Move JSON grammar -> regex grammar conversion to the router

This change moves the JSON grammar -> regex grammar conversion to the
router by adding a dependency on the `outlines-core` Rust crate. In
contrast to the Python implementation, the conversions are not LRU-cached
since they seem to be fast enough:

simple schema           time:   [5.8293 µs 5.8307 µs 5.8320 µs]
                        change: [-13.166% -12.884% -12.641%] (p = 0.00 < 0.05)
                        Performance has improved.

complex schema          time:   [14.875 µs 14.881 µs 14.887 µs]
                        change: [-2.1637% -1.9914% -1.7852%] (p = 0.00 < 0.05)
                        Performance has improved.

Using the schemas from:
https://github.com/dottxt-ai/outlines-core/blob/main/benchmarks/bench_json_schema.py

Cargo.lock

@@ -3005,6 +3005,17 @@ dependencies = [
  "num-traits",
 ]
 
+[[package]]
+name = "outlines-core"
+version = "0.1.0"
+source = "git+https://github.com/dottxt-ai/outlines-core.git?rev=ba10c619fc9bf3c487e43f49bdecb95a24bb465c#ba10c619fc9bf3c487e43f49bdecb95a24bb465c"
+dependencies = [
+ "anyhow",
+ "regex",
+ "serde-pyobject",
+ "serde_json",
+]
+
 [[package]]
 name = "overload"
 version = "0.1.1"
@@ -3952,6 +3963,16 @@ dependencies = [
  "serde_derive",
 ]
 
+[[package]]
+name = "serde-pyobject"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ca4b0aad8b225845739a0030a0d5cc2ae949c56a86a7daf9226c7df7c2016d16"
+dependencies = [
+ "pyo3",
+ "serde",
+]
+
 [[package]]
 name = "serde_cbor"
 version = "0.11.2"
@@ -3979,6 +4000,7 @@ version = "1.0.133"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c7fceb2473b9166b2294ef05efcb65a3db80803f0b03ef86a5fc88a2b85ee377"
 dependencies = [
+ "indexmap 2.6.0",
  "itoa",
  "memchr",
  "ryu",
@@ -4430,6 +4452,7 @@ dependencies = [
 name = "text-generation-router"
 version = "2.4.2-dev0"
 dependencies = [
+ "anyhow",
  "async-stream",
  "async-trait",
  "axum 0.7.9",
@@ -4453,6 +4476,7 @@ dependencies = [
  "once_cell",
  "opentelemetry 0.20.0",
  "opentelemetry-otlp",
+ "outlines-core",
  "pyo3",
  "rand",
  "regex",

crate-hashes.json

@@ -0,0 +1,3 @@
+{
+  "git+https://github.com/dottxt-ai/outlines-core.git?rev=ba10c619fc9bf3c487e43f49bdecb95a24bb465c#outlines-core@0.1.0": "1j9dcd831b0bmmjk2n4aag3x47qnqmkpg4gqpvwwyic7744llbfm"
+}

router/Cargo.toml

@@ -8,6 +8,7 @@ authors.workspace = true
 homepage.workspace = true
 
 [dependencies]
+anyhow = "1"
 async-trait = "0.1.74"
 async-stream = "0.3.5"
 axum = { version = "0.7", features = ["json"] }
@@ -22,6 +23,7 @@ metrics-exporter-prometheus = { workspace = true }
 nohash-hasher = "0.2.0"
 opentelemetry = { version = "0.20.0", features = ["rt-tokio"] }
 opentelemetry-otlp = "0.13.0"
+outlines-core = { git = "https://github.com/dottxt-ai/outlines-core.git", rev = "ba10c619fc9bf3c487e43f49bdecb95a24bb465c" }
 rand = "0.8.5"
 reqwest = { version = "0.11.20", features = [] }
 serde = "1.0.188"

router/src/validation.rs

@@ -9,6 +9,7 @@ use crate::{PyTokenizer, Tokenizer};
 use base64::{engine::general_purpose::STANDARD, Engine};
 use image::{ImageFormat, ImageReader};
 use jsonschema::{Draft, JSONSchema};
+use outlines_core::json_schema::to_regex as json_schema_to_regex;
 use rand::{thread_rng, Rng};
 use serde_json::Value;
 use std::io::Cursor;
@@ -351,11 +352,13 @@ impl Validation {
                                 "Grammar must have a 'properties' field".to_string(),
                             ))?;
 
