test_vllm.py

# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

import base64
import dataclasses
import logging
import os
import random
from dataclasses import dataclass, field
from pathlib import Path
from typing import Optional

import pytest

from tests.serve.common import (
    WORKSPACE_DIR,
    params_with_model_mark,
    run_serve_deployment,
)
from tests.serve.conftest import MULTIMODAL_IMG_URL, get_multimodal_test_image_bytes
from tests.serve.lora_utils import MinioLoraConfig
from tests.utils.constants import DefaultPort
from tests.utils.engine_process import EngineConfig
from tests.utils.payload_builder import (
    cached_tokens_chat_payload,
    chat_payload,
    chat_payload_default,
    chat_payload_with_logprobs,
    completion_payload_default,
    completion_payload_with_logprobs,
    metric_payload_default,
)
from tests.utils.payloads import LoraTestChatPayload, ToolCallingChatPayload

logger = logging.getLogger(__name__)


def _is_cuda13() -> bool:
    v = os.environ.get("CUDA_VERSION", "")
    # handles "13", "13.0", "13.0.1", etc.
    return v.startswith("13")


@dataclass
class VLLMConfig(EngineConfig):
    """Configuration for vLLM test scenarios"""

    stragglers: list[str] = field(default_factory=lambda: ["VLLM:EngineCore"])


vllm_dir = os.environ.get("VLLM_DIR") or os.path.join(
    WORKSPACE_DIR, "examples/backends/vllm"
)
LOCAL_VIDEO_TEST_PATH = Path(
    WORKSPACE_DIR, "lib/llm/tests/data/media/240p_10.mp4"
).resolve()
LOCAL_VIDEO_TEST_URI = LOCAL_VIDEO_TEST_PATH.as_uri()


