Merge pull request #397 from beep39/json-schema-for-exllamav3

Constrained generation with json schema for ExllamaV3

backends/exllamav3/grammar.py

@@ -0,0 +1,57 @@
+from typing import List
+import traceback
+
+from exllamav3 import (
+    Tokenizer,
+    Filter,
+    FormatronFilter,
+)
+from formatron.formatter import FormatterBuilder
+from formatron.schemas import json_schema
+from loguru import logger
+
+
+class ExLlamaV3Grammar:
+    """ExLlamaV3 class for various grammar filters/parsers."""
+
+    filters: List[Filter]
+
+    def __init__(self):
+        self.filters = []
+
+    def add_json_schema_filter(
+        self,
+        schema: dict,
+        tokenizer: Tokenizer,
+    ):
+        """Adds an ExllamaV3 filter based on a JSON schema."""
+
+        leading_character = "[" if schema.get("type") == "array" else "{"
+
+        try:
+            # Add fields required by formatron if not present
+            if "$id" not in schema:
+                schema["$id"] = "https://example.com/example.json"
+            if "$schema" not in schema:
+                schema["$schema"] = "http://json-schema.org/draft-07/schema#"
+
+            # Validate schema and create formatter
+            schema = json_schema.create_schema(schema)
+        except Exception:
+            traceback.print_exc()
+            logger.error(
+                "Skipping because the JSON schema couldn't be parsed. "
+                "Please read the above error for more information."
+            )
+            return
+
+        f = FormatterBuilder()
+        f.append_line(f"{f.json(schema)}")
+        self.filters.append(
+            FormatronFilter(tokenizer, eos_after_completed=True, formatter_builder=f)
+        )
+
+        # Additional constraint to force leading character
+        f = FormatterBuilder()
+        f.append_line(leading_character)
+        self.filters.append(FormatronFilter(tokenizer, formatter_builder=f))

backends/exllamav3/model.py

@@ -21,6 +21,7 @@ from exllamav3 import (
     Tokenizer,
 )
 from exllamav3.cache import CacheLayer_quant
+from backends.exllamav3.grammar import ExLlamaV3Grammar
 from loguru import logger
 
 from backends.base_model_container import BaseModelContainer
@@ -933,6 +934,7 @@ class ExllamaV3Container(BaseModelContainer):
         prompts = [prompt]
         stop_conditions = params.stop
         add_bos_token = unwrap(params.add_bos_token, self.hf_model.add_bos_token())
+        grammar_handler = ExLlamaV3Grammar()
 
         # Get multimodal embeddings if present
         mm_embeddings_content = mm_embeddings.content if mm_embeddings else []
@@ -974,6 +976,9 @@ class ExllamaV3Container(BaseModelContainer):
             request_id,
         )
 
+        if params.json_schema:
+            grammar_handler.add_json_schema_filter(params.json_schema, self.tokenizer)
+
         generation = {}
         job = AsyncJob(
             self.generator,
@@ -985,6 +990,7 @@ class ExllamaV3Container(BaseModelContainer):
             embeddings=mm_embeddings_content,
             return_top_tokens=params.logprobs,
             max_rq_tokens=self.max_rq_tokens,
+            filters=grammar_handler.filters,
         )
 
         generated_tokens = 0