From fe24e6084d1357f0ce74e93d56962a490b3395c8 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Sat, 6 Jul 2024 15:39:19 -0700
Subject: [PATCH 01/10] Support loading loras from huggingface in runtime

---
 vllm/lora/utils.py          | 42 +++++++++++++++++++++++++++++++++++++
 vllm/lora/worker_manager.py |  6 ++++--
 2 files changed, 46 insertions(+), 2 deletions(-)

diff --git a/vllm/lora/utils.py b/vllm/lora/utils.py
index ab3b99eee6f..8e8690087d1 100644
--- a/vllm/lora/utils.py
+++ b/vllm/lora/utils.py
@@ -1,5 +1,8 @@
+import os
 from typing import List, Optional, Set, Tuple, Type
 
+from huggingface_hub import snapshot_download
+from huggingface_hub.utils import RepositoryNotFoundError, EntryNotFoundError, HFValidationError
 from torch import nn
 from transformers import PretrainedConfig
 
@@ -105,3 +108,42 @@ def parse_fine_tuned_lora_name(name: str) -> Tuple[str, bool]:
             return ".".join(parts[2:-1]), parts[-1] == "lora_embedding_A"
 
     raise ValueError(f"{name} is unsupported LoRA weight")
+
+def get_lora_absolute_path(lora_path: str) -> str:
+    """
+    Resolves the given lora_path to an absolute local path.
+
+    If the lora_path is identified as a Hugging Face model identifier, it will download
+    the model and return the local snapshot path. Otherwise, it treats the lora_path
+    as a local file path and converts it to an absolute path.
+
+    Parameters:
+    lora_path (str): The path to the lora model, which can be an absolute path,
+                     a relative path, or a Hugging Face model identifier.
+
+    Returns:
+    str: The resolved absolute local path to the lora model.
+    """
+
+     # Check if the path is an absolute path. Return it no matter existing or not.
+    if os.path.isabs(lora_path):
+        return lora_path
+
+    # If the path starts with ~, expand the user home directory and return the path.
+    if lora_path.startswith('~'):
+        return os.path.expanduser(lora_path)
+
+    # Check if the expanded relative path exists locally.
+    if os.path.exists(lora_path):
+        return os.path.abspath(lora_path)
+
+    # If the path does not exist locally, assume it's a Hugging Face model identifier.
+    try:
+        local_snapshot_path = snapshot_download(repo_id=lora_path)
+    except (RepositoryNotFoundError, EntryNotFoundError, HFValidationError) as e:
+        # Handle errors that may occur during the download
+        # Return original path instead instead of throwing error here
+        print(f"Error downloading the Hugging Face model: {e}")
+        return lora_path
+
+    return local_snapshot_path
diff --git a/vllm/lora/worker_manager.py b/vllm/lora/worker_manager.py
index 3d0ef4252b0..0d75867d65d 100644
--- a/vllm/lora/worker_manager.py
+++ b/vllm/lora/worker_manager.py
@@ -13,6 +13,7 @@
 from vllm.lora.models import (LoRAModel, LoRAModelManager,
                               LRUCacheLoRAModelManager, create_lora_manager)
 from vllm.lora.request import LoRARequest
+from vllm.lora.utils import get_lora_absolute_path
 
 logger = init_logger(__name__)
 
@@ -89,8 +90,9 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
                         packed_modules_mapping[module])
                 else:
                     expected_lora_modules.append(module)
+            lora_local_path = get_lora_absolute_path(lora_request.lora_local_path)
             lora = self._lora_model_cls.from_local_checkpoint(
-                lora_request.lora_local_path,
+                lora_local_path,
                 expected_lora_modules,
                 max_position_embeddings=self.max_position_embeddings,
                 lora_model_id=lora_request.lora_int_id,
@@ -103,7 +105,7 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
             )
         except Exception as e:
             raise RuntimeError(
-                f"Loading lora {lora_request.lora_local_path} failed") from e
+                f"Loading lora {lora_local_path} failed") from e
         if lora.rank > self.lora_config.max_lora_rank:
             raise ValueError(
                 f"LoRA rank {lora.rank} is greater than max_lora_rank "

From 171044585fc1eb79388fea3499d920f112034873 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Mon, 8 Jul 2024 15:25:34 -0700
Subject: [PATCH 02/10] Add unit tests for get_lora_absolute_path

Format the code and pass the linter
---
 tests/lora/test_utils.py    | 46 ++++++++++++++++++++++++++++++++++++-
 vllm/lora/utils.py          | 20 +++++++++-------
 vllm/lora/worker_manager.py |  6 ++---
 3 files changed, 60 insertions(+), 12 deletions(-)

diff --git a/tests/lora/test_utils.py b/tests/lora/test_utils.py
index 4ff9715b4ca..7b5cfdb398d 100644
--- a/tests/lora/test_utils.py
+++ b/tests/lora/test_utils.py
@@ -1,9 +1,12 @@
 from collections import OrderedDict
+from unittest.mock import patch
 
 import pytest
+from huggingface_hub.utils import RepositoryNotFoundError
 from torch import nn
 