-                        // Serialize json to string
-                        ValidGrammar::Json(
-                            serde_json::to_string(&json)
-                                .map_err(|e| ValidationError::InvalidGrammar(e.to_string()))?,
-                        )
+                        // Do compilation in the router for performance. In the future, we
+                        // should also move regex -> automaton compilation in the router,
+                        // but this is not yet supported in pure Rust by outlines-core.
+                        let grammar_regex = json_schema_to_regex(&json, None, &json)
+                            .map_err(ValidationError::RegexFromSchema)?;
+
+                        ValidGrammar::Regex(grammar_regex.to_string())
                     }
                     GrammarType::Regex(regex) => ValidGrammar::Regex(regex),
                 };
@@ -810,6 +813,8 @@ pub enum ValidationError {
     Grammar,
     #[error("grammar is not valid: {0}")]
     InvalidGrammar(String),
+    #[error("cannot compile regex from schema: {0}")]
+    RegexFromSchema(anyhow::Error),
     #[error("base64 encoding is invalid: {0}")]
     InvalidBase64(#[from] base64::DecodeError),
     #[error("invalid image: {0}")]

server/text_generation_server/utils/logits_process.py

@@ -1,19 +1,19 @@
+from functools import lru_cache
 import math
+import time
 import torch
+from typing import List, Optional, DefaultDict
 
 from loguru import logger
 from typing import Dict, Union
 from text_generation_server.pb.generate_pb2 import GrammarType
 
 from outlines.fsm.guide import RegexGuide
-from outlines.fsm.json_schema import build_regex_from_schema
-from functools import lru_cache
-from typing import List, Optional, DefaultDict
-import time
 
 from transformers import (
     LogitsWarper,
     LogitsProcessor,
+    PreTrainedTokenizerBase,
     TemperatureLogitsWarper,
     TopKLogitsWarper,
     TopPLogitsWarper,
@@ -484,7 +484,13 @@ class GrammarLogitProcessor(LogitsProcessor):
     fsm_state: DefaultDict[int, int]
     fsm: RegexGuide
 
-    def __init__(self, tokenizer, device, grammar, grammar_type):
+    def __init__(
+        self,
+        tokenizer: Optional[PreTrainedTokenizerBase],
+        device: str,
+        grammar: str,
+        grammar_type: GrammarType,
+    ):
         self.device = device
         self.tokenizer = GrammarLogitProcessor._cached_adapt_tokenizer(tokenizer)
         self.fsm = GrammarLogitProcessor._cached_compile_fsm(
@@ -519,18 +525,20 @@ class GrammarLogitProcessor(LogitsProcessor):
     # TODO: move grammar compilation into the router
     @staticmethod
     @lru_cache(maxsize=32, typed=True)
-    def _cached_compile_fsm(grammar_type, schema, tokenizer):
+    def _cached_compile_fsm(
+        grammar_type: GrammarType,
+        schema: str,
+        tokenizer: Optional[PreTrainedTokenizerBase],
+    ):
         start_time = time.time()
         if grammar_type == GrammarType.GRAMMAR_TYPE_JSON:
-            try:
-                schema = build_regex_from_schema(schema)
-            # TODO: this is only here short term to avoid crashing the python server, mid term we want this in the rust/router layer
-            except Exception as e:
-                logger.error(f"Error compiling FSM, grammar won't be enforced \n{e}")
-                # allows everything
-                schema = "(.*?)"
-        elif grammar_type == GrammarType.GRAMMAR_TYPE_REGEX:
-            pass  # schema is already a regex just here for clarity
+            # JSON schema is compiled by the v3 router.
+            logger.error(
+                "Non-regex grammars must be compiled by the router, grammar won't be enforced"
+            )
+            # allows everything
+            schema = "(.*?)"
+
         fsm = RegexGuide.from_regex(schema, tokenizer)
         logger.debug(f"Compiled FSM in {time.time() - start_time:.2f}s")
         return fsm