Early exit for MoE LoRA kernels (#27131)

Signed-off-by: gnovack <gnovack@amazon.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
gnovack
2025-11-03 20:22:17 +08:00
committed by GitHub
co-authored by Jee Jee Li
parent 40b69e33e7
commit 294c805f1d
11 changed files with 123 additions and 34 deletions
+6
View File
@@ -134,6 +134,8 @@ def use_fused_moe_lora_kernel(
)
expert_ids = torch.empty((max_loras * max_num_m_blocks,), dtype=torch.int32)
num_tokens_post_padded = torch.empty((max_loras,), dtype=torch.int32)
adapter_enabled = torch.ones(max_loras + 1, dtype=torch.int32)
lora_ids = torch.arange(max_loras + 2, dtype=torch.int32)
# call kernel
ops.moe_lora_align_block_size(
@@ -147,6 +149,8 @@ def use_fused_moe_lora_kernel(
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
adapter_enabled,
lora_ids,
)
config = {
@@ -172,6 +176,8 @@ def use_fused_moe_lora_kernel(
num_tokens_post_padded,
max_lora_rank,
top_k_num,
lora_ids,
adapter_enabled,
config["BLOCK_SIZE_M"],
config["BLOCK_SIZE_N"],
config["BLOCK_SIZE_K"],
+4
View File
@@ -60,6 +60,8 @@ def test_moe_lora_align_block_size(
(max_loras * max_num_m_blocks,), num_experts, dtype=torch.int32, device="cuda"
)
num_tokens_post_pad = torch.zeros((max_loras,), dtype=torch.int32, device="cuda")
adapter_enabled = torch.ones((max_loras + 1,), dtype=torch.int32, device="cuda")
lora_ids = torch.arange(max_loras + 2, dtype=torch.int32, device="cuda")
# call kernel
ops.moe_lora_align_block_size(
@@ -73,6 +75,8 @@ def test_moe_lora_align_block_size(
sorted_token_ids,
expert_ids,
num_tokens_post_pad,
adapter_enabled,
lora_ids,
)
# verify values
+43 -7
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import vllm
from vllm.lora.request import LoRARequest
@@ -28,8 +29,17 @@ EXPECTED_LORA_OUTPUT = [
"SELECT poll_source FROM candidate GROUP BY poll_source ORDER BY count(*) DESC LIMIT 1", # noqa: E501
]
EXPECTED_BASE_MODEL_OUTPUT = [
"SELECT COUNT(Candidate_ID) FROM candidate",
"SELECT COUNT(Candidate_ID) FROM candidate",
"SELECT Candidate_ID, COUNT(*) as Total_Candidates\nFROM candidate\nINNER JOIN people ON candidate.People_ID = people.People_ID", # noqa: E501
"SELECT Candidate_ID, Poll_Source FROM candidate WHERE People_ID IN (SELECT People_ID FROM people) ORDER BY COUNT(*) DESC LIMIT 1", # noqa: E501
]
def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
def generate_and_test(
llm: vllm.LLM, lora_path: str, lora_id: list[int | None] | int | None
) -> None:
prompts = [
PROMPT_TEMPLATE.format(context="How many candidates are there?"),
PROMPT_TEMPLATE.format(context="Count the number of candidates."),
@@ -40,12 +50,18 @@ def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
context="Return the poll resource associated with the most candidates."
),
]
lora_request = None
if isinstance(lora_id, int):
lora_request = LoRARequest(str(lora_id), lora_id, lora_path)
elif isinstance(lora_id, list):
lora_request = [
LoRARequest(str(i), i, lora_path) if i is not None else None
for i in lora_id
]
sampling_params = vllm.SamplingParams(temperature=0, max_tokens=64)
outputs = llm.generate(
prompts,
sampling_params,
lora_request=LoRARequest(str(lora_id), lora_id, lora_path) if lora_id else None,
)
outputs = llm.generate(prompts, sampling_params, lora_request=lora_request)
# Print the outputs.
generated_texts: list[str] = []
for output in outputs:
@@ -55,7 +71,13 @@ def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
for i in range(len(EXPECTED_LORA_OUTPUT)):
assert generated_texts[i].startswith(EXPECTED_LORA_OUTPUT[i])
req_lora_id = lora_id[i] if isinstance(lora_id, list) else lora_id
expected_output = (
EXPECTED_LORA_OUTPUT[i]
if req_lora_id is not None
else EXPECTED_BASE_MODEL_OUTPUT[i]
)
assert generated_texts[i].startswith(expected_output)
def test_olmoe_lora(olmoe_lora_files):
@@ -75,6 +97,20 @@ def test_olmoe_lora(olmoe_lora_files):
generate_and_test(llm, olmoe_lora_files, lora_id=2)
def test_olmoe_lora_mixed(olmoe_lora_files):
llm = vllm.LLM(
MODEL_PATH,
max_model_len=1024,
enable_lora=True,
max_loras=4,
enforce_eager=True,
trust_remote_code=True,
enable_chunked_prefill=True,
)
generate_and_test(llm, olmoe_lora_files, lora_id=[1, None, 3, None])
@multi_gpu_test(num_gpus=2)
def test_olmoe_lora_tp2(olmoe_lora_files):
llm = vllm.LLM(