3-Linux-x86_64.sh \
+RUN curl -sLo ~/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-py310_4.10.3-Linux-x86_64.sh \
  && chmod +x ~/miniconda.sh \
  && ~/miniconda.sh -b -p ~/miniconda \
  && rm ~/miniconda.sh \
@@ -73,7 +73,7 @@ RUN --mount=target=/root/packages.txt,source=packages.txt \
     && rm -rf /var/lib/apt/lists/*
 
 RUN --mount=target=/root/on_startup.sh,source=on_startup.sh,readwrite \
-	bash /root/on_startup.sh
+        bash /root/on_startup.sh
 
 RUN mkdir /data && chown user:user /data
 
@@ -92,14 +92,20 @@ COPY --chown=user . $HOME/app
 
 RUN chmod +x start_server.sh
 
-COPY --chown=user login.html /home/user/miniconda/lib/python3.9/site-packages/jupyter_server/templates/login.html
+RUN download.sh
+
+COPY --chown=user login.html /home/user/miniconda/lib/python3.10/site-packages/jupyter_server/templates/login.html
+
+COPY --chown=user public.ipynb /data/public.ipynb
+
+# COPY --chown=user models.py /data/public.ipynb
 
 ENV PYTHONUNBUFFERED=1 \
-	GRADIO_ALLOW_FLAGGING=never \
-	GRADIO_NUM_PORTS=1 \
-	GRADIO_SERVER_NAME=0.0.0.0 \
-	GRADIO_THEME=huggingface \
-	SYSTEM=spaces \
-	SHELL=/bin/bash
-
-CMD ["./start_server.sh"]
+        GRADIO_ALLOW_FLAGGING=never \
+        GRADIO_NUM_PORTS=1 \
+        GRADIO_SERVER_NAME=0.0.0.0 \
+        GRADIO_THEME=huggingface \
+        SYSTEM=spaces \
+        SHELL=/bin/bash
+
+CMD ["./start_server.sh"]

ag4masses-public.ipynb

@@ -0,0 +1,473 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "executionInfo": {
+     "elapsed": 611,
+     "status": "ok",
+     "timestamp": 1733595497864,
+     "user": {
+      "displayName": "Tong Peng",
+      "userId": "14680520704856526492"
+     },
+     "user_tz": 300
+    },
+    "id": "-IHoHd-t5sLP",
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "import sys, os\n",
+    "\n",
+    "AG4MDIR='/home/user/ag4masses'\n",
+    "AGLIB=f'{AG4MDIR}/aglib'\n",
+    "AGDIR=f\"{AGLIB}/alphageometry\"\n",
+    "MELIAD_PATH=f\"{AGDIR}/meliad\"\n",
+    "DATA=f\"{AGLIB}/ag_ckpt_vocab\"\n",
+    "TESTDIR=f\"/data/ag4mtest\""
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "ASgGFu0NYHUH"
+   },
+   "source": [
+    "# Download Files"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Run this cell to refresh code and get the latest versions\n",
+    "# AG4MDIR and MELIAD_PATH are in /kaggle/working and will be saved as outputs of the Notebook\n",
+    "# !rm -fr {AG4MDIR} {MELIAD_PATH}"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "executionInfo": {
+     "elapsed": 5,
+     "status": "ok",
+     "timestamp": 1733594216384,
+     "user": {
+      "displayName": "Tong Peng",
+      "userId": "14680520704856526492"
+     },
+     "user_tz": 300
+    },
+    "id": "GgR_vO8XX9Vr",
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "if not os.path.exists(MELIAD_PATH):\n",
+    "    !git clone https://github.com/google-research/meliad.git {MELIAD_PATH}\n",
+    "\n",
+    "if not os.path.exists(AG4MDIR):\n",
+    "    !git clone https://github.com/tpgh24/ag4masses.git {AG4MDIR}\n",
+    "\n",
+    "# Temporarily modified files, upload into dataset tmpfiles\n",
+    "# !cp /kaggle/input/tmpfiles/numericals.py {AGDIR}\n",
+    "# !cp /kaggle/input/tmpfiles/alphageometry.py {AGDIR}\n",
+    "\n",
+    "if not os.path.exists(TESTDIR):\n",
+    "    !mkdir {TESTDIR}"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "executionInfo": {
+     "elapsed": 40247,
+     "status": "ok",
+     "timestamp": 1733594312243,
+     "user": {
+      "displayName": "Tong Peng",
+      "userId": "14680520704856526492"
+     },
+     "user_tz": 300
+    },
+    "id": "gP4zAZh2MHcv",
+    "outputId": "4796397b-8952-411e-bd33-8fd813865735",
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "if not os.path.exists(DATA):\n",
+    "  # download data: vocabulary, trained model\n",
+    "  !mkdir {DATA}\n",
+    "\n",
+    "  ### Convoluted process. This does no work, seems due to Google Drive restrictions\n",
+    "  # !gdown --folder https://bit.ly/alphageometry/\n",
+    "  #\n",
+    "  ### First got file links from Google Drive web UI, under Share menu. It will download HTML files with download button. Extract URL from the file.\n",
+    "  ### For checkpoint_10999999, because the file is big, there is an additional step asking for confirmation, got the final URL from the 2nd HTML,\n",
+    "  ### Link constructed using AI from HTML form.\n",
+    "  # !gdown https://drive.google.com/file/d/1mRd6J0UkeWoFUjeVB7BQi5lVNLvPBe31/view?usp=drive_link -O {AGLIB}/ag_ckpt_vocab/geometry.757.vocab\n",
+    "  # !gdown https://drive.google.com/file/d/1t-r3KfU8aDbS1UHpdyM3LH21rwSCIXTz/view?usp=drive_link -O {AGLIB}/ag_ckpt_vocab/geometry.757.model\n",
+    "  # !gdown https://drive.google.com/file/d/1qXkmmgoJ8oTYJdFV1xw0xGPpQj6SyOYA/view?usp=drive_link -O {AGLIB}/ag_ckpt_vocab/checkpoint_10999999\n",
+    "\n",
+    "  !wget -O {DATA}/geometry.757.vocab \"https://drive.usercontent.google.com/uc?id=1mRd6J0UkeWoFUjeVB7BQi5lVNLvPBe31&export=download\"\n",
+    "  !wget -O {DATA}/geometry.757.model \"https://drive.usercontent.google.com/uc?id=1t-r3KfU8aDbS1UHpdyM3LH21rwSCIXTz&export=download\"\n",
+    "  !wget -O {DATA}/checkpoint_10999999 \"https://drive.usercontent.google.com/download?id=1qXkmmgoJ8oTYJdFV1xw0xGPpQj6SyOYA&export=download&confirm=t&uuid=ae22f4de-cb77-4145-af5f-8cfbb59e867e\"\n",
+    "\n",