-from vllm.lora.utils import parse_fine_tuned_lora_name, replace_submodule
+from vllm.lora.utils import (get_lora_absolute_path,
+                             parse_fine_tuned_lora_name, replace_submodule)
 from vllm.utils import LRUCache
 
 
@@ -182,3 +185,44 @@ def test_lru_cache():
     assert 2 in cache
     assert 4 in cache
     assert 6 in cache
+
+
+# Unit tests for get_lora_absolute_path
+@patch('os.path.isabs', True)
+def test_get_lora_absolute_path_absolute(mock_isabs):
+    path = '/absolute/path/to/lora'
+    assert get_lora_absolute_path(path) == path
+
+
+@patch('os.path.expanduser', '/home/user/relative/path/to/lora')
+def test_get_lora_absolute_path_expanduser(mock_expanduser):
+    # Path with ~ that needs to be expanded
+    path = '~/relative/path/to/lora'
+    assert get_lora_absolute_path(path) == '/home/user/relative/path/to/lora'
+
+
+@patch('os.path.exists', True)
+@patch('os.path.abspath', '/absolute/path/to/lora')
+def test_get_lora_absolute_path_local_existing(mock_exists, mock_abspath):
+    # Relative path that exists locally
+    path = 'relative/path/to/lora'
+    assert get_lora_absolute_path(path) == '/absolute/path/to/lora'
+
+
+@patch('huggingface_hub.snapshot_download', '/mock/snapshot/path')
+@patch('os.path.exists', False)
+def test_get_lora_absolute_path_huggingface(mock_exists,
+                                            mock_snapshot_download):
+    # Hugging Face model identifier
+    path = 'org/repo'
+    assert get_lora_absolute_path(path) == '/mock/snapshot/path'
+
+
+@patch('huggingface_hub.snapshot_download',
+       side_effect=RepositoryNotFoundError)
+@patch('os.path.exists', False)
+def test_get_lora_absolute_path_huggingface_error(mock_exists,
+                                                  mock_snapshot_download):
+    # Hugging Face model identifier with download error
+    path = 'org/repo'
+    assert get_lora_absolute_path(path) == path
diff --git a/vllm/lora/utils.py b/vllm/lora/utils.py
index 8e8690087d1..6e2c4a1245c 100644
--- a/vllm/lora/utils.py
+++ b/vllm/lora/utils.py
@@ -2,7 +2,8 @@
 from typing import List, Optional, Set, Tuple, Type
 
 from huggingface_hub import snapshot_download
-from huggingface_hub.utils import RepositoryNotFoundError, EntryNotFoundError, HFValidationError
+from huggingface_hub.utils import (EntryNotFoundError, HFValidationError,
+                                   RepositoryNotFoundError)
 from torch import nn
 from transformers import PretrainedConfig
 
@@ -109,13 +110,15 @@ def parse_fine_tuned_lora_name(name: str) -> Tuple[str, bool]:
 
     raise ValueError(f"{name} is unsupported LoRA weight")
 
+
 def get_lora_absolute_path(lora_path: str) -> str:
     """
     Resolves the given lora_path to an absolute local path.
 
-    If the lora_path is identified as a Hugging Face model identifier, it will download
-    the model and return the local snapshot path. Otherwise, it treats the lora_path
-    as a local file path and converts it to an absolute path.
+    If the lora_path is identified as a Hugging Face model identifier,
+    it will download the model and return the local snapshot path.
+    Otherwise, it treats the lora_path as a local file path and
+    converts it to an absolute path.
 
     Parameters:
     lora_path (str): The path to the lora model, which can be an absolute path,
@@ -125,11 +128,11 @@ def get_lora_absolute_path(lora_path: str) -> str:
     str: The resolved absolute local path to the lora model.
     """
 
-     # Check if the path is an absolute path. Return it no matter existing or not.
+    # Check if the path is an absolute path. Return it no matter exists or not.
     if os.path.isabs(lora_path):
         return lora_path
 
-    # If the path starts with ~, expand the user home directory and return the path.
+    # If the path starts with ~, expand the user home directory.
     if lora_path.startswith('~'):
         return os.path.expanduser(lora_path)
 
@@ -137,10 +140,11 @@ def get_lora_absolute_path(lora_path: str) -> str:
     if os.path.exists(lora_path):
         return os.path.abspath(lora_path)
 
