first

.gitignore

@@ -0,0 +1,4 @@
+*.pt
+*.egg-info
+__pycache__
+wip

LICENSE.md

@@ -0,0 +1,9 @@
+MIT LICENSE
+
+Copyright 2024 Kyle Finn
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

README.md

@@ -0,0 +1,16 @@
+# torch2cpp
+
+Status: WIP
+
+Some features are supported, but most are not yet.
+
+Models are traced with torch.fx to extract the flattened AST of low level function calls.
+
+The AST graph is then traversed, generating C++ code in order.
+
+Temporary buffer shapes and ref-counts are tracked to enable compile-time scheduled memory re-use (~10x buffer reduction in common cases).
+
+Weights are stored in the binary as bfloat16, and unpacked to float32 at runtime. (Would like to investigate more options here for both storage and inference)
+
+The bundled tensor math lib uses compile-time shapes and in-place storage, so there is no dynamic memory allocation at all.
+

example/Makefile

@@ -0,0 +1,16 @@
+INCLUDES := $(shell python -m torch2cpp.includes)
+
+build/model.js : build/model.cpp
+	em++ -Os build/model.cpp -I$(INCLUDES) \
+		-o build/model.js -s MODULARIZE=1 -s EXPORT_NAME=load_model \
+		-s EXPORTED_FUNCTIONS=_model_step,_model_reset,_model_encode,_model_decode
+
+build/chat_cli : chat_cli.cpp build/model.cpp 
+	c++ -std=c++17 -Os -march=native -ffast-math \
+		build/model.cpp chat_cli.cpp -I$(INCLUDES) -o build/chat_cli
+
+.PHONY: model.js
+model.js: build/model.js
+
+.PHONY: chat_cli
+chat_cli: build/chat_cli

example/build/.gitignore

@@ -0,0 +1,2 @@
+*
+!.gitignore

example/chat_cli.cpp

@@ -0,0 +1,42 @@
+#include <string>
+#include <iostream>
+
+extern "C" {
+void model_reset();
+int model_step(int prevtok, float temperature);
+int model_encode(char const* str, int str_len, int * out, int out_len);
+int model_decode(int const* toks, int toks_len, char * out, int out_len);
+} // extern C
+
+int main(int argc, char ** argv)
+{
+    std::string prompt;
+    constexpr int max_tokens = 256;
+    int toks[max_tokens];
+    char decode[max_tokens];
+    float temperature = 1;
+    while(true)
+    {
+        std::getline(std::cin, prompt);
+        prompt += "\n";
+
+        int n_tok = model_encode(prompt.c_str(), prompt.size(), toks, max_tokens);
+        for(int i=0 ; i<n_tok-1 ; i++)
+        {
+            model_step(toks[i], 0);
+        }
+        for(int i=n_tok-1 ; i<max_tokens ; i++)
+        {
+            toks[i+1] = model_step(toks[i], temperature);
+            int strlen = model_decode(toks+i+1, 1, decode, max_tokens);
+            std::cout << std::string(decode, strlen);
+            bool exit = false;
+            for(int j=0 ; j<strlen ; j++)
+            {
+                exit |= (decode[j]=='\n');
+            }
+            if(exit) { break; }
+        }
+        std::cout << std::endl;
+    }
+}

example/sqrll2cpp.ipynb

@@ -0,0 +1,107 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "params=11,698,432\n"
+     ]
+    }
+   ],
+   "source": [
+    "import torch\n",
+    "from sqrll.sqrllm import SqrLLM\n",
+    "\n",
+    "model_file = '../../sqrll/example/models/model2048bpe.pt'\n",
+    "\n",
+    "model = SqrLLM.load(model_file).eval()\n",
+    "\n",
+    "params = sum(p.numel() for p in model.parameters())\n",
+    "print(f'{params=:,}')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "n_vocab=2048\n"
+     ]
+    }
+   ],
+   "source": [
+    "from tokenizers import Tokenizer\n",
+    "\n",
+    "token_file = '../../sqrll/example/models/tokenizer2048.json'\n",
+    "\n",
+    "tokenizer = Tokenizer.from_file(token_file)\n",
+    "n_vocab = tokenizer.get_vocab_size()\n",
+    "print(f'{n_vocab=}')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from torch2cpp import codegen\n",
+    "from sqrll import sqrll, sqrllm\n",
+    "\n",
+    "inputs = torch.tensor([[0]])\n",
+    "_, mem = model(inputs)\n",
+    "mem = [torch.zeros_like(m) for m in mem]\n",
+    "\n",
+    "with open('build/model.cpp', 'w') as out_file:\n",
+    "    codegen(\n",
+    "        model,\n",
+    "        out_file,\n",
+    "        args=[inputs, mem],\n",
+    "        tokenizer=tokenizer,\n",
+    "        autowrap_functions=[\n",
+    "            sqrll.sqrll_kernel,\n",
+    "            sqrllm.rms_norm,\n",
+    "        ],\n",
+    "        skip_weights=False,\n",
+    "    )"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": ".venv",
+   "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": 2
+}

pyproject.toml

@@ -0,0 +1,7 @@
+[project]
+name = "torch2cpp"
+version = "0.0.1"
+
+dependencies = [
+    "torch",
+]

setup.py

@@ -0,0 +1,2 @@
+import setuptools
+setuptools.setup()

src/torch2cpp/__init__.py

@@ -0,0 +1,2 @@
+
+from .torch2cpp import codegen