+    "!ls -l {DATA}"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Setup Env"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "executionInfo": {
+     "elapsed": 14190,
+     "status": "ok",
+     "timestamp": 1733594230570,
+     "user": {
+      "displayName": "Tong Peng",
+      "userId": "14680520704856526492"
+     },
+     "user_tz": 300
+    },
+    "id": "X8Aj3G0neT6K",
+    "outputId": "9538ceba-8065-44d6-a32f-35127e5f2575",
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Python packages for AlphaGeometry\n",
+    "!pip cache purge\n",
+    "!pip install --upgrade pip\n",
+    "!pip install --upgrade packaging setuptools setuptools_scm wheel \n",
+    "!pip install --require-hashes --no-deps -r {AG4MDIR}/alphageometry/requirements.txt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "executionInfo": {
+     "elapsed": 15694,
+     "status": "ok",
+     "timestamp": 1733594246256,
+     "user": {
+      "displayName": "Tong Peng",
+      "userId": "14680520704856526492"
+     },
+     "user_tz": 300
+    },
+    "id": "u9fuBSr2qEwN",
+    "outputId": "97bbce78-8b49-4d3b-a831-d188a4a9e536",
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Python packages for Nvidia gpu.\n",
+    "# The versions of Python packages used by AlphaGeometry seem to only work with Cuda 11, not 12\n",
+    "!pip install -U \"jax==0.4.6\" \"jaxlib[cuda11_cudnn86]==0.4.6\" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html\n",
+    "!pip install nvidia-cuda-runtime-cu11\n",
+    "!pip install nvidia-pyindex\n",
+    "# !pip list"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Linux packages for Nvidia gpu.\n",
+    "# The versions of Python packages used by AlphaGeometry seem to only work with Cuda 11, not 12\n",
+    "!apt-get update\n",
+    "!DEBIAN_FRONTEND=noninteractive apt install -y cuda-11-8\n",
+    "# !apt list|grep cuda"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "id": "fChy49CNhf01",
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Information about Nvidia drivers\n",
+    "!nvcc --version\n",
+    "!nvidia-smi"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "jUWvch7kYhxt"
+   },
+   "source": [
+    "# Execution"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "#!! cannot have ' in the script, including in comments\n",
+    "jobScript = '''\n",
+    "# !/bin/bash\n",
+    "set -e\n",
+    "set -x\n",
+    "\n",
+    "# stdout, solution is written here\n",
+    "OUTFILE=$TESTDIR/${PROB}.out\n",
+    "# stderr, a lot of information, error message, log etc.\n",
+    "ERRFILE=$TESTDIR/${PROB}.log\n",
+    "\n",
+    "# stdout and stderr are written to both ERRFILF and console\n",
+    "exec >$ERRFILE 2>&1\n",
+    "\n",
+    "echo PROB=$PROB\n",
+    "echo PROB_FILE=$PROBFILE\n",
+    "echo MODEL=$MODEL\n",
+    "\n",
+    "# Directory where output files go\n",
+    "echo TESTDIR=$TESTDIR\n",
+    "# Directory containing AG4Masses source files\n",
+    "echo AG4MDIR=$AG4MDIR\n",
+    "# Directory containing external libraries including ag_ckpt_vocab and meliad\n",
+    "echo AGLIB=$AGLIB\n",
+    "\n",
+    "AGDIR=$AG4MDIR/alphageometry\n",
+    "export PYTHONPATH=$PYTHONPATH:$AGDIR:$AGLIB\n",
+    "\n",
+    "echo BATCH_SIZE=$BATCH_SIZE\n",
+    "echo BEAM_SIZE=$BEAM_SIZE\n",
+    "echo DEPTH=$DEPTH\n",
+    "echo NWORKERS=$NWORKERS\n",
+    "\n",
+    "echo ERRFILE=$ERRFILE\n",
+    "echo OUTFILE=$OUTFILE\n",
+    "\n",
+    "DATA=$AGLIB/ag_ckpt_vocab\n",
+    "MELIAD_PATH=$AGLIB/meliad\n",
+    "export PYTHONPATH=$PYTHONPATH:$MELIAD_PATH\n",
+    "\n",
+    "DDAR_ARGS=( \\\n",
+    "  --defs_file=$AGDIR/defs.txt \\\n",
+    "  --rules_file=$AGDIR/rules.txt \\\n",
+    ")\n",
+    "\n",
+    "SEARCH_ARGS=(\n",
+    "  --beam_size=$BEAM_SIZE\n",
+    "  --search_depth=$DEPTH\n",
+    ")\n",
+    "\n",
+    "LM_ARGS=(\n",
+    "  --ckpt_path=$DATA \\\n",
+    "  --vocab_path=$DATA/geometry.757.model \\\n",
+    "  --gin_search_paths=$MELIAD_PATH/transformer/configs,$AGDIR \\\n",
+    "  --gin_file=base_htrans.gin \\\n",
+    "  --gin_file=size/medium_150M.gin \\\n",
+    "  --gin_file=options/positions_t5.gin \\\n",
+    "  --gin_file=options/lr_cosine_decay.gin \\\n",
+    "  --gin_file=options/seq_1024_nocache.gin \\\n",
+    "  --gin_file=geometry_150M_generate.gin \\\n",
+    "  --gin_param=DecoderOnlyLanguageModelGenerate.output_token_losses=True \\\n",
+    "  --gin_param=TransformerTaskConfig.batch_size=$BATCH_SIZE \\\n",
+    "  --gin_param=TransformerTaskConfig.sequence_length=128 \\\n",
+    "  --gin_param=Trainer.restore_state_variables=False\n",
+    ");\n",
+    "\n",
+    "true \"==========================================\"\n",
+    "\n",
+    "python -m alphageometry \\\n",
+    "--alsologtostderr \\\n",
+    "--problems_file=$PROB_FILE \\\n",
+    "--problem_name=$PROB \\\n",
+    "--mode=$MODEL \\\n",
+    "\"${DDAR_ARGS[@]}\" \\\n",
+    "\"${SEARCH_ARGS[@]}\" \\\n",
+    "\"${LM_ARGS[@]}\" \\\n",
+    "--out_file=$OUTFILE \\\n",
+    "--n_workers=$NWORKERS 2>&1\n",
+    "\n",
+    "'''"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "os.environ[\"TESTDIR\"]=TESTDIR\n",
+    "os.environ[\"AG4MDIR\"]=AG4MDIR\n",
+    "os.environ[\"AGLIB\"]=AGLIB\n",
+    "\n",
+    "# BATCH_SIZE: number of outputs for each LM query\n",
+    "# BEAM_SIZE: size of the breadth-first search queue\n",
+    "# DEPTH: search depth (number of auxilary points to add)\n",
+    "# NWORKERS: number of parallel run worker processes.\n",