-    # If the path does not exist locally, assume it's a Hugging Face model identifier.
+    # If the path does not exist locally, assume it's a Hugging Face repo.
     try:
         local_snapshot_path = snapshot_download(repo_id=lora_path)
-    except (RepositoryNotFoundError, EntryNotFoundError, HFValidationError) as e:
+    except (RepositoryNotFoundError, EntryNotFoundError,
+            HFValidationError) as e:
         # Handle errors that may occur during the download
         # Return original path instead instead of throwing error here
         print(f"Error downloading the Hugging Face model: {e}")
diff --git a/vllm/lora/worker_manager.py b/vllm/lora/worker_manager.py
index 0d75867d65d..8ea92ccc151 100644
--- a/vllm/lora/worker_manager.py
+++ b/vllm/lora/worker_manager.py
@@ -90,7 +90,8 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
                         packed_modules_mapping[module])
                 else:
                     expected_lora_modules.append(module)
-            lora_local_path = get_lora_absolute_path(lora_request.lora_local_path)
+            lora_local_path = get_lora_absolute_path(
+                lora_request.lora_local_path)
             lora = self._lora_model_cls.from_local_checkpoint(
                 lora_local_path,
                 expected_lora_modules,
@@ -104,8 +105,7 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
                 embedding_padding_modules=self.embedding_padding_modules,
             )
         except Exception as e:
-            raise RuntimeError(
-                f"Loading lora {lora_local_path} failed") from e
+            raise RuntimeError(f"Loading lora {lora_local_path} failed") from e
         if lora.rank > self.lora_config.max_lora_rank:
             raise ValueError(
                 f"LoRA rank {lora.rank} is greater than max_lora_rank "

From 0030b9ffce2b2e8b8e9f04a7b108f5877193cec0 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Sat, 6 Jul 2024 15:45:12 -0700
Subject: [PATCH 03/10] Rename lora_local_path to lora_path

---
 tests/core/test_scheduler.py              | 4 ++--
 vllm/entrypoints/openai/serving_engine.py | 4 ++--
 vllm/lora/request.py                      | 2 +-
 vllm/lora/worker_manager.py               | 7 +++----
 vllm/transformers_utils/tokenizer.py      | 5 ++---
 vllm/worker/model_runner.py               | 2 +-
 6 files changed, 11 insertions(+), 13 deletions(-)

diff --git a/tests/core/test_scheduler.py b/tests/core/test_scheduler.py
index bae958211cb..4ca2260b5e0 100644
--- a/tests/core/test_scheduler.py
+++ b/tests/core/test_scheduler.py
@@ -462,7 +462,7 @@ def test_prefill_schedule_max_lora():
                                            lora_request=LoRARequest(
                                                lora_name=str(i),
                                                lora_int_id=i + 1,
-                                               lora_local_path="abc"))
+                                               lora_path="abc"))
         waiting.append(seq_group)
     # Add two more requests to verify lora is prioritized.
     # 0: Lora, 1: Lora, 2: regular, 3: regular
@@ -760,7 +760,7 @@ def test_schedule_swapped_max_loras():
                                            lora_request=LoRARequest(
                                                lora_name=str(i),
                                                lora_int_id=i + 1,
-                                               lora_local_path="abc"))
+                                               lora_path="abc"))
         scheduler._allocate_and_set_running(seq_group)
         append_new_token_seq_group(60, seq_group, 1)
         scheduler._swap_out(seq_group, blocks_to_swap_out)
diff --git a/vllm/entrypoints/openai/serving_engine.py b/vllm/entrypoints/openai/serving_engine.py
index 14c1df89e06..b54480e12e0 100644
--- a/vllm/entrypoints/openai/serving_engine.py
+++ b/vllm/entrypoints/openai/serving_engine.py
@@ -33,7 +33,7 @@ class PromptAdapterPath:
 @dataclass
 class LoRAModulePath:
     name: str
-    local_path: str
+    path: str
 
 
 class OpenAIServing:
@@ -68,7 +68,7 @@ def __init__(
                 LoRARequest(
                     lora_name=lora.name,
                     lora_int_id=i,
-                    lora_local_path=lora.local_path,
+                    lora_path=lora.path,
                 ) for i, lora in enumerate(lora_modules, start=1)
             ]
 
diff --git a/vllm/lora/request.py b/vllm/lora/request.py
index 2d10d037760..858ae40736c 100644
--- a/vllm/lora/request.py
+++ b/vllm/lora/request.py
@@ -20,7 +20,7 @@ class LoRARequest(AdapterRequest):
 
     lora_name: str
     lora_int_id: int
-    lora_local_path: str
+    lora_path: str
     long_lora_max_len: Optional[int] = None
     __hash__ = AdapterRequest.__hash__
 
