Skip to content

Improve error handling for requests with unloaded LoRA path(s) #7642

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Jul 2, 2025
Merged

Improve error handling for requests with unloaded LoRA path(s) #7642

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
9f22632
Handle requests with unloaded LoRA.
lifuhuang Jun 30, 2025
3899b29
Minor
lifuhuang Jun 30, 2025
8139e8d
Merge remote-tracking branch 'origin/main' into lifuhuang/lora-error
lifuhuang Jun 30, 2025
2db1b9a
Fix typo.
lifuhuang Jun 30, 2025
dae680d
Fix
lifuhuang Jun 30, 2025
e838b47
Merge branch 'main' into lifuhuang/lora-error
lifuhuang Jun 30, 2025
3dc0509
Merge branch 'main' into lifuhuang/lora-error
lifuhuang Jun 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 22 additions & 0 deletions python/sglang/srt/managers/tokenizer_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -241,6 +241,10 @@ def __init__(
revision=server_args.revision,
)

# Initialize loaded loRA adapters with the initial lora paths in the server_args.
# This list will be updated when new LoRA adapters are loaded or unloaded dynamically.
self.loaded_lora_adapters: Dict[str, str] = self.server_args.lora_paths.copy()

# Store states
self.no_create_loop = False
self.rid_to_state: Dict[str, ReqState] = {}
Expand Down Expand Up @@ -444,6 +448,22 @@ async def generate_request(
"with --enable-return-hidden-states (ServerArgs.enable_return_hidden_states)."
)

# Verify requested LoRA adapters are loaded.
if obj.lora_path:
requested_adapters = (
set(obj.lora_path)
if isinstance(obj.lora_path, list)
else {obj.lora_path}
)
unloaded_adapters = (
requested_adapters - self.loaded_lora_adapters.keys()
)
if unloaded_adapters:
raise ValueError(
f"The following requested LoRA adapters are not loaded: {unloaded_adapters}\n"
f"Loaded adapters: {self.loaded_lora_adapters}."
)

if self.log_requests:
max_length, skip_names, _ = self.log_request_metadata
logger.info(
Expand Down Expand Up @@ -992,6 +1012,7 @@ async def load_lora_adapter(

async with self.model_update_lock.writer_lock:
result = (await self.update_lora_adapter_communicator(obj))[0]
self.loaded_lora_adapters = result.loaded_adapters
return result

async def unload_lora_adapter(
Expand All @@ -1013,6 +1034,7 @@ async def unload_lora_adapter(

async with self.model_update_lock.writer_lock:
result = (await self.update_lora_adapter_communicator(obj))[0]
self.loaded_lora_adapters = result.loaded_adapters
return result

async def get_weights_by_name(
Expand Down
122 changes: 105 additions & 17 deletions test/srt/models/lora/test_lora_update.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,14 +42,14 @@
class OperationType(Enum):
LOAD = "load"
UNLOAD = "unload"
NOOP = "noop"
FORWARD = "forward"
EXPECT_ERROR = "expect_error"


@dataclass
class Operation:
type: OperationType
data: Optional[str]
data: Union[str, tuple[str, str]]