+    "# \n",
+    "# Memory usage is affected by BATCH_SIZE, NWORKER and complexity of the problem.\n",
+    "# Larger NWORKER and BATCH_SIZE tends to cause out of memory issue\n",
+    "#\n",
+    "# The results in Google paper can be obtained by setting BATCH_SIZE=32, BEAM_SIZE=512, DEPTH=16\n",
+    "#\n",
+    "# 1/2025: Kaggle free version provides GPU T4x2, 4 virtual CPUs, 29G RAM. Can set \n",
+    "#   NWORKERS=2\n",
+    "#   CUDA_VISIBLE_DEVICES=0,1\n",
+    "\n",
+    "os.environ[\"BATCH_SIZE\"]=\"16\"\n",
+    "os.environ[\"BEAM_SIZE\"]=\"64\"\n",
+    "os.environ[\"DEPTH\"]=\"8\"\n",
+    "os.environ[\"NWORKERS\"]=\"2\"\n",
+    "\n",
+    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0,1\"\n",
+    "\n",
+    "# test problems can be uploaded into a dataset, e.g. for dataset \"tmpfiles\", \"/kaggle/input/tmpfiles/test-problems.txt\"\n",
+    "os.environ[\"PROB_FILE\"]=f\"{AG4MDIR}/data/ag4m_problems.txt\"\n",
+    "PROB=\"imo-2024-q4\"\n",
+    "os.environ[\"PROB\"]=PROB\n",
+    "# alphageometry|ddar\n",
+    "os.environ[\"MODEL\"]=\"alphageometry\"\n",
+    "\n",
+    "# In an interactive Kaggle session, run the job in background, so we can do other things in the Notebook.\n",
+    "# For long jobs, commit the Notebook and run in Batch mode.\n",
+    "# An interactive session will be terminated after about 20 minutes of idle time.\n",
+    "if os.environ[\"KAGGLE_KERNEL_RUN_TYPE\"]==\"Batch\":\n",
+    "    os.system(f\"echo '{jobScript}'|bash\")\n",
+    "else:\n",
+    "    os.system(f\"echo '{jobScript}'|bash &\")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "#!cat /kaggle/input/tmpfiles/test-problems.txt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# In an interactive Kaggle session, run this to see the log file. We can cancel this cell's execution\n",
+    "# to do other things in the Notebook\n",
+    "if os.environ[\"KAGGLE_KERNEL_RUN_TYPE\"] != \"Batch\":\n",
+    "    !tail -f {TESTDIR}/{PROB}.log"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Command to kill the background job in an interactive session\n",
+    "# !pkill -P `ps -ef|grep -- '-m alphageometry'|grep -v grep|awk '{printf \"%d,%d\", $2, $3}'`\n",
+    "\n",
+    "# Command to check processes\n",
+    "### |cat to show full commandline\n",
+    "# !ps -eo pid,ppid,pgid,%cpu,cmd |cat\n",
+    "# !ps -ef|cat"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# Command to check progress of a running job in an interactive session\n",
+    "# !bash {AG4MDIR}/utils/checkprog.sh {TESTDIR}/{PROB}.log"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "trusted": true
+   },
+   "outputs": [],
+   "source": [
+    "# In Batch run, after the job completes, list output files\n",
+    "!ls -ltr {TESTDIR}"
+   ]
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "authorship_tag": "ABX9TyOcsgkfOgCk5oTpUiS6zrgo",
+   "collapsed_sections": [
+    "pW2KIijZBAdh"
+   ],
+   "gpuType": "T4",
+   "provenance": []
+  },
+  "kaggle": {
+   "accelerator": "nvidiaTeslaT4",
+   "dataSources": [],
+   "dockerImageVersionId": 30823,
+   "isGpuEnabled": true,
+   "isInternetEnabled": true,
+   "language": "python",
+   "sourceType": "notebook"
+  },
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

alphageometry.py

@@ -0,0 +1,778 @@
+# Copyright 2023 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Run DD+AR or AlphaGeometry solver.
+
+Please refer to README.md for detailed instructions.
+"""
+
+import time
+import traceback
+
+from absl import app
+from absl import flags
+from absl import logging
+import ddar
+import graph as gh
+import lm_inference as lm
+import pretty as pt
+import problem as pr
+
+#=============
+import sys, os, math, re
+import multiprocessing
+model = None # global variable used in multi-processing workers
+
+_GIN_SEARCH_PATHS = flags.DEFINE_list(
+    'gin_search_paths',
+    ['third_party/py/meliad/transformer/configs'],
+    'List of paths where the Gin config files are located.',
+)
+_GIN_FILE = flags.DEFINE_multi_string(
+    'gin_file', ['base_htrans.gin'], 'List of Gin config files.'
+)
+_GIN_PARAM = flags.DEFINE_multi_string(
+    'gin_param', None, 'Newline separated list of Gin parameter bindings.'
+)
+
+_PROBLEMS_FILE = flags.DEFINE_string(
+    'problems_file',
+    'imo_ag_30.txt',
+    'text file contains the problem strings. See imo_ag_30.txt for example.',
+)
+_PROBLEM_NAME = flags.DEFINE_string(
+    'problem_name',
+    'imo_2000_p1',
+    'name of the problem to solve, must be in the problem_file.',
+)
+_MODE = flags.DEFINE_string(
+    'mode', 'ddar', 'either `ddar` (DD+AR) or `alphageometry`')
+_DEFS_FILE = flags.DEFINE_string(
+    'defs_file',
+    'defs.txt',
+    'definitions of available constructions to state a problem.',
+)
+_RULES_FILE = flags.DEFINE_string(
+    'rules_file', 'rules.txt', 'list of deduction rules used by DD.'
+)
+_CKPT_PATH = flags.DEFINE_string('ckpt_path', '', 'checkpoint of the LM model.')
+_VOCAB_PATH = flags.DEFINE_string(
+    'vocab_path', '', 'path to the LM vocab file.'
+)
+_OUT_FILE = flags.DEFINE_string(
+    'out_file', '', 'path to the solution output file.'
+)  # pylint: disable=line-too-long
+_BEAM_SIZE = flags.DEFINE_integer(
+    'beam_size', 1, 'beam size of the proof search.'
+)  # pylint: disable=line-too-long
+_SEARCH_DEPTH = flags.DEFINE_integer(
+    'search_depth', 1, 'search depth of the proof search.'
