Use tempfile for batch

rl/llm/engines/client.py

@@ -1,8 +1,9 @@
-import io
 import json
 import re
+import tempfile
 import time
 from abc import ABC, abstractmethod
+from pathlib import Path
 from typing import TYPE_CHECKING
 
 import tqdm
@@ -35,7 +36,7 @@ class ClientEngine(InferenceEngine, ABC):
     def generate(self, prompt: ChatInput) -> InferenceOutput:
         pass
 
-    def batch_generate(self, prompts: list[ChatInput]) -> InferenceOutput:
+    def batch_generate(self, prompts: list[ChatInput]) -> list[InferenceOutput]:
         return thread_map(
             self.generate,
             prompts,
@@ -342,38 +343,43 @@ def generate(self, prompt: ChatInput) -> InferenceOutput:
         return self.batch_generate([prompt])[0]
 
     def batch_generate(self, prompts: list[ChatInput]) -> list[InferenceOutput]:
-        # Create in-memory JSONL file
-        jsonl_file = io.StringIO()
-        for i, prompt in enumerate(prompts):
-            body_kwargs = {
-                "model": self.llm_config.model_name_or_path,
-                "messages": prompt,
-            }
-            if self.llm_config.max_new_tokens is not None:
-                body_kwargs["max_tokens"] = self.llm_config.max_new_tokens
-            if self.llm_config.temperature is not None:
-                body_kwargs["temperature"] = self.llm_config.temperature
-            if EngineFeature.JSON_OUTPUT in self.enabled_features:
-                body_kwargs["response_format"] = {"type": "json_object"}
-            request = {
-                "custom_id": f"request-{i}",
-                "method": "POST",
-                "url": "/v1/chat/completions",
-                "body": body_kwargs,
-            }
-            jsonl_file.write(json.dumps(request) + "\n")
-
-        jsonl_file.seek(0)
-        batch_input_file = self.client.files.create(file=jsonl_file, purpose="batch")
+        with tempfile.NamedTemporaryFile(
+            mode="w+", suffix=".jsonl", delete=False
+        ) as temp_file:
+            for i, prompt in enumerate(prompts):
+                body_kwargs = {
+                    "model": self.llm_config.model_name_or_path,
+                    "messages": prompt,
+                }
+                if self.llm_config.max_new_tokens is not None:
+                    body_kwargs["max_tokens"] = self.llm_config.max_new_tokens
+                if self.llm_config.temperature is not None:
+                    body_kwargs["temperature"] = self.llm_config.temperature
+                if EngineFeature.JSON_OUTPUT in self.enabled_features:
+                    body_kwargs["response_format"] = {"type": "json_object"}
+                request = {
+                    "custom_id": f"request-{i}",
+                    "method": "POST",
+                    "url": "/v1/chat/completions",
+                    "body": body_kwargs,
+                }
+                json.dump(request, temp_file)
+                temp_file.write("\n")
+            temp_file_path = temp_file.name
+
+        # Upload file
+        with Path(temp_file_path).open("rb") as file:
+            batch_input_file = self.client.files.create(file=file, purpose="batch")
 
+        # Create batch
         batch = self.client.batches.create(
             input_file_id=batch_input_file.id,
             endpoint="/v1/chat/completions",
             completion_window="24h",
         )
 
         # Poll for status and update progress bar
-        pbar = tqdm.tqdm(total=len(prompts), desc="Polling batch status")
+        pbar = tqdm(total=len(prompts), desc="Polling batch status")
         while batch.status not in ["completed", "failed", "expired"]:
             time.sleep(5)  # Poll every 5 seconds
             batch = self.client.batches.retrieve(batch.id)