@dataclass
Expand All @@ -62,7 +62,7 @@ class TestCase:
max_new_tokens: int = 32


def create_batch_data(adapters: Union[str, list]) -> dict:
def create_batch_data(adapters: Union[str, list]) -> List[tuple[str, str]]:
if not isinstance(adapters, list):
adapters = [adapters]
return [(prompt, adapter) for prompt in PROMPTS for adapter in adapters]
Expand All @@ -80,6 +80,26 @@ def create_batch_data(adapters: Union[str, list]) -> dict:
],
initial_adapters=["philschmid/code-llama-3-1-8b-text-to-sql-lora"],
op_sequence=[
Operation(
type=OperationType.FORWARD,
data=create_batch_data("philschmid/code-llama-3-1-8b-text-to-sql-lora"),
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data(
"Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16"
),
"not loaded",
),
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data("pbevan11/llama-3.1-8b-ocr-correction"),
"not loaded",
),
),
Operation(
type=OperationType.LOAD,
data="Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16",
Expand All @@ -102,6 +122,13 @@ def create_batch_data(adapters: Union[str, list]) -> dict:
type=OperationType.UNLOAD,
data="philschmid/code-llama-3-1-8b-text-to-sql-lora",
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data("philschmid/code-llama-3-1-8b-text-to-sql-lora"),
"not loaded",
),
),
Operation(
type=OperationType.FORWARD,
data=create_batch_data(
Expand All @@ -115,6 +142,15 @@ def create_batch_data(adapters: Union[str, list]) -> dict:
type=OperationType.UNLOAD,
data="Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16",
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data(
"Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16"
),
"not loaded",
),
),
Operation(
type=OperationType.FORWARD,
data=create_batch_data("pbevan11/llama-3.1-8b-ocr-correction"),
Expand Down Expand Up @@ -149,6 +185,22 @@ def create_batch_data(adapters: Union[str, list]) -> dict:
type=OperationType.FORWARD,
data=create_batch_data("philschmid/code-llama-3-1-8b-text-to-sql-lora"),
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data(
"Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16"
),
"not loaded",
),
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data("pbevan11/llama-3.1-8b-ocr-correction"),
"not loaded",
),
),
Operation(
type=OperationType.LOAD,
data="pbevan11/llama-3.1-8b-ocr-correction",
Expand All @@ -157,6 +209,13 @@ def create_batch_data(adapters: Union[str, list]) -> dict:
type=OperationType.UNLOAD,
data="philschmid/code-llama-3-1-8b-text-to-sql-lora",
),
Operation(
type=OperationType.EXPECT_ERROR,
data=(
create_batch_data("philschmid/code-llama-3-1-8b-text-to-sql-lora"),
"not loaded",
),
),
Operation(
type=OperationType.FORWARD,
data=create_batch_data("pbevan11/llama-3.1-8b-ocr-correction"),
Expand Down Expand Up @@ -332,19 +391,31 @@ def forward(
prompts: List[str],
lora_paths: List[str],
max_new_tokens: int = 32,
expected_error: str = None,
):
"""
Perform a batch forward pass with the current set of loaded LoRA adapters.
"""
response = self.handle.batch_forward(
prompts=prompts,
lora_paths=lora_paths,
max_new_tokens=max_new_tokens,
)
output_strs = response.output_strs
try:
response = self.handle.batch_forward(
prompts=prompts,
lora_paths=lora_paths,
max_new_tokens=max_new_tokens,
)
except ValueError as e:
if expected_error:
error_message = str(e)
self.testcase.assertIn(expected_error, error_message)
print(f"Received error as expected: {error_message}")
return error_message

raise e

self.testcase.assertEqual(len(response.output_strs), len(prompts))
output = response.output_strs
print(f"output_strs: {output}")

print(f"output_strs: {output_strs}")
return output_strs
return output


class LoRAUpdateServerTestSession(LoRAUpdateTestSessionBase):
Expand Down Expand Up @@ -426,6 +497,7 @@ def forward(
prompts: List[str],
lora_paths: List[str],
max_new_tokens: int = 32,
expected_error: str = None,
):
"""
Perform a batch forward pass with the current set of loaded LoRA adapters.
Expand All @@ -442,11 +514,18 @@ def forward(
},
},
)
self.testcase.assertTrue(response.ok)
output_strs = [r["text"] for r in response.json()]

print(f"output_strs: {output_strs}")
return output_strs
if expected_error:
self.testcase.assertEqual(response.status_code, 400)
self.testcase.assertIn(expected_error, response.text)
output = response.text
print(f"Received error as expected: {response.text}")
return output
else:
self.testcase.assertTrue(response.ok)
output = [r["text"] for r in response.json()]
self.testcase.assertEqual(len(output), len(prompts))
print(f"output_strs: {output}")
return output


# Factory function to create the appropriate LoRA test session based on mode
Expand Down Expand Up @@ -535,14 +614,23 @@ def _run_operation_sequence(
max_new_tokens=max_new_tokens,
)
forward_outputs.append(result)
elif op_type == OperationType.EXPECT_ERROR:
input_data, expected_error = data
prompts, adapters = zip(*input_data)
result = session.forward(
prompts=list(prompts),
lora_paths=list(adapters),
max_new_tokens=max_new_tokens,
expected_error=expected_error,
)

return forward_outputs

def test_dynamic_adapter_updates(self):
for case_idx, test_case in enumerate(TEST_CASES, start=1):
for mode in [
LoRAUpdateTestSessionMode.SERVER,
LoRAUpdateTestSessionMode.ENGINE,
LoRAUpdateTestSessionMode.SERVER,
]:
print("=" * 100)
print(f"Starting test case {case_idx} in {mode.value} mode.")
Expand Down