+)  # pylint: disable=line-too-long
+
+#===================================
+_N_WORKSERS = flags.DEFINE_integer(
+    'n_workers', 1, 'number of workers'
+)# pylint: disable=line-too-long
+
+DEFINITIONS = None  # contains definitions of construction actions
+RULES = None  # contains rules of deductions
+
+
+def natural_language_statement(logical_statement: pr.Dependency) -> str:
+  """Convert logical_statement to natural language.
+
+  Args:
+    logical_statement: pr.Dependency with .name and .args
+
+  Returns:
+    a string of (pseudo) natural language of the predicate for human reader.
+  """
+  names = [a.name.upper() for a in logical_statement.args]
+  names = [(n[0] + '_' + n[1:]) if len(n) > 1 else n for n in names]
+  return pt.pretty_nl(logical_statement.name, names)
+
+
+def proof_step_string(
+    proof_step: pr.Dependency, refs: dict[tuple[str, ...], int], last_step: bool
+) -> str:
+  """Translate proof to natural language.
+
+  Args:
+    proof_step: pr.Dependency with .name and .args
+    refs: dict(hash: int) to keep track of derived predicates
+    last_step: boolean to keep track whether this is the last step.
+
+  Returns:
+    a string of (pseudo) natural language of the proof step for human reader.
+  """
+  premises, [conclusion] = proof_step
+
+  premises_nl = ' & '.join(
+      [
+          natural_language_statement(p) + ' [{:02}]'.format(refs[p.hashed()])
+          for p in premises
+      ]
+  )
+
+  if not premises:
+    premises_nl = 'similarly'
+
+  refs[conclusion.hashed()] = len(refs)
+
+  conclusion_nl = natural_language_statement(conclusion)
+  if not last_step:
+    conclusion_nl += ' [{:02}]'.format(refs[conclusion.hashed()])
+
+  return f'{premises_nl} \u21d2 {conclusion_nl}'
+
+
+def write_solution(g: gh.Graph, p: pr.Problem, out_file: str) -> None:
+  """Output the solution to out_file.
+
+  Args:
+    g: gh.Graph object, containing the proof state.
+    p: pr.Problem object, containing the theorem.
+    out_file: file to write to, empty string to skip writing to file.
+  """
+  setup, aux, proof_steps, refs = ddar.get_proof_steps(
+      g, p.goal, merge_trivials=False
+  )
+
+  solution = '\n=========================='
+  solution += '\n * From theorem premises:\n'
+  premises_nl = []
+  for premises, [points] in setup:
+    solution += ' '.join([p.name.upper() for p in points]) + ' '
+    if not premises:
+      continue
+    premises_nl += [
+        natural_language_statement(p) + ' [{:02}]'.format(refs[p.hashed()])
+        for p in premises
+    ]
+  solution += ': Points\n' + '\n'.join(premises_nl)
+
+  solution += '\n\n * Auxiliary Constructions:\n'
+  aux_premises_nl = []
+  for premises, [points] in aux:
+    solution += ' '.join([p.name.upper() for p in points]) + ' '
+    aux_premises_nl += [
+        natural_language_statement(p) + ' [{:02}]'.format(refs[p.hashed()])
+        for p in premises
+    ]
+  solution += ': Points\n' + '\n'.join(aux_premises_nl)
+
+  # some special case where the deduction rule has a well known name.
+  r2name = {
+      'r32': '(SSS)',
+      'r33': '(SAS)',
+      'r34': '(Similar Triangles)',
+      'r35': '(Similar Triangles)',
+      'r36': '(ASA)',
+      'r37': '(ASA)',
+      'r38': '(Similar Triangles)',
+      'r39': '(Similar Triangles)',
+      'r40': '(Congruent Triangles)',
+      'a00': '(Distance chase)',
+      'a01': '(Ratio chase)',
+      'a02': '(Angle chase)',
+  }
+
+  solution += '\n\n * Proof steps:\n'
+  for i, step in enumerate(proof_steps):
+    _, [con] = step
+    nl = proof_step_string(step, refs, last_step=i == len(proof_steps) - 1)
+    rule_name = r2name.get(con.rule_name, '')
+    nl = nl.replace('\u21d2', f'{rule_name}\u21d2 ')
+    solution += '{:03}. '.format(i + 1) + nl + '\n'
+
+  solution += '==========================\n'
+  logging.info(solution)
+  if out_file:
+    with open(out_file, 'w') as f:
+      f.write(solution)
+    logging.info('Solution written to %s.', out_file)
+
+
+def get_lm(ckpt_init: str, vocab_path: str) -> lm.LanguageModelInference:
+  lm.parse_gin_configuration(
+      _GIN_FILE.value, _GIN_PARAM.value, gin_paths=_GIN_SEARCH_PATHS.value
+  )
+
+  return lm.LanguageModelInference(vocab_path, ckpt_init, mode='beam_search')
+
+
+def run_ddar(g: gh.Graph, p: pr.Problem, out_file: str) -> bool:
+  """Run DD+AR.
+
+  Args:
+    g: gh.Graph object, containing the proof state.
+    p: pr.Problem object, containing the problem statement.
+    out_file: path to output file if solution is found.
+
+  Returns:
+    Boolean, whether DD+AR finishes successfully.
+  """
+  ddar.solve(g, RULES, p, max_level=1000)
+
+  goal_args = g.names2nodes(p.goal.args)
+  if not g.check(p.goal.name, goal_args):
+    logging.info('DD+AR failed to solve the problem.')
+    return False
+
+  write_solution(g, p, out_file)
+
+  gh.nm.draw(
+      g.type2nodes[gh.Point],
+      g.type2nodes[gh.Line],
+      g.type2nodes[gh.Circle],
+      g.type2nodes[gh.Segment])
+  return True
+
+
+def translate_constrained_to_constructive(
+    point: str, name: str, args: list[str]
+) -> tuple[str, list[str]]:
+  """Translate a predicate from constraint-based to construction-based.
+
+  Args:
+    point: str: name of the new point
+    name: str: name of the predicate, e.g., perp, para, etc.
+    args: list[str]: list of predicate args.
+
+  Returns:
+    (name, args): translated to constructive predicate.
