Fix LoRA buffer contamination during adapter eviction (#8103)

python/sglang/srt/lora/mem_pool.py

@@ -188,10 +188,18 @@ class LoRAMemoryPool:
         lora_adapter: LoRAAdapter,
         lora_modules: Dict[int, Dict[str, BaseLayerWithLoRA]],
     ):
-        def check_lora_weight_shape(buffer_view: torch.Tensor, weight: torch.Tensor):
+        def load_lora_weight_tensor(
-            assert (
+            buffer_view: torch.Tensor, weight: Optional[torch.Tensor]
-                buffer_view.shape == weight.shape
+        ):
-            ), f"LoRA buffer shape {buffer_view.shape} does not match weight shape {weight.shape}."
+            if weight is None:
+                # If the particular weight is not present in the adapter, we initialize the buffer to zero
+                # to avoid contamination from the residual weight of the evicted adapters.
+                buffer_view.zero_()
+            else:
+                assert (
+                    buffer_view.shape == weight.shape
+                ), f"LoRA buffer shape {buffer_view.shape} does not match weight shape {weight.shape}."
+                buffer_view.copy_(weight)
 
         if uid is None:
             for i in range(self.num_layer):
@@ -203,8 +211,12 @@ class LoRAMemoryPool:
         lora_rank = lora_adapter.config.hf_config["r"]
         for layer_id in range(self.num_layer):
             layer_weights = lora_adapter.layers[layer_id].weights
-            temp_A_buffer: Dict[str, torch.Tensor] = {}
+            temp_A_buffer: Dict[str, Optional[torch.Tensor]] = {
-            temp_B_buffer: Dict[str, torch.Tensor] = {}
+                weight_name: None for weight_name in self.A_buffer
+            }
+            temp_B_buffer: Dict[str, Optional[torch.Tensor]] = {
+                weight_name: None for weight_name in self.B_buffer
+            }
             for name, weights in layer_weights.items():
                 if "lora_A" in name:
                     lora_weight_name = get_weight_name(
@@ -220,6 +232,14 @@ class LoRAMemoryPool:
             if self.tp_size > 1:
                 cur_layer_modules = lora_modules[layer_id]
                 for module_name, module in cur_layer_modules.items():
+                    weight_name = get_weight_name(
+                        module_name, self.lora_weight_names, LoRAType.LORA_A
+                    )
+
+                    if temp_A_buffer[weight_name] is None:
+                        # Skip weight slicing if the weight is not present in the adapter
+                        continue
+
                     if "qkv_proj" in module_name:
                         temp_A_buffer["qkv_proj"] = module.slice_lora_a_weights(
                             temp_A_buffer["qkv_proj"], self.tp_rank
@@ -231,9 +251,10 @@ class LoRAMemoryPool:
                             )
                         )
                     else:
-                        weight_name = get_weight_name(
+                        # TODO (lifuhuang): Ideally, we should call `get_weight_name` separately for both A and B.
-                            module_name, self.lora_weight_names, LoRAType.LORA_A
+                        # Currently, we're reusing A's weight name as a workaround, relying on the fact that A and
-                        )
+                        # B share the same name except for `qkv_proj`. We should clean this up once we deprecate the
+                        # FlashInfer LoRA backend.
                         temp_A_buffer[weight_name] = module.slice_lora_a_weights(
                             temp_A_buffer[weight_name], self.tp_rank
                         )
@@ -246,8 +267,7 @@ class LoRAMemoryPool:
                 buffer_view = self.A_buffer[name][layer_id][buffer_id][
                     : lora_rank * c, :
                 ]
-                check_lora_weight_shape(buffer_view, weights)
+                load_lora_weight_tensor(buffer_view, weights)
-                buffer_view.copy_(weights)
 
             for name, weights in temp_B_buffer.items():
                 c = get_stacked_multiply(name)
@@ -256,14 +276,15 @@ class LoRAMemoryPool:
                         buffer_view = self.B_buffer[name][layer_id][stacked_id][
                             buffer_id
                         ][:, :lora_rank]
-                        check_lora_weight_shape(buffer_view, weights[stacked_id])
+                        weight_slice = (
-                        buffer_view.copy_(weights[stacked_id])
+                            weights[stacked_id] if weights is not None else None
+                        )
+                        load_lora_weight_tensor(buffer_view, weight_slice)
                 else:
                     buffer_view = self.B_buffer[name][layer_id][0][buffer_id][
                         :, :lora_rank
                     ]
-                    check_lora_weight_shape(buffer_view, weights)
+                    load_lora_weight_tensor(buffer_view, weights)
-                    buffer_view.copy_(weights)
 
