[CI/Build] Add gpt-oss LoRA test (#27870)

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

.buildkite/test-amd.yaml

@@ -441,7 +441,7 @@ steps:
       --ignore=lora/test_llm_with_multi_loras.py \
       --ignore=lora/test_olmoe_tp.py \
       --ignore=lora/test_deepseekv2_tp.py \
-      --ignore=lora/test_gptoss.py \
+      --ignore=lora/test_gptoss_tp.py \
       --ignore=lora/test_qwen3moe_tp.py
   parallelism: 4
 
@@ -1217,6 +1217,8 @@ steps:
     - pytest -v -s -x lora/test_llama_tp.py
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
+    - pytest -v -s -x lora/test_gptoss_tp.py
+
 
 - label: Weight Loading Multiple GPU Test  # 33min
   timeout_in_minutes: 45

.buildkite/test-pipeline.yaml

@@ -417,7 +417,7 @@ steps:
       --ignore=lora/test_llm_with_multi_loras.py \
       --ignore=lora/test_olmoe_tp.py \
       --ignore=lora/test_deepseekv2_tp.py \
-      --ignore=lora/test_gptoss.py \
+      --ignore=lora/test_gptoss_tp.py \
       --ignore=lora/test_qwen3moe_tp.py
 
   parallelism: 4
@@ -1119,6 +1119,7 @@ steps:
     - pytest -v -s -x lora/test_llama_tp.py
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
+    - pytest -v -s -x lora/test_gptoss_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

tests/lora/conftest.py

@@ -237,7 +237,7 @@ def deepseekv2_lora_files():
 
 @pytest.fixture(scope="session")
 def gptoss20b_lora_files():
-    return snapshot_download(repo_id="LevinZheng/gpt-oss-20b-lora-adapter")
+    return snapshot_download(repo_id="jeeejeee/gpt-oss-20b-lora-adapter-text2sql")
 
 
 @pytest.fixture(scope="session")

tests/lora/test_deepseekv2_tp.py

@@ -1,6 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+# NOTE To avoid overloading the CI pipeline, this test script will
+# not be triggered on CI and is primarily intended for local testing
+# and verification.
+
 import vllm
 from vllm.lora.request import LoRARequest
 

tests/lora/test_gptoss.py

@@ -1,52 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import vllm
-from vllm.lora.request import LoRARequest
-
-MODEL_PATH = "openai/gpt-oss-20b"
-
-PROMPT_TEMPLATE = "<｜begin▁of▁sentence｜>You are a helpful assistant.\n\nUser: {context}\n\nAssistant:"  # noqa: E501
-
-
-def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> list[str]:
-    prompts = [
-        PROMPT_TEMPLATE.format(context="Who are you?"),
-    ]
-    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,
-    )
-    # Print the outputs.
-    generated_texts: list[str] = []
-    for output in outputs:
-        prompt = output.prompt
-        generated_text = output.outputs[0].text.strip()
-        generated_texts.append(generated_text)
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-    return generated_texts
-
-
-# FIXME: Load gpt-oss adapter
-def test_gptoss20b_lora(gptoss20b_lora_files):
-    # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI,
-    # Otherwise, the lora-test will fail due to CUDA OOM.
-    llm = vllm.LLM(
-        MODEL_PATH,
-        enable_lora=True,
-        max_loras=4,
-        trust_remote_code=True,
-    )
-
-    expected_lora_output = [
-        "I am an AI language model developed by OpenAI. "
-        "I am here to help you with any questions or "
-        "tasks you may have."
-    ]
-
-    output1 = do_sample(llm, gptoss20b_lora_files, lora_id=1)
-    print(output1)
-    for i in range(len(expected_lora_output)):
-        assert output1[i].startswith(expected_lora_output[i])