+  """
+  if name in ['T', 'perp']:
+    a, b, c, d = args
+    if point in [c, d]:
+      a, b, c, d = c, d, a, b
+    if point == b:
+      a, b = b, a
+    if point == d:
+      c, d = d, c
+    if a == c and a == point:
+      return 'on_dia', [a, b, d]
+    return 'on_tline', [a, b, c, d]
+
+  elif name in ['P', 'para']:
+    a, b, c, d = args
+    if point in [c, d]:
+      a, b, c, d = c, d, a, b
+    if point == b:
+      a, b = b, a
+    return 'on_pline', [a, b, c, d]
+
+  elif name in ['D', 'cong']:
+    a, b, c, d = args
+    if point in [c, d]:
+      a, b, c, d = c, d, a, b
+    if point == b:
+      a, b = b, a
+    if point == d:
+      c, d = d, c
+    if a == c and a == point:
+      return 'on_bline', [a, b, d]
+    if b in [c, d]:
+      if b == d:
+        c, d = d, c  # pylint: disable=unused-variable
+      return 'on_circle', [a, b, d]
+    return 'eqdistance', [a, b, c, d]
+
+  elif name in ['C', 'coll']:
+    a, b, c = args
+    if point == b:
+      a, b = b, a
+    if point == c:
+      a, b, c = c, a, b
+    return 'on_line', [a, b, c]
+
+  elif name in ['^', 'eqangle']:
+    a, b, c, d, e, f = args
+
+    if point in [d, e, f]:
+      a, b, c, d, e, f = d, e, f, a, b, c
+
+    x, b, y, c, d = b, c, e, d, f
+    if point == b:
+      a, b, c, d = b, a, d, c
+
+    if point == d and x == y:  # x p x b = x c x p
+      return 'angle_bisector', [point, b, x, c]
+
+    if point == x:
+      return 'eqangle3', [x, a, b, y, c, d]
+
+    return 'on_aline', [a, x, b, c, y, d]
+
+  elif name in ['cyclic', 'O']:
+    a, b, c = [x for x in args if x != point]
+    return 'on_circum', [point, a, b, c]
+
+  return name, args
+
+
+def check_valid_args(name: str, args: list[str]) -> bool:
+  """Check whether a predicate is grammarically correct.
+
+  Args:
+    name: str: name of the predicate
+    args: list[str]: args of the predicate
+
+  Returns:
+    bool: whether the predicate arg count is valid.
+  """
+  if name == 'perp':
+    if len(args) != 4:
+      return False
+    a, b, c, d = args
+    if len({a, b}) < 2:
+      return False
+    if len({c, d}) < 2:
+      return False
+  elif name == 'para':
+    if len(args) != 4:
+      return False
+    a, b, c, d = args
+    if len({a, b, c, d}) < 4:
+      return False
+  elif name == 'cong':
+    if len(args) != 4:
+      return False
+    a, b, c, d = args
+    if len({a, b}) < 2:
+      return False
+    if len({c, d}) < 2:
+      return False
+  elif name == 'coll':
+    if len(args) != 3:
+      return False
+    a, b, c = args
+    if len({a, b, c}) < 3:
+      return False
+  elif name == 'cyclic':
+    if len(args) != 4:
+      return False
+    a, b, c, d = args
+    if len({a, b, c, d}) < 4:
+      return False
+  elif name == 'eqangle':
+    if len(args) != 8:
+      return False
+    a, b, c, d, e, f, g, h = args
+    if len({a, b, c, d}) < 3:
+      return False
+    if len({e, f, g, h}) < 3:
+      return False
+  return True
+
+
+def try_translate_constrained_to_construct(string: str, g: gh.Graph) -> str:
+  """Whether a string of aux construction can be constructed.
+
+  Args:
+    string: str: the string describing aux construction.
+    g: gh.Graph: the current proof state.
+
+  Returns:
+    str: whether this construction is valid. If not, starts with "ERROR:".
+  """
+  if string[-1] != ';':
+    return 'ERROR: must end with ;'
+
+  logging.info(f'PID={os.getpid()}: !! try_translate_constrained_to_construct: string=%s', string)
+
+  # sometimes the LM may return ill-formed result with multiple colons.
+  # example:
+  #
+  # napoleon2
+  # a1 a2 a3 = triangle; c3 = s_angle a1 a2 c3 30, s_angle a2 a1 c3 150; c1 = s_angle a2 a3 c1 30, s_angle a3 a2 c1 150; c2 = s_angle a3 a1 c2 30, s_angle a1 a3 c2 150 ? cong c1 c2 c1 c3
+  #
+  # in the process, 
+  # I0210 17:58:01.513668 140016515833856 alphageometry.py:550] Decoding from {S} a : ; b : ; c : ; d : ^ a d a b 5. pi / 6. 00 ^ b d b a 1. pi / 6. 01 ; e : ^ b e b c 5. pi / 6. 02 ^ c e c b 1. pi / 6. 03 ; f : ^ a f a c 1. pi / 6. 04 ^ c f c a 5. pi / 6. 05 ? D e f e d {F1} x00 g : C a b g 06 D a g b g 07 ; x00 h : C c b h 08 D c h b h 09 ; x00
+  # I0210 18:01:38.182158 140016515833856 alphageometry.py:384] !! try_translate_constrained_to_construct: string=i : C a c i 10 D a i c i 11 ? V d f {F1} x00 j : D g j h j 12 D h j i j 13 ;
+
+  #XXX
+  # str_parts = string.split(' : ')
+  # if len(str_parts) != 2:
+  #   return f'ERROR: string has multiple colons: |{string}|'
+  mch = re.match('(.*?)( \? | \. \{)', string)
+  if mch :
+    strFixed = mch.group(1) + ';'
+    logging.info(f'ID={os.getpid()}: Bad LM output: {string}. Changed to {strFixed}')
+    string = strFixed
+
+  # sometimes the constraint in string is empty:
+  # 0407 17:11:35.470240 126383800963072 alphageometry.py:394] !! try_translate_constrained_to_construct: string=j : ;
+  hdprem = string.split(' : ')
+  if len(hdprem) !=2 or hdprem[1].strip()==';' :
+    logging.info(f'ID={os.getpid()}: Bad LM output: {string}. ERROR')
+    return f'ERROR: Bad LM output: {string}'
+  head, prem_str = hdprem
+  point = head.strip()
+
+  if len(point) != 1 or point == ' ':
+    return f'ERROR: invalid point name {point}'
+
+  existing_points = [p.name for p in g.all_points()]
+  if point in existing_points:
+    return f'ERROR: point {point} already exists.'
+
+  prem_toks = prem_str.split()[:-1]  # remove the EOS ' ;'
+  prems = [[]]
+
+  for i, tok in enumerate(prem_toks):
+    if tok.isdigit():
+      if i < len(prem_toks) - 1:
+        prems.append([])
+    else:
+      prems[-1].append(tok)
+
+  if len(prems) > 2:
+    return 'ERROR: there cannot be more than two predicates.'
+
+  clause_txt = point + ' = '
+  constructions = []
+
+  for prem in prems:
+    name, *args = prem
+
+    if point not in args:
+      return f'ERROR: {point} not found in predicate args.'