# vLLM test configurations
# NOTE: pytest.mark.gpu_1 tests take ~5.5 minutes total to run sequentially (with models pre-cached)
# TODO: Now that these tests use dynamic ports and each config has VRAM markers,
# optimize the runtime by bin-packing multiple engine deployments in parallel on the same GPU.
# A future collector/launcher can sum profiled_vram_gib values to decide how many tests fit
# concurrently without exceeding available VRAM.
vllm_configs = {
    "aggregated": VLLMConfig(
        name="aggregated",
        directory=vllm_dir,
        script_name="agg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(
                360
            ),  # ~8.5x observed 42.2s; bumped for GPU-parallel headroom
            pytest.mark.pre_merge,
        ],
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
            chat_payload(
                "Can you write me a song?",
                repeat_count=1,
                expected_response=["song"],
                temperature=0.0,
                max_tokens=32,
                extra_body={
                    "stop": ["song"],
                    "include_stop_str_in_output": True,
                },
            ),
            metric_payload_default(min_num_requests=6, backend="vllm"),
        ],
    ),
    "aggregated_logprobs": VLLMConfig(
        name="aggregated_logprobs",
        directory=vllm_dir,
        script_name="agg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(120),  # ~5x observed 24.3s; CI machines are slower
            pytest.mark.post_merge,
        ],
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_with_logprobs(
                repeat_count=2,
                expected_response=["AI", "knock", "joke"],
                max_tokens=30,
                temperature=0.0,
                top_logprobs=3,
            ),
            completion_payload_with_logprobs(
                repeat_count=2,
                expected_response=["AI", "knock", "joke"],
                max_tokens=30,
                temperature=0.0,
                logprobs=5,
            ),
        ],
    ),
    "aggregated_lmcache": VLLMConfig(
        name="aggregated_lmcache",
        directory=vllm_dir,
        script_name="agg_lmcache.sh",
        marks=[
            pytest.mark.lmcache,
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(360),  # ~7x observed 49.0s; old value before profiling
            pytest.mark.pre_merge,
            pytest.mark.skipif(
                _is_cuda13(),
                reason="lmcache does not support CUDA 13 as of v0.3.11",
            ),
        ],
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
            metric_payload_default(min_num_requests=6, backend="vllm"),
            metric_payload_default(min_num_requests=6, backend="lmcache"),
        ],
    ),
    "aggregated_lmcache_multiproc": VLLMConfig(
        name="aggregated_lmcache_multiproc",
        directory=vllm_dir,
        script_name="agg_lmcache_multiproc.sh",
        marks=[
            pytest.mark.lmcache,
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(360),  # ~7x observed 49.3s; old value before profiling
            pytest.mark.pre_merge,
            pytest.mark.skipif(
                _is_cuda13(),
                reason="lmcache does not support CUDA 13 as of v0.3.11",
            ),
        ],
        model="Qwen/Qwen3-0.6B",
        env={
            "PROMETHEUS_MULTIPROC_DIR": f"/tmp/prometheus_multiproc_test_{os.getpid()}_{random.randint(0, 10000)}"
        },
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
            metric_payload_default(min_num_requests=6, backend="vllm"),
            metric_payload_default(min_num_requests=6, backend="lmcache"),
        ],
    ),
    "agg-request-plane-tcp": VLLMConfig(
        name="agg-request-plane-tcp",
        directory=vllm_dir,
        script_name="agg_request_planes.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(
                360
            ),  # ~8x observed 43.0s; bumped for GPU-parallel headroom
            pytest.mark.pre_merge,
        ],
        model="Qwen/Qwen3-0.6B",
        script_args=["--tcp"],
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
    "agg-request-plane-http": VLLMConfig(
        name="agg-request-plane-http",
        directory=vllm_dir,
        script_name="agg_request_planes.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(
                360
            ),  # ~8.5x observed 42.3s; bumped for GPU-parallel headroom
            pytest.mark.pre_merge,
        ],
        model="Qwen/Qwen3-0.6B",
        script_args=["--http"],
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
    "agg-router": VLLMConfig(
        name="agg-router",
        directory=vllm_dir,
        script_name="agg_router.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.pre_merge,
            pytest.mark.skip(reason="DYN-2263"),
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_default(
                expected_log=[
                    r"ZMQ listener .* received batch with \d+ events \(seq=\d+(?:, [^)]*)?\)",
                    r"Event processor for worker_id \d+ processing event: Stored\(",
                    r"Selected worker: worker_type=\w+, worker_id=\d+ dp_rank=.*?, logit: ",
                ]
            )
        ],
        env={
            "DYN_LOG": "dynamo_llm::kv_router::publisher=trace,dynamo_kv_router::scheduling::selector=info",
        },
    ),
    "agg-router-approx": VLLMConfig(
        name="agg-router-approx",
        directory=vllm_dir,
        script_name="agg_router_approx.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.pre_merge,
            pytest.mark.skip(reason="DYN-2264"),
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            # Test approximate KV routing (--no-kv-events mode)
            # Repeated requests should show cache-aware routing in logs
            chat_payload_default(
                repeat_count=3,
                expected_log=[
                    # Verify scheduler is selecting workers with cache awareness
                    r"Selected worker: worker_type=\w+, worker_id=\d+ dp_rank=.*?, logit: ",
                    # After first request, should see cached blocks being tracked
                    r"with \d+ cached blocks",
                ],
            ),
            # Also test with cached tokens payload to verify usage field
            cached_tokens_chat_payload(