diff --git a/vllm/lora/worker_manager.py b/vllm/lora/worker_manager.py
index 8ea92ccc151..fdb69a9ef0e 100644
--- a/vllm/lora/worker_manager.py
+++ b/vllm/lora/worker_manager.py
@@ -90,10 +90,9 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
                         packed_modules_mapping[module])
                 else:
                     expected_lora_modules.append(module)
-            lora_local_path = get_lora_absolute_path(
-                lora_request.lora_local_path)
+            lora_path = get_lora_absolute_path(lora_request.lora_path)
             lora = self._lora_model_cls.from_local_checkpoint(
-                lora_local_path,
+                lora_path,
                 expected_lora_modules,
                 max_position_embeddings=self.max_position_embeddings,
                 lora_model_id=lora_request.lora_int_id,
@@ -105,7 +104,7 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
                 embedding_padding_modules=self.embedding_padding_modules,
             )
         except Exception as e:
-            raise RuntimeError(f"Loading lora {lora_local_path} failed") from e
+            raise RuntimeError(f"Loading lora {lora_path} failed") from e
         if lora.rank > self.lora_config.max_lora_rank:
             raise ValueError(
                 f"LoRA rank {lora.rank} is greater than max_lora_rank "
diff --git a/vllm/transformers_utils/tokenizer.py b/vllm/transformers_utils/tokenizer.py
index f5684dbf127..ea74a19baa2 100644
--- a/vllm/transformers_utils/tokenizer.py
+++ b/vllm/transformers_utils/tokenizer.py
@@ -134,14 +134,13 @@ def get_lora_tokenizer(lora_request: LoRARequest, *args,
     if lora_request is None:
         return None
     try:
-        tokenizer = get_tokenizer(lora_request.lora_local_path, *args,
-                                  **kwargs)
+        tokenizer = get_tokenizer(lora_request.lora_path, *args, **kwargs)
     except OSError as e:
         # No tokenizer was found in the LoRA folder,
         # use base model tokenizer
         logger.warning(
             "No tokenizer found in %s, using base model tokenizer instead. "
-            "(Exception: %s)", lora_request.lora_local_path, e)
+            "(Exception: %s)", lora_request.lora_path, e)
         tokenizer = None
     return tokenizer
 
diff --git a/vllm/worker/model_runner.py b/vllm/worker/model_runner.py
index 75a2607d0d9..3168cdf38fc 100644
--- a/vllm/worker/model_runner.py
+++ b/vllm/worker/model_runner.py
@@ -687,7 +687,7 @@ def profile_run(self) -> None:
                     dummy_lora_request = LoRARequest(
                         lora_name=f"warmup_{lora_id}",
                         lora_int_id=lora_id,
-                        lora_local_path="/not/a/real/path",
+                        lora_path="/not/a/real/path",
                     )
                     self.lora_manager.add_dummy_lora(dummy_lora_request,
                                                      rank=LORA_WARMUP_RANK)

From e3c3523f5dd3355a6795e6f39a871bae8b499905 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Mon, 8 Jul 2024 19:09:35 -0700
Subject: [PATCH 04/10] Resolve test failures

---
 tests/lora/test_utils.py | 43 +++++++++++++++++++++++++---------------
 vllm/lora/utils.py       | 11 +++++-----
 2 files changed, 33 insertions(+), 21 deletions(-)

diff --git a/tests/lora/test_utils.py b/tests/lora/test_utils.py
index 7b5cfdb398d..5a9b60d23d5 100644
--- a/tests/lora/test_utils.py
+++ b/tests/lora/test_utils.py
@@ -2,7 +2,7 @@
 from unittest.mock import patch
 
 import pytest
-from huggingface_hub.utils import RepositoryNotFoundError
+from huggingface_hub.utils import HfHubHTTPError
 from torch import nn
 
 from vllm.lora.utils import (get_lora_absolute_path,
@@ -188,41 +188,52 @@ def test_lru_cache():
 
 
 # Unit tests for get_lora_absolute_path
-@patch('os.path.isabs', True)
+@patch('os.path.isabs')
 def test_get_lora_absolute_path_absolute(mock_isabs):
     path = '/absolute/path/to/lora'
+    mock_isabs.return_value = True
     assert get_lora_absolute_path(path) == path
 
 
-@patch('os.path.expanduser', '/home/user/relative/path/to/lora')
+@patch('os.path.expanduser')
 def test_get_lora_absolute_path_expanduser(mock_expanduser):
     # Path with ~ that needs to be expanded
     path = '~/relative/path/to/lora'
-    assert get_lora_absolute_path(path) == '/home/user/relative/path/to/lora'
+    absolute_path = '/home/user/relative/path/to/lora'
+    mock_expanduser.return_value = absolute_path
+    assert get_lora_absolute_path(path) == absolute_path
 