+
+    if not check_valid_args(pt.map_symbol(name), args):
+      return 'ERROR: Invalid predicate ' + name + ' ' + ' '.join(args)
+
+    for a in args:
+      if a != point and a not in existing_points:
+        return f'ERROR: point {a} does not exist.'
+
+    try:
+      name, args = translate_constrained_to_constructive(point, name, args)
+    except:  # pylint: disable=bare-except
+      return 'ERROR: Invalid predicate ' + name + ' ' + ' '.join(args)
+
+    if name == 'on_aline':
+      if args.count(point) > 1:
+        return f'ERROR: on_aline involves twice {point}'
+
+    constructions += [name + ' ' + ' '.join(args)]
+
+  clause_txt += ', '.join(constructions)
+  clause = pr.Clause.from_txt(clause_txt)
+
+  try:
+    g.copy().add_clause(clause, 0, DEFINITIONS)
+  except:  # pylint: disable=bare-except
+    return 'ERROR: ' + traceback.format_exc()
+
+  return clause_txt
+
+
+def insert_aux_to_premise(pstring: str, auxstring: str) -> str:
+  """Insert auxiliary constructs from proof to premise.
+
+  Args:
+    pstring: str: describing the problem to solve.
+    auxstring: str: describing the auxiliar construction.
+
+  Returns:
+    str: new pstring with auxstring inserted before the conclusion.
+  """
+  setup, goal = pstring.split(' ? ')
+  return setup + '; ' + auxstring + ' ? ' + goal
+
+
+class BeamQueue:
+  """Keep only the top k objects according to their values."""
+
+  def __init__(self, max_size: int = 512):
+    self.queue = []
+    self.max_size = max_size
+
+  def add(self, node: object, val: float) -> None:
+    """Add a new node to this queue."""
+
+    if len(self.queue) < self.max_size:
+      self.queue.append((val, node))
+      return
+
+    # Find the minimum node:
+    min_idx, (min_val, _) = min(enumerate(self.queue), key=lambda x: x[1])
+
+    # replace it if the new node has higher value.
+    if val > min_val:
+      self.queue[min_idx] = (val, node)
+
+  def __iter__(self):
+    for val, node in self.queue:
+      yield val, node
+
+  def __len__(self) -> int:
+    return len(self.queue)
+
+def bqsearch_init(worker_id):
+    # When using spawn or forkserver start method for multiprocessing.Pool, need to re-initialize
+    flags.FLAGS(sys.argv)
+    logging.use_absl_handler()
+    logging.set_verbosity(logging.INFO)
+    sys.setrecursionlimit(10000)
+
+    # Global variables initialized in main(). Need to re-initialize
+    #
+    # definitions of terms used in our domain-specific language.
+    global DEFINITIONS, RULES
+    DEFINITIONS = pr.Definition.from_txt_file(_DEFS_FILE.value, to_dict=True)
+    # load inference rules used in DD.
+    RULES = pr.Theorem.from_txt_file(_RULES_FILE.value, to_dict=True)
+
+    wkrpid = os.getpid()
+    logging.info('Worker %d initializing. PID=%d', worker_id, wkrpid)
+
+    if 'CUDA_VISIBLE_DEVICES' in os.environ and os.environ['CUDA_VISIBLE_DEVICES'].strip():
+      os.environ['CUDA_VISIBLE_DEVICES']=f"{worker_id}"
+      logging.info('Worker %d: CUDA_VISIBLE_DEVICES=%s', worker_id, os.environ['CUDA_VISIBLE_DEVICES'])
+
+    global model
+    model = get_lm(_CKPT_PATH.value, _VOCAB_PATH.value)
+    return wkrpid
+
+def bqsearch(i_nd, srch_inputs, out_file) -> tuple[int, bool, list]: # ( iNode, solved, [ (node, score) ] )
+    pid = os.getpid()
+    logging.info(f'Worker PID={pid} called for beam search node {i_nd}')
+
+    prev_score, (g, string, pstring) = srch_inputs
+    logging.info(f'Worker PID={pid}: Beam-searching and Decoding from {string}')
+    outputs = model.beam_decode(string, eos_tokens=[';'])
+
+    # translate lm output to the constructive language.
+    # so that we can update the graph representing proof states:
+    translations = [
+        try_translate_constrained_to_construct(o, g)
+        for o in outputs['seqs_str']
+    ]
+
+    # couple the lm outputs with its translations
+    candidates = zip(outputs['seqs_str'], translations, outputs['scores'])
+
+    # bring the highest scoring candidate first
+    candidates = reversed(list(candidates))
+
+    ret = []
+    for lm_out, translation, score in candidates:
+      logging.info(f'Worker PID={pid}: LM output (score={score}): "{lm_out}"')
+      logging.info(f'Worker PID={pid}: Translation: "{translation}"')
+
+      if translation.startswith('ERROR:'):
+        # the construction is invalid.
+        continue
+
+      # Update the constructive statement of the problem with the aux point:
+      candidate_pstring = insert_aux_to_premise(pstring, translation)
+
+      #XXX
+      logging.info(f'Worker PID={pid}: string=|{string}| lm_out=|{lm_out}|')
+      logging.info(f'Worker PID={pid}: Solving: "{candidate_pstring}"')
+      p_new = pr.Problem.from_txt(candidate_pstring)
+
+      # This is the new proof state graph representation:
+      g_new, _ = gh.Graph.build_problem(p_new, DEFINITIONS)
+
+      try:
+        if run_ddar(g_new, p_new, out_file):
+          logging.info(f'Worker PID={pid}: Solved.')
+          return (i_nd, True, None)
+      except Exception as e:
+          logging.info(f'Worker PID={pid}: Error in run_ddar: {e}')
+
+      # Add the candidate to the beam queue.
+      ret.append( [
+          # The string for the new node is old_string + lm output +
+          # the special token asking for a new auxiliary point ' x00':
+          # node
+          (g_new, string + ' ' + lm_out + ' x00', candidate_pstring),
+          # the score of each node is sum of score of all nodes
+          # on the path to itself. For beam search, there is no need to
+          # normalize according to path length because all nodes in beam
+          # is of the same path length.
+          # val
+          prev_score + score ]
+      )
+
+    logging.info(f'Worker PID={pid} beam search node {i_nd}: returning')
+    return (i_nd, False, ret)
+
+def run_alphageometry(
+    #XX model: lm.LanguageModelInference,
+    p: pr.Problem,
+    search_depth: int,
+    beam_size: int,
+    out_file: str,
+) -> bool:
+  """Simplified code to run AlphaGeometry proof search.