                repeat_count=3,
                expected_log=[
                    # Verify routing decision shows cache hits
                    r"with \d+ cached blocks",
                ],
            ),
        ],
        env={
            "DYN_LOG": "dynamo_kv_router::scheduling::selector=info",
        },
    ),
    "disaggregated": VLLMConfig(
        name="disaggregated",
        directory=vllm_dir,
        script_name="disagg.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.pre_merge,
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
    "deepep": VLLMConfig(
        name="deepep",
        directory=vllm_dir,
        script_name="dsr1_dep.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.vllm,
            pytest.mark.h100,
            pytest.mark.nightly,
            # TODO: profile to get max_vram and timeout
        ],
        model="deepseek-ai/DeepSeek-V2-Lite",
        script_args=[
            "--model",
            "deepseek-ai/DeepSeek-V2-Lite",
            "--num-nodes",
            "1",
            "--node-rank",
            "0",
            "--gpus-per-node",
            "2",
        ],
        timeout=700,
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
    # NOTE: Pack all workers on 1 GPU for lower CI resource requirements
    # NOTE: disagg_multimodal_e_pd.sh uses explicit --gpu-memory-utilization via
    # DYN_ENCODE_GPU_MEM / DYN_PD_GPU_MEM env vars in single-GPU mode.
    # PD worker honors build_gpu_mem_args for parallel execution.
    "multimodal_e_pd_qwen": VLLMConfig(
        name="multimodal_e_pd_qwen",
        directory=vllm_dir,
        script_name="disagg_multimodal_e_pd.sh",
        marks=[
            pytest.mark.gpu_1,
            # No profiled_vram_gib / requested_vllm_kv_cache_bytes: single-GPU mode
            # uses hardcoded fractions (encode=0.1, PD=0.7) that scale with GPU size.
            pytest.mark.timeout(340),  # ~5x observed 68.4s; 2B model loads slower on CI
            pytest.mark.pre_merge,
        ],
        model="Qwen/Qwen3-VL-2B-Instruct",
        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
        request_payloads=[
            chat_payload(
                [
                    {
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
                    },
                    {
                        "type": "image_url",
                        "image_url": {"url": MULTIMODAL_IMG_URL},
                    },
                ],
                repeat_count=1,
                # With proper prompt templating, the model actually only returns "green",
                # verified behavior with native vLLM.
                expected_response=["green"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    "multimodal_agg_frontend_decoding": VLLMConfig(
        name="multimodal_agg_frontend_decoding",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        # post_merge because needs real NIXL not stub
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(9.6),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_710_490_000
            ),  # KV cache cap (2x safety over min=855_244_800)
            pytest.mark.timeout(220),  # ~5x observed 43.7s; 2B model loads slower on CI
            pytest.mark.post_merge,
        ],
        model="Qwen/Qwen2-VL-2B-Instruct",
        # Pass --frontend-decoding to enable Rust frontend image decoding + NIXL RDMA transfer
        script_args=[
            "--model",
            "Qwen/Qwen2-VL-2B-Instruct",
            "--frontend-decoding",
        ],
        request_payloads=[
            chat_payload(
                [
                    {
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
                    },
                    {
                        "type": "image_url",
                        "image_url": {"url": MULTIMODAL_IMG_URL},
                    },
                ],
                repeat_count=1,
                expected_response=["green"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    # NOTE: Pack all workers on 1 GPU for lower CI resource requirements.
    # NOTE: disagg_multimodal_epd.sh uses --kv-cache-memory-bytes=512MB for P/D
    # workers. Per vLLM CacheConfig, kv_cache_memory_bytes (when not-None) ignores
    # gpu_memory_utilization (ref: https://docs.vllm.ai/en/stable/api/vllm/config/cache/),
    # so KV cache overrides have no effect. Regardless of GPU_MEM
    # fractions (0.1/0.4/0.4), the 3 workers combined consistently use ~17.6 GiB
    # total on this GPU.
    # NOTE: disagg_multimodal_epd.sh uses explicit --gpu-memory-utilization via
    # DYN_ENCODE_GPU_MEM / DYN_PREFILL_GPU_MEM / DYN_DECODE_GPU_MEM env vars.
    # P/D workers honor build_gpu_mem_args for parallel execution.
    "multimodal_disagg_qwen": VLLMConfig(
        name="multimodal_disagg_qwen",
        directory=vllm_dir,
        script_name="disagg_multimodal_epd.sh",
        marks=[
            pytest.mark.gpu_1,
            # No profiled_vram_gib / requested_vllm_kv_cache_bytes: single-GPU mode
            # uses hardcoded fractions via DYN_*_GPU_MEM that scale with GPU size.
            pytest.mark.pre_merge,
        ],
        model="Qwen/Qwen3-VL-2B-Instruct",
        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
        timeout=300,
        request_payloads=[
            chat_payload(
                [
                    {
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
                    },
                    {
                        "type": "image_url",
                        "image_url": {"url": MULTIMODAL_IMG_URL},
                    },
                ],
                repeat_count=1,
                expected_response=["green"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    "multimodal_agg_qwen": VLLMConfig(
        name="multimodal_agg_qwen",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(19.9),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                922_354_000
            ),  # KV cache cap (2x safety over min=461_176_832)
            pytest.mark.timeout(
                360
            ),  # ~7x observed 50.0s; 7B model loads ~48s on CI (A10G/L4)
            pytest.mark.post_merge,
        ],
        model="Qwen/Qwen2.5-VL-7B-Instruct",
        script_args=["--model", "Qwen/Qwen2.5-VL-7B-Instruct"],
        delayed_start=0,
        timeout=360,
        request_payloads=[
            chat_payload(
                [