 
-@patch('os.path.exists', True)
-@patch('os.path.abspath', '/absolute/path/to/lora')
-def test_get_lora_absolute_path_local_existing(mock_exists, mock_abspath):
+@patch('os.path.exists')
+@patch('os.path.abspath')
+def test_get_lora_absolute_path_local_existing(mock_abspath, mock_exist):
     # Relative path that exists locally
     path = 'relative/path/to/lora'
-    assert get_lora_absolute_path(path) == '/absolute/path/to/lora'
+    absolute_path = '/absolute/path/to/lora'
+    mock_exist.return_value = True
+    mock_abspath.return_value = absolute_path
+    assert get_lora_absolute_path(path) == absolute_path
 
 
-@patch('huggingface_hub.snapshot_download', '/mock/snapshot/path')
-@patch('os.path.exists', False)
-def test_get_lora_absolute_path_huggingface(mock_exists,
+@patch('huggingface_hub.snapshot_download')
+@patch('os.path.exists')
+def test_get_lora_absolute_path_huggingface(mock_exist,
                                             mock_snapshot_download):
     # Hugging Face model identifier
     path = 'org/repo'
-    assert get_lora_absolute_path(path) == '/mock/snapshot/path'
+    absolute_path = '/mock/snapshot/path'
+    mock_exist.return_value = False
+    mock_snapshot_download.return_value = absolute_path
+    assert get_lora_absolute_path(path) == absolute_path
 
 
-@patch('huggingface_hub.snapshot_download',
-       side_effect=RepositoryNotFoundError)
-@patch('os.path.exists', False)
-def test_get_lora_absolute_path_huggingface_error(mock_exists,
+@patch('huggingface_hub.snapshot_download')
+@patch('os.path.exists')
+def test_get_lora_absolute_path_huggingface_error(mock_exist,
                                                   mock_snapshot_download):
     # Hugging Face model identifier with download error
     path = 'org/repo'
+    mock_exist.return_value = False
+    mock_snapshot_download.side_effect = HfHubHTTPError(
+        "failed to query model info")
     assert get_lora_absolute_path(path) == path
diff --git a/vllm/lora/utils.py b/vllm/lora/utils.py
index 6e2c4a1245c..923fbce1f3f 100644
--- a/vllm/lora/utils.py
+++ b/vllm/lora/utils.py
@@ -1,9 +1,9 @@
 import os
 from typing import List, Optional, Set, Tuple, Type
 
-from huggingface_hub import snapshot_download
-from huggingface_hub.utils import (EntryNotFoundError, HFValidationError,
-                                   RepositoryNotFoundError)
+import huggingface_hub
+from huggingface_hub.utils import (EntryNotFoundError, HfHubHTTPError,
+                                   HFValidationError, RepositoryNotFoundError)
 from torch import nn
 from transformers import PretrainedConfig
 
@@ -142,8 +142,9 @@ def get_lora_absolute_path(lora_path: str) -> str:
 
     # If the path does not exist locally, assume it's a Hugging Face repo.
     try:
-        local_snapshot_path = snapshot_download(repo_id=lora_path)
-    except (RepositoryNotFoundError, EntryNotFoundError,
+        local_snapshot_path = huggingface_hub.snapshot_download(
+            repo_id=lora_path)
+    except (HfHubHTTPError, RepositoryNotFoundError, EntryNotFoundError,
             HFValidationError) as e:
         # Handle errors that may occur during the download
         # Return original path instead instead of throwing error here

From 14557220883c2b78c6787a75fa610d23de3b9506 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Tue, 9 Jul 2024 00:03:31 -0700
Subject: [PATCH 05/10] Update vllm/lora/utils.py

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
---
 vllm/lora/utils.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/vllm/lora/utils.py b/vllm/lora/utils.py
index 923fbce1f3f..2735014090b 100644
--- a/vllm/lora/utils.py
+++ b/vllm/lora/utils.py
@@ -111,7 +111,7 @@ def parse_fine_tuned_lora_name(name: str) -> Tuple[str, bool]:
     raise ValueError(f"{name} is unsupported LoRA weight")
 
 
-def get_lora_absolute_path(lora_path: str) -> str:
+def get_adapter_absolute_path(lora_path: str) -> str:
     """
     Resolves the given lora_path to an absolute local path.
 