+
+  We removed all optimizations that are infrastructure-dependent, e.g.
+  parallelized model inference on multi GPUs,
+  parallelized DD+AR on multiple CPUs,
+  parallel execution of LM and DD+AR,
+  shared pool of CPU workers across different problems, etc.
+
+  Many other speed optimizations and abstractions are also removed to
+  better present the core structure of the proof search.
+
+  Args:
+    model: Interface with inference-related endpoints to JAX's model.
+    p: pr.Problem object describing the problem to solve.
+    search_depth: max proof search depth.
+    beam_size: beam size of the proof search.
+    out_file: path to output file if solution is found.
+
+  Returns:
+    boolean of whether this is solved.
+  """
+  # translate the problem to a string of grammar that the LM is trained on.
+  string = p.setup_str_from_problem(DEFINITIONS)
+  # special tokens prompting the LM to generate auxiliary points.
+  string += ' {F1} x00'
+  # the graph to represent the proof state.
+  g, _ = gh.Graph.build_problem(p, DEFINITIONS)
+
+  # First we run the symbolic engine DD+AR:
+  if run_ddar(g, p, out_file):
+    return True
+  
+  # ?? when pickling graph for some problems, the default recursion limit 1000 is not enough,
+  #    got 'maximum recursion depth exceeded while pickling an object' error
+  sys.setrecursionlimit(10000)
+
+  # beam search for the proof
+  # each node in the search tree is a 3-tuple:
+  # (<graph representation of proof state>,
+  #  <string for LM to decode from>,
+  #  <original problem string>)
+  beam_queue = BeamQueue(max_size=beam_size)
+  # originally the beam search tree starts with a single node (a 3-tuple):
+  beam_queue.add(
+      node=(g, string, p.txt()), val=0.0  # value of the root node is simply 0.
+  )
+
+  pool = None
+  if _N_WORKSERS.value == 1:
+    bqsearch_init(0)
+  else:
+    # Default is 'fork' on Linux, does not work with CUDA. Need to use 'spawn' or 'forkserver'
+    multiprocessing.set_start_method('spawn')
+    pool = multiprocessing.Pool(_N_WORKSERS.value)
+
+    logging.info("Initializing workers")
+    wkrpids = pool.map(bqsearch_init, range(_N_WORKSERS.value))
+    logging.info("Worker PIDs: " + str(wkrpids))
+
+  for depth in range(search_depth):
+    logging.info(
+        'Depth %s. There are %i nodes to expand:', depth, len(beam_queue)
+    )
+    for _, (_, string, _) in beam_queue:
+      logging.info(string)
+
+    new_queue = BeamQueue(max_size=beam_size)  # to replace beam_queue.
+    if _N_WORKSERS.value==1:
+      for i, srch_inputs in enumerate(beam_queue):
+        _, solved, res = bqsearch(i, srch_inputs, out_file)
+        if solved:
+          return True
+        for node, val in res:
+          # Add the candidate to the beam queue.
+          new_queue.add(node, val)
+          # Note that the queue only maintain at most beam_size nodes
+          # so this new node might possibly be dropped depending on its value.
+    else:      
+      jobs = [pool.apply_async(bqsearch, (i, srch_inputs, out_file)) for i, srch_inputs in enumerate(beam_queue)]
+
+      n_done = 0
+      while n_done < len(beam_queue):
+        for i, jobres in enumerate(jobs):
+          if jobres and jobres.ready():            
+            n_done += 1
+            jobs[i] = None
+            _, solved, res = jobres.get()
+            if solved:
+              # Clean up resources
+              pool.terminate()
+              pool.join()
+              return True
+            for node, val in res:
+              # Add the candidate to the beam queue.
+              new_queue.add(node, val)
+              # Note that the queue only maintain at most beam_size nodes
+              # so this new node might possibly be dropped depending on its value.            
+        time.sleep(1)  # Adjust wait time as needed
+
+    # replace the old queue with new queue before the new proof search depth.
+    beam_queue = new_queue
+
+  # Clean up resources
+  if pool:
+    pool.terminate()
+    pool.join()
+  return False
+
+def main(_):
+  global DEFINITIONS
+  global RULES
+
+  # definitions of terms used in our domain-specific language.
+  DEFINITIONS = pr.Definition.from_txt_file(_DEFS_FILE.value, to_dict=True)
+  # load inference rules used in DD.
+  RULES = pr.Theorem.from_txt_file(_RULES_FILE.value, to_dict=True)
+
+  # when using the language model,
+  # point names will be renamed to alphabetical a, b, c, d, e, ...
+  # instead of staying with their original names,
+  # in order to match the synthetic training data generation.
+  need_rename = _MODE.value != 'ddar'
+
+  # load problems from the problems_file,
+  problems = pr.Problem.from_txt_file(
+      _PROBLEMS_FILE.value, to_dict=True, translate=need_rename
+  )
+
+  if _PROBLEM_NAME.value not in problems:
+    raise ValueError(
+        f'Problem name `{_PROBLEM_NAME.value}` '
+        + f'not found in `{_PROBLEMS_FILE.value}`'
+    )
+
+  this_problem = problems[_PROBLEM_NAME.value]
+
+  if _MODE.value == 'ddar':
+    g, _ = gh.Graph.build_problem(this_problem, DEFINITIONS)
+    run_ddar(g, this_problem, _OUT_FILE.value)
+
+  elif _MODE.value == 'alphageometry':
+    #XX model = get_lm(_CKPT_PATH.value, _VOCAB_PATH.value)
+    run_alphageometry(
+        #XX model,
+        this_problem,
+        _SEARCH_DEPTH.value,
+        _BEAM_SIZE.value,
+        _OUT_FILE.value,
+    )
+
+  else:
+    raise ValueError(f'Unknown FLAGS.mode: {_MODE.value}')
+
+
+if __name__ == '__main__':
+  app.run(main)

download.sh

@@ -0,0 +1,31 @@
+HOME=/home/user
+AG4MDIR=$HOME/ag4masses
+# Directory containing external libraries including ag_ckpt_vocab and meliad
+AGLIB=$HOME/aglib
+
+AGDIR=$AG4MDIR/alphageometry
+
+DATA=$AGLIB/ag_ckpt_vocab
+MELIAD_PATH=$AGLIB/meliad
+
+cd /home/user
+mkdir aglib
+cd aglib
+git clone https://github.com/tpgh24/ag4masses.git
+cd alphageometry
+git clone https://github.com/google-research/meliad.git
+
+cd ..
