[CI] Fix model swap test failures

- Responses API: Qwen3-1.7B too small for reasoning + tool calling,
  bump to Qwen3-4B (~8 GiB, fits on 17GB MIG slice)
- MM cache stats: llava-onevision produces more multimodal tokens per
  image than llava-1.5, increase max_model_len from 4096 to 8192

Signed-off-by: Kevin H. Luu <khluu000@gmail.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.buildkite/scripts/run-benchmarks.sh

@@ -18,7 +18,7 @@ vllm bench throughput --input-len 256 --output-len 256 --output-json throughput_
 bench_throughput_exit_code=$?
 
 # run server-based benchmarks and upload the result to buildkite
-vllm serve meta-llama/Llama-2-7b-chat-hf &
+vllm serve Qwen/Qwen3-0.6B &
 server_pid=$!
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 
@@ -28,10 +28,10 @@ vllm bench serve \
     --backend vllm \
     --dataset-name sharegpt \
     --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json \
-    --model meta-llama/Llama-2-7b-chat-hf \
+    --model Qwen/Qwen3-0.6B \
     --num-prompts 20 \
     --endpoint /v1/completions \
-    --tokenizer meta-llama/Llama-2-7b-chat-hf \
+    --tokenizer Qwen/Qwen3-0.6B \
     --save-result \
     2>&1 | tee benchmark_serving.txt
 bench_serving_exit_code=$?

.buildkite/test_areas/misc.yaml

@@ -106,7 +106,7 @@ steps:
      # for basic
     - python3 basic/offline_inference/chat.py
     - python3 basic/offline_inference/generate.py --model facebook/opt-125m
-    - python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
+    - python3 basic/offline_inference/generate.py --model meta-llama/Llama-3.1-8B-Instruct --cpu-offload-gb 4
     - python3 basic/offline_inference/classify.py
     - python3 basic/offline_inference/embed.py
     - python3 basic/offline_inference/score.py

tests/entrypoints/llm/test_mm_cache_stats.py

@@ -63,8 +63,8 @@ def test_mm_cache_stats(
     caplog_vllm,
 ):
     llm = LLM(
-        model="llava-hf/llava-1.5-7b-hf",
-        max_model_len=4096,
+        model="llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
+        max_model_len=8192,
         max_num_seqs=5,
         enforce_eager=True,
         mm_processor_cache_type=mm_processor_cache_type,

tests/entrypoints/llm/test_prompt_validation.py

@@ -21,7 +21,7 @@ def test_out_of_vocab_token():
 
 def test_require_mm_embeds():
     llm = LLM(
-        model="llava-hf/llava-1.5-7b-hf",
+        model="llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
         enforce_eager=True,
         enable_mm_embeds=False,
     )

tests/entrypoints/openai/responses/test_parsable_context.py

@@ -20,7 +20,7 @@ from .conftest import (
 
 logger = logging.getLogger(__name__)
 
-MODEL_NAME = "Qwen/Qwen3-8B"
+MODEL_NAME = "Qwen/Qwen3-4B"
 
 _PYTHON_TOOL_INSTRUCTION = (
     "You must use the Python tool to execute code. "

tests/entrypoints/openai/responses/test_simple.py

@@ -9,7 +9,7 @@ from tests.utils import RemoteOpenAIServer
 
 from .conftest import validate_streaming_event_stack
 
-MODEL_NAME = "Qwen/Qwen3-8B"
+MODEL_NAME = "Qwen/Qwen3-4B"
 
 
 @pytest.fixture(scope="module")

tests/entrypoints/openai/speech_to_text/test_transcription_validation.py

@@ -71,6 +71,7 @@ async def test_basic_audio(mary_had_lamb, model_name, rocm_aiter_fa_attention):
 
     if model_name.startswith("mistralai"):
         server_args += MISTRAL_FORMAT_ARGS
+        server_args += ["--max-model-len", "4096"]
 
     add_attention_backend(server_args, rocm_aiter_fa_attention)
 

tests/lora/test_add_lora.py

@@ -14,8 +14,8 @@ from vllm.lora.request import LoRARequest
 from vllm.sampling_params import SamplingParams
 from vllm.utils.async_utils import merge_async_iterators
 
-MODEL_PATH = "zai-org/chatglm3-6b"
-LORA_RANK = 64
+MODEL_PATH = "Qwen/Qwen3-0.6B"
+LORA_RANK = 8
 DEFAULT_MAX_LORAS = 4 * 3
 
 
@@ -54,7 +54,7 @@ async def requests_processing_time(llm, lora_requests: list[LoRARequest]) -> flo
 
 
 @pytest.mark.asyncio
-async def test_add_lora(chatglm3_lora_files):
+async def test_add_lora(qwen3_lora_files):
     """
     The add_lora function is used to preload some LoRA adapters into the
     engine in anticipation of future requests using these adapters. To test
@@ -64,7 +64,7 @@ async def test_add_lora(chatglm3_lora_files):
     We measure the request processing time in both cases and expect the time
     to be lesser in the case with add_lora() calls.
     """
-    lora_requests: list[LoRARequest] = get_lora_requests(chatglm3_lora_files)
+    lora_requests: list[LoRARequest] = get_lora_requests(qwen3_lora_files)
 
     max_loras = len(set([lr.lora_int_id for lr in lora_requests]))
     # Create engine in eager-mode. Due to high max_loras, the CI can
@@ -76,7 +76,6 @@ async def test_add_lora(chatglm3_lora_files):
         max_lora_rank=LORA_RANK,
         max_model_len=128,
         gpu_memory_utilization=0.8,  # avoid OOM
-        trust_remote_code=True,
         enforce_eager=True,
     )
 

tests/v1/e2e/general/test_kv_sharing_fast_prefill.py

@@ -87,6 +87,7 @@ def test_kv_sharing_fast_prefill(
 
         llm = LLM(
             model="google/gemma-3n-E2B-it",
+            max_model_len=16384,
             enforce_eager=enforce_eager,
             compilation_config=compilation_config,
             seed=SEED,