                    {
                        "type": "text",
                        "text": "What colors are in the following image? Respond only with the colors.",
                    },
                    {
                        "type": "image_url",
                        "image_url": {"url": MULTIMODAL_IMG_URL},
                    },
                ],
                repeat_count=1,
                expected_response=["purple"],
                max_tokens=100,
            ),
        ],
    ),
    "multimodal_agg_llava": VLLMConfig(
        name="multimodal_agg_llava",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(14.9),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                922_354_000
            ),  # KV cache cap (2x safety over min=461_176_832)
            pytest.mark.timeout(
                300
            ),  # ~7x observed 42.7s; 7B model loads ~48s on CI (A10G/L4)
            pytest.mark.nightly,
            # https://github.com/ai-dynamo/dynamo/issues/4501
            pytest.mark.xfail(strict=False),
        ],
        model="llava-hf/llava-1.5-7b-hf",
        script_args=["--model", "llava-hf/llava-1.5-7b-hf"],
        delayed_start=0,
        timeout=360,
        request_payloads=[
            # HTTP URL test
            chat_payload(
                [
                    {"type": "text", "text": "What is in this image?"},
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": "http://images.cocodataset.org/test2017/000000155781.jpg"
                        },
                    },
                ],
                repeat_count=1,
                expected_response=["bus"],
                temperature=0.0,
            ),
            # String content test - verifies string → array conversion for multimodal templates
            chat_payload_default(
                repeat_count=1,
                expected_response=[],  # Just validate no error
            ),
        ],
    ),
    # Video multimodal tests for CI use the canonical aggregated multimodal launcher.
    "multimodal_video_agg": VLLMConfig(
        name="multimodal_video_agg",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen3-VL-2B-Instruct",
        delayed_start=60,  # Video models require longer loading time
        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct"],
        timeout=600,  # 10 minutes for video processing overhead
        request_payloads=[
            chat_payload(
                [
                    {"type": "text", "text": "Describe the video in detail"},
                    {
                        "type": "video_url",
                        "video_url": {"url": LOCAL_VIDEO_TEST_URI},
                    },
                ],
                repeat_count=1,
                expected_response=["red", "static", "still"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    "multimodal_video_disagg": VLLMConfig(
        name="multimodal_video_disagg",
        directory=vllm_dir,
        script_name="disagg_multimodal_epd.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.pre_merge,
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen3-VL-2B-Instruct",
        delayed_start=60,  # Video models require longer loading time
        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
        timeout=600,  # 10 minutes for video processing overhead
        request_payloads=[
            chat_payload(
                [
                    {"type": "text", "text": "Describe the video in detail"},
                    {
                        "type": "video_url",
                        "video_url": {"url": LOCAL_VIDEO_TEST_URI},
                    },
                ],
                repeat_count=1,
                expected_response=["red", "static", "still"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    # Audio multimodal tests for nightly CI pipeline
    # These tests validate audio inference capabilities with Qwen2-Audio model
    "multimodal_audio_agg": VLLMConfig(
        name="multimodal_audio_agg",
        directory=os.path.join(WORKSPACE_DIR, "examples/multimodal"),
        script_name="audio_agg.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.nightly,
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen2-Audio-7B-Instruct",
        delayed_start=60,  # Audio models require longer loading time
        script_args=["--model", "Qwen/Qwen2-Audio-7B-Instruct"],
        timeout=600,  # 10 minutes for audio processing overhead
        request_payloads=[
            chat_payload(
                [
                    {"type": "text", "text": "What is recited in the audio?"},
                    {
                        "type": "audio_url",
                        "audio_url": {
                            "url": "https://raw.githubusercontent.com/yuekaizhang/Triton-ASR-Client/main/datasets/mini_en/wav/1221-135766-0002.wav"
                        },
                    },
                ],
                repeat_count=1,
                expected_response=["Hester", "Pynne"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    "multimodal_audio_disagg": VLLMConfig(
        name="multimodal_audio_disagg",
        directory=os.path.join(WORKSPACE_DIR, "examples/multimodal"),
        script_name="audio_disagg.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.nightly,
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen2-Audio-7B-Instruct",
        delayed_start=60,  # Audio models require longer loading time
        script_args=["--model", "Qwen/Qwen2-Audio-7B-Instruct"],
        timeout=600,  # 10 minutes for audio processing overhead
        request_payloads=[
            chat_payload(
                [
                    {"type": "text", "text": "What is recited in the audio?"},
                    {
                        "type": "audio_url",
                        "audio_url": {
                            "url": "https://raw.githubusercontent.com/yuekaizhang/Triton-ASR-Client/main/datasets/mini_en/wav/1221-135766-0002.wav"
                        },
                    },
                ],
                repeat_count=1,
                expected_response=["Hester", "Pynne"],
                temperature=0.0,
                max_tokens=100,
            )
        ],
    ),
    "aggregated_toolcalling": VLLMConfig(
        name="aggregated_toolcalling",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.multimodal,
            pytest.mark.nightly,