+# mkdir aglib
+gdown --folder https://bit.ly/alphageometry
+export DATA=ag_ckpt_vocab
+
+pip cache purge
+pip install --upgrade pip
+pip install --upgrade packaging setuptools setuptools_scm wheel 
+pip install --require-hashes --no-deps -r /home/user/ag4masses/alphageometry/requirements.txt
+
+# some patch for cpu
+cp models.py /home/user/ag4masses/alphageometry/models.py
+cp alphageometry.py /home/user/ag4masses/alphageometry/alphageometry.py
+cp ag4masses-public.ipynb /data/ag4masses-public.ipynb

models.py

@@ -0,0 +1,178 @@
+# Copyright 2023 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Transformer language model generate mode."""
+
+from typing import Any, Tuple
+import beam_search
+import decoder_stack
+import gin
+import jax
+import jax.numpy as jnp
+from transformer import models
+
+
+@gin.configurable
+class DecoderOnlyLanguageModelGenerate(models.DecoderOnlyLanguageModel):
+  """Decoder only language modeling in inference mode."""
+
+  decoder_factory = decoder_stack.DecoderStackGenerate
+
+  num_heads: int = gin.REQUIRED
+  head_size: int = gin.REQUIRED
+
+  def get_fake_input(self) -> dict[str, Any]:
+    fake_input_dict = super().get_fake_input()
+    b = self.task_config.batch_size
+    n = self.num_heads
+    h = self.head_size
+    fake_input_dict.update({
+        'dstate': tuple(
+            [{
+                'current_index': jnp.array([0] * b, dtype=jnp.int32),
+                'keys': jnp.zeros((b, 2048, n, h), dtype=jnp.float32),
+                'values': jnp.zeros((b, 2048, n, h), dtype=jnp.float32),
+                'recurrent_kvq': None,
+                'relative_position_bias': jnp.zeros(
+                    (b, n, 1, 1024), dtype=jnp.float32
+                ),
+            }]
+            * 12
+        ),
+        'eos': jnp.zeros([1024], dtype=jnp.float32),
+        'mask': jnp.ones([1024], dtype=jnp.float32),
+        'length': 1,
+        'temperature': 1.0,
+    })
+    return fake_input_dict
+
+  def __call__(self, inputs: ...) -> tuple[Any, dict[str, Any]]:
+    # Make sure this code is not used on untested cases.
+    if self.mode not in ['init', 'beam_search']:
+      raise ValueError(f'{type(self)} cannot do mode {self.mode}')
+    if self.decoder.supports_generate():
+      raise ValueError(f'{type(self)}.decoder cannot supports_generate()')
+
+    self.decoder(
+        input_tokens=inputs['targets'][:, 0:1],
+        target_tokens=None,
+        start_of_sequence=inputs['start_of_sequence'],
+    )
+
+    b = inputs['targets'].shape[0]
+    no_start_of_seq = jnp.array([False] * b, dtype=jnp.bool_)
+
+    # This fn is used in both beam_search or topk_sampling.
+    def tokens_to_logits_fn(
+        input_token: jnp.ndarray, dstate: tuple[dict[str, jnp.ndarray], ...]
+    ) -> tuple[jnp.ndarray, tuple[dict[str, jnp.ndarray], ...]]:
+      (logits, dstate, _) = self.decoder(
+          input_tokens=input_token,
+          target_tokens=None,
+          start_of_sequence=no_start_of_seq,
+          decoder_state=dstate,
+      )
+      return logits[:, -1, :], dstate
+
+    last_token = jax.lax.dynamic_slice_in_dim(
+        inputs['targets'], inputs['length'] - 1, 1, axis=1
+    )
+
+    # last token is used to seed beam_search
+    inputs['targets'] = inputs['targets'][:, 0:-1]
+    dstate = jax.lax.cond(
+        inputs['start_of_sequence'][0],
+        lambda: self.generate(inputs)[0],
+        lambda: inputs['dstate'],
+    )
+
+    # Then we run beam search, init with last_token & dstate.
+    finished_seqs, finished_scores, dstate = beam_search.beam_search_flat(
+        last_token,
+        dstate,
+        tokens_to_logits_fn,
+        max_decode_len=512,
+        eos=inputs['eos'].reshape((1, 1, -1)),
+        mask=inputs['mask'].reshape((1, 1, -1)),
+    )
+
+    return 0.0, {
+        'finished_seqs': finished_seqs,
+        'finished_scores': finished_scores,
+        'dstate': dstate,
+    }
+
+  def generate(
+      self, inputs: ...
+  ) -> tuple[tuple[dict[str, jnp.ndarray, ...], ...], jnp.ndarray]:
+    """Generate an output sequence.
+
+    Args:
+      inputs: the same as argument to _call_.
+
+    Returns:
+      An array of generated tokens of shape (batch_size, sequence_length).
+    """
+    input_tokens = inputs['targets']  # [b,seq_len]
+    start_of_sequence = inputs['start_of_sequence']  # [b]
+    target_tokens = jnp.pad(input_tokens[:, 1:], [(0, 0), (0, 1)])
+    batch_size = target_tokens.shape[0]
+
+    # Assuming all sequences start at the same time.
+    start0 = inputs['start_of_sequence'][0]
+    dstate = jax.lax.cond(
+        start0,
+        lambda: self.decoder.init_decoder_state_vanilla(  # pylint: disable=g-long-lambda
+            1024, start_of_sequence
+        ),
+        lambda: inputs['dstate'],
+    )
+
+    first_token = input_tokens[:, 0:1]
+    no_start_of_seq = jnp.array([False] * batch_size, dtype=jnp.bool_)
+    temperature = 1
+    if 'temperature' in inputs:
+      temperature = inputs['temperature']
+
+    num_steps = inputs['length']
+    if self.mode == 'beam_search':
+      num_steps -= 1
+
+    def cond_fn(scan_state) -> jnp.bool_:
+      _, _, i, _ = scan_state
+      return i < num_steps
+
+    def loop_fn(scan_state: Any) -> Tuple[Any, Any, Any, Any]:
+      (dstate, input_token, i, _) = scan_state
+
+      (logits, dstate, _) = self.decoder(
+          input_tokens=input_token,
+          target_tokens=None,
+          start_of_sequence=no_start_of_seq,
+          decoder_state=dstate,
+      )
+
+      logits = logits / temperature
+      output_token = jax.lax.dynamic_slice_in_dim(target_tokens, i, 1, axis=1)
+
+      return (dstate, output_token, i + 1, logits)
+
+    # Scan over the sequence length.
+    dummy_logits = jnp.zeros((batch_size, 1, 1024))
+    initial_scan_state = (dstate, first_token, 0, dummy_logits)
+    dstate, _, _, logits = jax.lax.while_loop(
+        cond_fn, loop_fn, initial_scan_state
+    )
+    return dstate, logits