From a6a30b99c1d6a5a709ce51ee653f48e6999d3000 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Tue, 9 Jul 2024 00:09:31 -0700
Subject: [PATCH 06/10] Address some feedback from reviewer

---
 tests/lora/test_utils.py    | 28 ++++++++++++++--------------
 vllm/lora/utils.py          |  4 ++--
 vllm/lora/worker_manager.py |  4 ++--
 3 files changed, 18 insertions(+), 18 deletions(-)

diff --git a/tests/lora/test_utils.py b/tests/lora/test_utils.py
index 5a9b60d23d5..db02bacdb64 100644
--- a/tests/lora/test_utils.py
+++ b/tests/lora/test_utils.py
@@ -5,7 +5,7 @@
 from huggingface_hub.utils import HfHubHTTPError
 from torch import nn
 
-from vllm.lora.utils import (get_lora_absolute_path,
+from vllm.lora.utils import (get_adapter_absolute_path,
                              parse_fine_tuned_lora_name, replace_submodule)
 from vllm.utils import LRUCache
 
@@ -187,53 +187,53 @@ def test_lru_cache():
     assert 6 in cache
 
 
-# Unit tests for get_lora_absolute_path
+# Unit tests for get_adapter_absolute_path
 @patch('os.path.isabs')
-def test_get_lora_absolute_path_absolute(mock_isabs):
+def test_get_adapter_absolute_path_absolute(mock_isabs):
     path = '/absolute/path/to/lora'
     mock_isabs.return_value = True
-    assert get_lora_absolute_path(path) == path
+    assert get_adapter_absolute_path(path) == path
 
 
 @patch('os.path.expanduser')
-def test_get_lora_absolute_path_expanduser(mock_expanduser):
+def test_get_adapter_absolute_path_expanduser(mock_expanduser):
     # Path with ~ that needs to be expanded
     path = '~/relative/path/to/lora'
     absolute_path = '/home/user/relative/path/to/lora'
     mock_expanduser.return_value = absolute_path
-    assert get_lora_absolute_path(path) == absolute_path
+    assert get_adapter_absolute_path(path) == absolute_path
 
 
 @patch('os.path.exists')
 @patch('os.path.abspath')
-def test_get_lora_absolute_path_local_existing(mock_abspath, mock_exist):
+def test_get_adapter_absolute_path_local_existing(mock_abspath, mock_exist):
     # Relative path that exists locally
     path = 'relative/path/to/lora'
     absolute_path = '/absolute/path/to/lora'
     mock_exist.return_value = True
     mock_abspath.return_value = absolute_path
-    assert get_lora_absolute_path(path) == absolute_path
+    assert get_adapter_absolute_path(path) == absolute_path
 
 
 @patch('huggingface_hub.snapshot_download')
 @patch('os.path.exists')
-def test_get_lora_absolute_path_huggingface(mock_exist,
-                                            mock_snapshot_download):
+def test_get_adapter_absolute_path_huggingface(mock_exist,
+                                               mock_snapshot_download):
     # Hugging Face model identifier
     path = 'org/repo'
     absolute_path = '/mock/snapshot/path'
     mock_exist.return_value = False
     mock_snapshot_download.return_value = absolute_path
-    assert get_lora_absolute_path(path) == absolute_path
+    assert get_adapter_absolute_path(path) == absolute_path
 
 
 @patch('huggingface_hub.snapshot_download')
 @patch('os.path.exists')
-def test_get_lora_absolute_path_huggingface_error(mock_exist,
-                                                  mock_snapshot_download):
+def test_get_adapter_absolute_path_huggingface_error(mock_exist,
+                                                     mock_snapshot_download):
     # Hugging Face model identifier with download error
     path = 'org/repo'
     mock_exist.return_value = False
     mock_snapshot_download.side_effect = HfHubHTTPError(
         "failed to query model info")
-    assert get_lora_absolute_path(path) == path
+    assert get_adapter_absolute_path(path) == path
diff --git a/vllm/lora/utils.py b/vllm/lora/utils.py
index 2735014090b..4513337299e 100644
--- a/vllm/lora/utils.py
+++ b/vllm/lora/utils.py
@@ -145,10 +145,10 @@ def get_adapter_absolute_path(lora_path: str) -> str:
         local_snapshot_path = huggingface_hub.snapshot_download(
             repo_id=lora_path)
     except (HfHubHTTPError, RepositoryNotFoundError, EntryNotFoundError,
-            HFValidationError) as e:
+            HFValidationError):
         # Handle errors that may occur during the download
         # Return original path instead instead of throwing error here
-        print(f"Error downloading the Hugging Face model: {e}")
+        logger.exception("Error downloading the HuggingFace model")
         return lora_path
 
     return local_snapshot_path
diff --git a/vllm/lora/worker_manager.py b/vllm/lora/worker_manager.py
index fdb69a9ef0e..724c308a07a 100644
--- a/vllm/lora/worker_manager.py
+++ b/vllm/lora/worker_manager.py
@@ -13,7 +13,7 @@
 from vllm.lora.models import (LoRAModel, LoRAModelManager,
                               LRUCacheLoRAModelManager, create_lora_manager)
 from vllm.lora.request import LoRARequest
-from vllm.lora.utils import get_lora_absolute_path
+from vllm.lora.utils import get_adapter_absolute_path
 
 logger = init_logger(__name__)
 