     def get_tensor(
         self, weight_name: str, layer_id: int, lora_type: LoRAType

test/srt/models/lora/test_lora_eviction.py

@@ -0,0 +1,111 @@
+# Copyright 2023-2024 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+import multiprocessing as mp
+import unittest
+from typing import Dict, List, Tuple
+
+import torch
+
+from sglang.test.runners import SRTRunner
+from sglang.test.test_utils import CustomTestCase
+
+PROMPTS = [
+    "AI is a field of computer science focused on",
+    """
+    ### Instruction:
+    Compose a SQL query that uses the following table: users, and returns the user_id and name of all users whose name that does not have a duplicate in the table.
+    ### Response:
+    SELECT user_id, name FROM users WHERE name LIKE 'A%';
+    """,
+]
+
+ADAPTERS = [
+    "faridlazuarda/valadapt-llama-3.1-8B-it-chinese",  # target_modules = q, v
+    "philschmid/code-llama-3-1-8b-text-to-sql-lora",  # target_modules = q, k, v, o, gate, up, down
+]
+
+BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B-Instruct"
+
+
+class TestLoRAEviction(CustomTestCase):
+    def test_lora_eviction_with_different_target_modules(self):
+        """
+        Test LoRA eviction with different target modules.
+
+        This test runs inference against two LoRA adapters in different orders to force eviction behavior, and ensures
+        that the outputs of the same (adapter, prompt) pair are consistent across runs.
+        """
+        output_history = {}
+        self._run_test(ADAPTERS, output_history, reverse=False)
+        self._run_test(ADAPTERS, output_history, reverse=True)
+
+    def _run_test(
+        self,
+        lora_paths: List[str],
+        output_history: Dict[Tuple[str, str], str],
+        reverse: bool,
+        repeat: int = 2,
+    ):
+        max_new_tokens = 256
+        backend = "triton"
+        torch_dtype = torch.float16
+        base_path = BASE_MODEL
+        assert len(lora_paths) >= 2
+
+        # Initialize runners
+        with SRTRunner(
+            base_path,
+            torch_dtype=torch_dtype,
+            model_type="generation",
+            lora_paths=lora_paths,
+            max_loras_per_batch=1,
+            lora_backend=backend,
+            disable_radix_cache=True,
+        ) as srt_runner:
+            adapter_sequence = lora_paths if not reverse else lora_paths[::-1]
+
+            for i in range(repeat):
+                for j, adapter in enumerate(adapter_sequence):
+                    print(
+                        f"\n========== Testing LoRA eviction with adapter '{adapter}' (#{j+1}/{len(adapter_sequence)}), reversed: {reverse}, repeat: {i+1}/{repeat} ---"
+                    )
+                    for prompt in PROMPTS:
+                        print("\nprompt:\n", prompt)
+                        srt_outputs = srt_runner.forward(
+                            [prompt],
+                            max_new_tokens=max_new_tokens,
+                            lora_paths=[adapter],
+                        )
+                        output = srt_outputs.output_strs[0].strip()
+                        print("\noutput:\n", output)
+
+                        prev_output = output_history.get((adapter, prompt))
+                        if prev_output is not None:
+                            self.assertEqual(
+                                prev_output,
+                                output,
+                                f"Output mismatch for adapter {adapter} and prompt '{prompt}' on repeat {j + 1}, previous: '{prev_output}', current: '{output}'.",
+                            )
+                        else:
+                            output_history[(adapter, prompt)] = output
+
+
+if __name__ == "__main__":
+    try:
+        mp.set_start_method("spawn")
+    except RuntimeError:
+        pass
+
+    unittest.main(warnings="ignore")

test/srt/run_suite.py

@@ -14,6 +14,7 @@ class TestFile:
 suites = {
     "per-commit": [
         TestFile("models/lora/test_lora.py", 200),
+        TestFile("models/lora/test_lora_eviction.py", 120),
         TestFile("models/lora/test_lora_backend.py", 99),
         TestFile("models/lora/test_multi_lora_backend.py", 60),
         TestFile("models/lora/test_lora_cuda_graph.py", 250),