        ],  # TODO: profile to get max_vram and timeout
        model="Qwen/Qwen3-VL-30B-A3B-Instruct-FP8",
        script_args=[
            "--model",
            "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8",
            "--max-model-len",
            "10000",
            "--dyn-tool-call-parser",
            "hermes",
        ],
        delayed_start=0,
        timeout=600,
        request_payloads=[
            ToolCallingChatPayload(
                body={
                    "messages": [
                        {
                            "role": "user",
                            "content": [
                                {
                                    "type": "text",
                                    "text": "Describe what you see in this image in detail.",
                                },
                                {
                                    "type": "image_url",
                                    "image_url": {"url": MULTIMODAL_IMG_URL},
                                },
                            ],
                        }
                    ],
                    "tools": [
                        {
                            "type": "function",
                            "function": {
                                "name": "describe_image",
                                "description": "Provides detailed description of objects and scenes in an image",
                                "parameters": {
                                    "type": "object",
                                    "properties": {
                                        "objects": {
                                            "type": "array",
                                            "items": {"type": "string"},
                                            "description": "List of objects detected in the image",
                                        },
                                        "scene": {
                                            "type": "string",
                                            "description": "Overall scene description",
                                        },
                                    },
                                    "required": ["objects", "scene"],
                                },
                            },
                        }
                    ],
                    "tool_choice": "auto",
                    "max_tokens": 1024,
                },
                repeat_count=1,
                expected_response=["purple"],  # Validate image understanding
                expected_log=[],
                expected_tool_name="describe_image",  # Validate tool call happened
            )
        ],
    ),
    # TODO: Enable this test case when we have 4 GPUs runners.
    # "multimodal_disagg": VLLMConfig(
    #     name="multimodal_disagg",
    #     directory=os.path.join(WORKSPACE_DIR, "examples/multimodal"),
    #     script_name="disagg.sh",
    #     marks=[pytest.mark.gpu_4, pytest.mark.vllm],
    #     model="llava-hf/llava-1.5-7b-hf",
    #     delayed_start=45,
    #     script_args=["--model", "llava-hf/llava-1.5-7b-hf"],
    # ),
    "completions_only": VLLMConfig(
        name="completions_only",
        directory=vllm_dir,
        script_name="agg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(18.3),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                4_074_898_000
            ),  # KV cache cap (2x safety over min=2_037_448_704)
            pytest.mark.timeout(
                420
            ),  # 7B model loads ~48s on CI (A10G/L4) vs ~15s locally
            pytest.mark.post_merge,
        ],
        model="deepseek-ai/deepseek-llm-7b-base",
        script_args=[
            "--model",
            "deepseek-ai/deepseek-llm-7b-base",
            "--dyn-endpoint-types",
            "completions",
        ],
        request_payloads=[
            completion_payload_default(),
        ],
    ),
    "multi_node_tp_headless": VLLMConfig(
        name="multi_node_tp_headless",
        directory=os.path.join(WORKSPACE_DIR, "tests/serve"),
        script_name="multi_node_tp_headless.sh",
        marks=[
            pytest.mark.gpu_2,
            pytest.mark.pre_merge,
            # TODO: profile to get max_vram
            pytest.mark.timeout(300),
        ],
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload_default(),
            completion_payload_default(),
        ],
    ),
    "guided_decoding": VLLMConfig(
        name="guided_decoding",
        directory=vllm_dir,
        script_name="agg.sh",
        marks=[
            pytest.mark.gpu_1,
            pytest.mark.profiled_vram_gib(3.8),  # actual profiled peak with kv-bytes
            pytest.mark.requested_vllm_kv_cache_bytes(
                1_119_388_000
            ),  # KV cache cap (2x safety over min=559_693_824)
            pytest.mark.timeout(110),  # ~5x observed 22.3s; CI machines are slower
            pytest.mark.pre_merge,
        ],
        model="Qwen/Qwen3-0.6B",
        request_payloads=[
            chat_payload(
                "Generate a person with name and age",
                repeat_count=1,
                expected_response=['"name"', '"age"'],
                temperature=0.0,
                max_tokens=100,
                extra_body={
                    "guided_json": {
                        "type": "object",
                        "properties": {
                            "name": {"type": "string"},
                            "age": {"type": "integer"},
                        },
                        "required": ["name", "age"],
                    }
                },
            ),
            chat_payload(
                "Generate a color name (red, blue, or green)",
                repeat_count=1,
                expected_response=["red", "blue", "green"],
                temperature=0.0,
                max_tokens=20,
                extra_body={"guided_regex": r"(red|blue|green)"},
            ),
            chat_payload(
                "Generate a color name (red, blue, or green)",
                repeat_count=1,
                expected_response=["red", "blue", "green"],
                temperature=0.0,
                max_tokens=20,
                extra_body={"guided_choice": ["red", "blue", "green"]},
            ),
        ],
    ),
}