@@ -90,7 +90,7 @@ def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
                         packed_modules_mapping[module])
                 else:
                     expected_lora_modules.append(module)
-            lora_path = get_lora_absolute_path(lora_request.lora_path)
+            lora_path = get_adapter_absolute_path(lora_request.lora_path)
             lora = self._lora_model_cls.from_local_checkpoint(
                 lora_path,
                 expected_lora_modules,

From a16ec5dd6143cf3714b1bb574c24502ffa3df12d Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Fri, 12 Jul 2024 15:34:21 -0700
Subject: [PATCH 07/10] Add test for loading lora from huggingface

---
 tests/lora/conftest.py              |  6 +++++
 tests/lora/test_lora_huggingface.py | 39 +++++++++++++++++++++++++++++
 2 files changed, 45 insertions(+)
 create mode 100644 tests/lora/test_lora_huggingface.py

diff --git a/tests/lora/conftest.py b/tests/lora/conftest.py
index bda123bf131..a962efb6900 100644
--- a/tests/lora/conftest.py
+++ b/tests/lora/conftest.py
@@ -163,6 +163,12 @@ def sql_lora_files():
     return snapshot_download(repo_id="yard1/llama-2-7b-sql-lora-test")
 
 
+@pytest.fixture(scope="session")
+def sql_lora_huggingface_id():
+    # huggingface repo id is used to test lora runtime downloading.
+    return "yard1/llama-2-7b-sql-lora-test"
+
+
 @pytest.fixture(scope="session")
 def mixtral_lora_files():
     # Note: this module has incorrect adapter_config.json to test
diff --git a/tests/lora/test_lora_huggingface.py b/tests/lora/test_lora_huggingface.py
new file mode 100644
index 00000000000..e2daf9d1351
--- /dev/null
+++ b/tests/lora/test_lora_huggingface.py
@@ -0,0 +1,39 @@
+from typing import List
+
+import pytest
+
+from vllm.lora.models import LoRAModel
+from vllm.lora.utils import get_adapter_absolute_path
+from vllm.model_executor.models.llama import LlamaForCausalLM
+
+# Provide absolute path and huggingface lora ids
+lora_fixture_name = ["sql_lora_files", "sql_lora_huggingface_id"]
+
+
+@pytest.mark.parametrize("lora_fixture_name", lora_fixture_name)
+def test_load_checkpoints_from_huggingface(lora_fixture_name, request):
+    lora_name = request.getfixturevalue(lora_fixture_name)
+    supported_lora_modules = LlamaForCausalLM.supported_lora_modules
+    packed_modules_mapping = LlamaForCausalLM.packed_modules_mapping
+    embedding_modules = LlamaForCausalLM.embedding_modules
+    embed_padding_modules = LlamaForCausalLM.embedding_padding_modules
+    expected_lora_modules: List[str] = []
+    for module in supported_lora_modules:
+        if module in packed_modules_mapping:
+            expected_lora_modules.extend(packed_modules_mapping[module])
+        else:
+            expected_lora_modules.append(module)
+
+    lora_path = get_adapter_absolute_path(lora_name)
+
+    # lora loading should work for either absolute path and hugggingface id.
+    lora_model = LoRAModel.from_local_checkpoint(
+        lora_path,
+        expected_lora_modules,
+        lora_model_id=1,
+        device="cpu",
+        embedding_modules=embedding_modules,
+        embedding_padding_modules=embed_padding_modules)
+
+    # Assertions to ensure the model is loaded correctly
+    assert lora_model is not None, "LoRAModel is not loaded correctly"

From 10ee978e878dbcd3915f249b69ec404ff6defb0b Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Fri, 12 Jul 2024 17:56:13 -0700
Subject: [PATCH 08/10] Update tests/lora/conftest.py

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
---
 tests/lora/conftest.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/tests/lora/conftest.py b/tests/lora/conftest.py
index a962efb6900..c4efee21f09 100644
--- a/tests/lora/conftest.py
+++ b/tests/lora/conftest.py
@@ -168,6 +168,9 @@ def sql_lora_huggingface_id():
     # huggingface repo id is used to test lora runtime downloading.
     return "yard1/llama-2-7b-sql-lora-test"
 
+@pytest.fixture(scope="session")
+def sql_lora_files(sql_lora_huggingface_id):
+    return snapshot_download(repo_id=sql_lora_huggingface_id)
 
 @pytest.fixture(scope="session")
 def mixtral_lora_files():

From 0ea0d0d26ba5d4f6434be865e79ab77ec9cd8aad Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Sun, 14 Jul 2024 23:59:17 -0700
Subject: [PATCH 09/10] Add deprecation notice for lora_local_path

