[Misc][LoRA] Fix LoRA weight mapper (#11495)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

tests/lora/test_lora_checkpoints.py

@@ -74,7 +74,7 @@ def test_load_checkpoints(
                 embedding_padding_modules=embed_padding_modules)
 
 
-def test_lora_weights_mapping(baichuan_lora_files, ):
+def test_lora_weights_mapping(baichuan_lora_files):
     supported_lora_modules = BaiChuanBaseForCausalLM.supported_lora_modules
     packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
     embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
@@ -86,10 +86,14 @@ def test_lora_weights_mapping(baichuan_lora_files, ):
         else:
             expected_lora_modules.append(module)
 
-    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
-        "model.": "language_model.model.",
-    }, )
-
+    hf_to_vllm_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            "model.": "language_model.model.",
+        },
+        orig_to_new_substr={
+            ".layers.": ".baichuan_layers.",
+        },
+    )
     lora_model = LoRAModel.from_local_checkpoint(
         baichuan_lora_files,
         expected_lora_modules,
@@ -101,3 +105,4 @@ def test_lora_weights_mapping(baichuan_lora_files, ):
     )
     for name in lora_model.loras:
         assert name.startswith(hf_to_vllm_mapper.orig_to_new_prefix["model."])
+        assert ".baichuan_layers." in name

tests/lora/test_qwen2vl.py

@@ -22,7 +22,7 @@ IMAGE_ASSETS = [
 
 # After fine-tuning with LoRA, all generated content should start begin `A`.
 EXPECTED_OUTPUT = [
-    "A stop sign stands prominently in the foreground, with a traditional Chinese gate and a black SUV in the background, illustrating a blend of modern and cultural elements.",  # noqa: E501
+    "A red stop sign stands prominently in the foreground, with a traditional Chinese gate and a black SUV in the background, illustrating a blend of modern and cultural elements.",  # noqa: E501
     "A majestic skyscraper stands tall, partially obscured by a vibrant canopy of cherry blossoms, against a clear blue sky.",  # noqa: E501
 ]
 
@@ -76,3 +76,7 @@ def test_qwen2vl_lora(qwen2vl_lora_files):
     output1 = do_sample(llm, qwen2vl_lora_files, lora_id=1)
     for i in range(len(EXPECTED_OUTPUT)):
         assert EXPECTED_OUTPUT[i].startswith(output1[i])
+
+    output2 = do_sample(llm, qwen2vl_lora_files, lora_id=2)
+    for i in range(len(EXPECTED_OUTPUT)):
+        assert EXPECTED_OUTPUT[i].startswith(output2[i])

vllm/lora/models.py

@@ -231,7 +231,8 @@ class LoRAModel(AdapterModel):
             with safetensors.safe_open(lora_tensor_path,
                                        framework="pt") as f:  # type: ignore
                 for lora_module in f.keys():  # noqa
-                    module_name, _, _ = parse_fine_tuned_lora_name(lora_module)
+                    module_name, _, _ = parse_fine_tuned_lora_name(
+                        lora_module, weights_mapper)
                     part_name = module_name.split(".")[-1]
                     if part_name not in expected_lora_modules:
                         unexpected_modules.append(module_name)

vllm/lora/utils.py

@@ -1,4 +1,3 @@
-import copy
 import os
 import re
 from typing import List, Optional, Set, Tuple, Type, Union
@@ -32,7 +31,6 @@ from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.models.utils import WeightsMapper
-from vllm.utils import print_warning_once
 
 logger = init_logger(__name__)
 
@@ -112,36 +110,28 @@ def parse_fine_tuned_lora_name(
             is_bias whether the tensor is lora bias.
     """
 
-    w_mapper = None
-    if weights_mapper:
-        w_mapper = copy.deepcopy(weights_mapper)
-        # TODO: Currently only supports mapping for prefix, mapping for
-        # substr and subfix will be supported in the future.
-        for attr, mapping in [
-            ("orig_to_new_substr", w_mapper.orig_to_new_substr),
-            ("orig_to_new_suffix", w_mapper.orig_to_new_suffix),
-        ]:
-            if mapping:
-                print_warning_once(
-                    f"vLLM currently does not support mapping of LoRA weights "
-                    f"for {mapping}.")
-                setattr(w_mapper, attr, {})
+    # LoRA weight qualified name always starts with `base_model.model.`,
+    # so we remove the prefix `base_model.model.` to make the following
+    # mapping correctly.
+    if "base_model.model." in name:
+        name = name.replace("base_model.model.", "")
+        name = weights_mapper._map_name(name) if weights_mapper else name
+        # recover the prefix `base_model.model.`
+        name = "base_model.model." + name
 
-    mapper = (lambda name: w_mapper._map_name(name)
-              if w_mapper is not None else name)
     parts = name.split(".")
     if parts[-1] == "weight" and (parts[-2] == "lora_A"
                                   or parts[-2] == "lora_B"):
         new_name = ".".join(parts[2:-2])
-        return mapper(new_name), parts[-2] == "lora_A", False
+        return new_name, parts[-2] == "lora_A", False
 
     if parts[-1] == "lora_embedding_A" or parts[-1] == "lora_embedding_B":
         new_name = ".".join(parts[2:-1])
-        return mapper(new_name), parts[-1] == "lora_embedding_A", False
+        return new_name, parts[-1] == "lora_embedding_A", False
 
     if parts[-1] == "bias":
         new_name = ".".join(parts[2:-2])
-        return mapper(new_name), False, True
+        return new_name, False, True
 
     raise ValueError(f"{name} is unsupported LoRA weight")
 

vllm/lora/worker_manager.py

@@ -91,6 +91,8 @@ class WorkerLoRAManager(AbstractWorkerManager):
                         packed_modules_mapping[module])
                 else:
                     expected_lora_modules.append(module)
+
+            expected_lora_modules = list(set(expected_lora_modules))
             lora_path = get_adapter_absolute_path(lora_request.lora_path)
 
             # For some models like Qwen2VL, we need to use hf_to_vllm_mapper