Fix memory issue with llama.cpp LLM pipeline, closes #824

src/python/txtai/pipeline/llm/llama.py

@@ -8,7 +8,7 @@
 
 # Conditional import
 try:
-    from llama_cpp import Llama
+    import llama_cpp as llama
 
     LLAMA_CPP = True
 except ImportError:
@@ -45,11 +45,8 @@ def __init__(self, path, template=None, **kwargs):
         # Check if this is a local path, otherwise download from the HF Hub
         path = path if os.path.exists(path) else self.download(path)
 
-        # Default GPU layers if not already set
-        kwargs["n_gpu_layers"] = kwargs.get("n_gpu_layers", -1 if kwargs.get("gpu", os.environ.get("LLAMA_NO_METAL") != "1") else 0)
-
         # Create llama.cpp instance
-        self.llm = Llama(path, n_ctx=0, verbose=kwargs.pop("verbose", False), **kwargs)
+        self.llm = self.create(path, **kwargs)
 
     def stream(self, texts, maxlength, stream, stop, **kwargs):
         for text in texts:
@@ -79,6 +76,39 @@ def download(self, path):
         # Download and cache file
         return hf_hub_download(repo_id="/".join(parts[:repo]), filename="/".join(parts[repo:]))
 
+    def create(self, path, **kwargs):
+        """
+        Creates a new llama.cpp model instance.
+
+        Args:
+            path: path to model
+            kwargs: additional keyword args
+
+        Returns:
+            llama.cpp instance
+        """
+
+        # Default n_ctx=0 if not already set. This sets n_ctx = n_ctx_train.
+        kwargs["n_ctx"] = kwargs.get("n_ctx", 0)
+
+        # Default GPU layers if not already set
+        kwargs["n_gpu_layers"] = kwargs.get("n_gpu_layers", -1 if kwargs.get("gpu", os.environ.get("LLAMA_NO_METAL") != "1") else 0)
+
+        # Default verbose flag
+        kwargs["verbose"] = kwargs.get("verbose", False)
+
+        # Create llama.cpp instance
+        try:
+            return llama.Llama(model_path=path, **kwargs)
+        except ValueError as e:
+            # Fallback to default n_ctx when not enough memory for n_ctx = n_ctx_train
+            if not kwargs["n_ctx"]:
+                kwargs.pop("n_ctx")
+                return llama.Llama(model_path=path, **kwargs)
+
+            # Raise exception if n_ctx manually specified
+            raise e
+
     def messages(self, messages, maxlength, stream, stop, **kwargs):
         """
         Processes a list of messages.

test/python/testpipeline/testaudio/testmicrophone.py

@@ -70,7 +70,7 @@ def int16(self, data):
                 return (data * absmax + offset).clip(i.min, i.max).astype(np.int16)
 
         # Mock input stream
-        inputstream.return_value = RawInputStream()
+        inputstream.side_effect = RawInputStream
 
         # Create microphone pipeline and read data
         pipeline = Microphone()

test/python/testpipeline/testllm/testllama.py

@@ -4,6 +4,8 @@
 
 import unittest
 
+from unittest.mock import patch
+
 from txtai.pipeline import LLM
 
 
@@ -12,6 +14,46 @@ class TestLlama(unittest.TestCase):
     llama.cpp tests.
     """
 
+    @patch("llama_cpp.Llama")
+    def testContext(self, llama):
+        """
+        Test n_ctx with llama.cpp
+        """
+
+        class Llama:
+            """
+            Mock llama.cpp instance to test invalid context
+            """
+
+            def __init__(self, **kwargs):
+                if kwargs.get("n_ctx") == 0 or kwargs.get("n_ctx", 0) >= 10000:
+                    raise ValueError("Failed to create context")
+
+                # Save parameters
+                self.params = kwargs
+
+        # Mock llama.cpp instance
+        llama.side_effect = Llama
+
+        # Model to test
+        path = "TheBloke/TinyLlama-1.1B-Chat-v0.3-GGUF/tinyllama-1.1b-chat-v0.3.Q2_K.gguf"
+
+        # Test omitting n_ctx falls back to default settings
+        llm = LLM(path)
+        self.assertNotIn("n_ctx", llm.generator.llm.params)
+
+        # Test n_ctx=0 falls back to default settings
+        llm = LLM(path, n_ctx=0)
+        self.assertNotIn("n_ctx", llm.generator.llm.params)
+
+        # Test n_ctx manually set
+        llm = LLM(path, n_ctx=1024)
+        self.assertEqual(llm.generator.llm.params["n_ctx"], 1024)
+
+        # Mock a value for n_ctx that's too big
+        with self.assertRaises(ValueError):
+            llm = LLM(path, n_ctx=10000)
+
     def testGeneration(self):
         """
         Test generation with llama.cpp