tests/lora/test_gptoss_tp.py

@@ -0,0 +1,106 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import vllm
+from vllm.lora.request import LoRARequest
+
+from ..utils import multi_gpu_test
+
+MODEL_PATH = "openai/gpt-oss-20b"
+
+PROMPT_TEMPLATE = """<|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI.
+Knowledge cutoff: 2024-06
+Current date: 2025-10-29
+
+Reasoning: medium
+
+# Valid channels: analysis, commentary, final. Channel must be included for every message.<|end|><|start|>user<|message|>I want you to act as a SQL terminal in front of an example database, you need only to return the sql command to me.Below is an instruction that describes a task, Write a response that appropriately completes the request.
+"
+##Instruction:
+farm contains tables such as city, farm, farm_competition, competition_record. Table city has columns such as City_ID, Official_Name, Status, Area_km_2, Population, Census_Ranking. City_ID is the primary key.
+Table farm has columns such as Farm_ID, Year, Total_Horses, Working_Horses, Total_Cattle, Oxen, Bulls, Cows, Pigs, Sheep_and_Goats. Farm_ID is the primary key.
+Table farm_competition has columns such as Competition_ID, Year, Theme, Host_city_ID, Hosts. Competition_ID is the primary key.
+Table competition_record has columns such as Competition_ID, Farm_ID, Rank. Competition_ID is the primary key.
+The Host_city_ID of farm_competition is the foreign key of City_ID of city.
+The Farm_ID of competition_record is the foreign key of Farm_ID of farm.
+The Competition_ID of competition_record is the foreign key of Competition_ID of farm_competition.
+
+
+###Input:
+{context}
+
+###Response:<|end|><|start|>assistant<|channel|>final<|message|>"""  # noqa: E501
+
+EXPECTED_LORA_OUTPUT = [
+    "SELECT AVG(Working_Horses) FROM farm WHERE Total_Horses > 5000;",
+    "SELECT AVG(Working_Horses) FROM farm WHERE Total_Horses > 5000;",
+    "SELECT MAX(Cows) AS Max_Cows, MIN(Cows) AS Min_Cows FROM farm;",
+    "SELECT MAX(Cows) AS Max_Cows, MIN(Cows) AS Min_Cows FROM farm;",
+]
+
+
+def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
+    prompts = [
+        PROMPT_TEMPLATE.format(
+            context="What is the average number of working horses of farms with more than 5000 total number of horses?"  # noqa: E501
+        ),  # noqa: E501
+        PROMPT_TEMPLATE.format(
+            context="Give the average number of working horses on farms with more than 5000 total horses."  # noqa: E501
+        ),  # noqa: E501
+        PROMPT_TEMPLATE.format(
+            context="What are the maximum and minimum number of cows across all farms."
+        ),
+        PROMPT_TEMPLATE.format(
+            context="Return the maximum and minimum number of cows across all farms."
+        ),
+    ]
+    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,
+    )
+    # Print the outputs.
+    generated_texts: list[str] = []
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text.strip()
+        generated_texts.append(generated_text)
+        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])
+
+
+def test_gpt_oss_lora(gptoss20b_lora_files):
+    llm = vllm.LLM(
+        MODEL_PATH,
+        max_model_len=1024,
+        enable_lora=True,
+        max_loras=4,
+        max_lora_rank=8,
+        compilation_config=vllm.config.CompilationConfig(  # Avoid OOM
+            cudagraph_specialize_lora=False,
+        ),
+    )
+
+    generate_and_test(llm, gptoss20b_lora_files, lora_id=1)
+    generate_and_test(llm, gptoss20b_lora_files, lora_id=2)
+
+
+@multi_gpu_test(num_gpus=2)
+def test_gpt_oss_lora_tp2(gptoss20b_lora_files):
+    llm = vllm.LLM(
+        MODEL_PATH,
+        max_model_len=1024,
+        enable_lora=True,
+        max_loras=2,
+        max_lora_rank=8,
+        tensor_parallel_size=2,
+        compilation_config=vllm.config.CompilationConfig(  # Avoid OOM
+            cudagraph_specialize_lora=False,
+        ),
+    )
+
+    generate_and_test(llm, gptoss20b_lora_files, lora_id=1)
+    generate_and_test(llm, gptoss20b_lora_files, lora_id=2)

tests/lora/test_qwen3moe_tp.py

@@ -1,6 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+
+# NOTE To avoid overloading the CI pipeline, this test script will not
+# be triggered on CI and is primarily intended for local testing and verification.
+
 import vllm
 from vllm.lora.request import LoRARequest