---
 tests/lora/conftest.py |  7 ++-----
 vllm/lora/request.py   | 42 +++++++++++++++++++++++++++++++++++++++---
 2 files changed, 41 insertions(+), 8 deletions(-)

diff --git a/tests/lora/conftest.py b/tests/lora/conftest.py
index c4efee21f09..0bcae5b0c96 100644
--- a/tests/lora/conftest.py
+++ b/tests/lora/conftest.py
@@ -158,20 +158,17 @@ def dummy_model_gate_up() -> nn.Module:
     return model
 
 
-@pytest.fixture(scope="session")
-def sql_lora_files():
-    return snapshot_download(repo_id="yard1/llama-2-7b-sql-lora-test")
-
-
 @pytest.fixture(scope="session")
 def sql_lora_huggingface_id():
     # huggingface repo id is used to test lora runtime downloading.
     return "yard1/llama-2-7b-sql-lora-test"
 
+
 @pytest.fixture(scope="session")
 def sql_lora_files(sql_lora_huggingface_id):
     return snapshot_download(repo_id=sql_lora_huggingface_id)
 
+
 @pytest.fixture(scope="session")
 def mixtral_lora_files():
     # Note: this module has incorrect adapter_config.json to test
diff --git a/vllm/lora/request.py b/vllm/lora/request.py
index 858ae40736c..a0b9fd6cd41 100644
--- a/vllm/lora/request.py
+++ b/vllm/lora/request.py
@@ -1,4 +1,5 @@
-from dataclasses import dataclass
+import warnings
+from dataclasses import dataclass, field
 from typing import Optional
 
 from vllm.adapter_commons.request import AdapterRequest
@@ -20,10 +21,25 @@ class LoRARequest(AdapterRequest):
 
     lora_name: str
     lora_int_id: int
-    lora_path: str
+    lora_path: str = ""
+    lora_local_path: Optional[str] = field(default=None, repr=False)
     long_lora_max_len: Optional[int] = None
     __hash__ = AdapterRequest.__hash__
 
+    def __post_init__(self):
+        if 'lora_local_path' in self.__dict__:
+            warnings.warn(
+                "The 'lora_local_path' attribute is deprecated "
+                "and will be removed in a future version. "
+                "Please use 'lora_path' instead.",
+                DeprecationWarning,
+                stacklevel=2)
+            if not self.lora_path:
+                self.lora_path = self.lora_local_path or ""
+
+        # Ensure lora_path is not empty
+        assert self.lora_path, "lora_path can not be empty"
+
     @property
     def adapter_id(self):
         return self.lora_int_id
@@ -32,6 +48,26 @@ def adapter_id(self):
     def name(self):
         return self.lora_name
 
+    @property
+    def path(self):
+        return self.lora_path
+
     @property
     def local_path(self):
-        return self.lora_local_path
+        warnings.warn(
+            "The 'local_path' attribute is deprecated "
+            "and will be removed in a future version. "
+            "Please use 'path' instead.",
+            DeprecationWarning,
+            stacklevel=2)
+        return self.lora_path
+
+    @local_path.setter
+    def local_path(self, value):
+        warnings.warn(
+            "The 'local_path' attribute is deprecated "
+            "and will be removed in a future version. "
+            "Please use 'path' instead.",
+            DeprecationWarning,
+            stacklevel=2)
+        self.lora_path = value

From ba2a04de6cb894d96c5141bbca438b3b7b44a080 Mon Sep 17 00:00:00 2001
From: Jiaxin Shan <seedjeffwan@gmail.com>
Date: Wed, 17 Jul 2024 18:02:43 -0700
Subject: [PATCH 10/10] Fix failures in test_long_context.

The problem comes from new added field and the test didn't explicitly set the field but using position information.
---
 tests/lora/test_long_context.py | 2 +-
 vllm/lora/request.py            | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/tests/lora/test_long_context.py b/tests/lora/test_long_context.py
index 853fd9fb3ce..389a3ccbc17 100644
--- a/tests/lora/test_long_context.py
+++ b/tests/lora/test_long_context.py
@@ -29,7 +29,7 @@ def _create_lora_request(lora_id, long_context_infos):
     context_len = long_context_infos[lora_id]["context_length"]
     scaling_factor = context_len_to_scaling_factor[context_len]
     return LoRARequest(context_len, lora_id,
-                       long_context_infos[lora_id]["lora"],
+                       long_context_infos[lora_id]["lora"], None,
                        4096 * scaling_factor)
 
 
diff --git a/vllm/lora/request.py b/vllm/lora/request.py
index a0b9fd6cd41..5d791424fbe 100644
--- a/vllm/lora/request.py
+++ b/vllm/lora/request.py
@@ -38,7 +38,7 @@ def __post_init__(self):
                 self.lora_path = self.lora_local_path or ""
 
         # Ensure lora_path is not empty
-        assert self.lora_path, "lora_path can not be empty"
+        assert self.lora_path, "lora_path cannot be empty"
 
     @property
     def adapter_id(self):