@pytest.fixture(params=params_with_model_mark(vllm_configs))
def vllm_config_test(request):
    """Fixture that provides different vLLM test configurations"""
    return vllm_configs[request.param]


@pytest.mark.vllm
@pytest.mark.e2e
def test_serve_deployment(
    vllm_config_test,
    request,
    runtime_services_dynamic_ports,
    dynamo_dynamic_ports,
    predownload_models,
    image_server,
):
    """
    Test dynamo serve deployments with different graph configurations.
    """
    config = dataclasses.replace(
        vllm_config_test, frontend_port=dynamo_dynamic_ports.frontend_port
    )
    run_serve_deployment(config, request, ports=dynamo_dynamic_ports)


@pytest.mark.vllm
@pytest.mark.e2e
@pytest.mark.gpu_2
@pytest.mark.nightly
@pytest.mark.timeout(360)  # Match VLLMConfig.timeout for this multimodal deployment
def test_multimodal_b64(
    request,
    runtime_services_dynamic_ports,
    dynamo_dynamic_ports,
    predownload_models,
):
    """
    Test multimodal inference with base64 url passthrough.

    This test is separate because it loads the required image at runtime
    (not collection time), ensuring it only fails when actually executed.
    """
    # Load B64 image at test execution time (uses real PNG even if MULTIMODAL_IMG is LFS pointer)
    b64_img = base64.b64encode(get_multimodal_test_image_bytes()).decode()

    # Create payload with B64 image
    b64_payload = chat_payload(
        [
            {
                "type": "text",
                "text": "What colors are in the following image? Respond only with the colors.",
            },
            {
                "type": "image_url",
                "image_url": {"url": f"data:image/png;base64,{b64_img}"},
            },
        ],
        repeat_count=1,
        expected_response=["purple"],
        max_tokens=100,
    )

    # Create test config
    config = VLLMConfig(
        name="test_multimodal_b64",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[],  # markers at function-level
        model="Qwen/Qwen2.5-VL-7B-Instruct",
        script_args=["--model", "Qwen/Qwen2.5-VL-7B-Instruct"],
        delayed_start=0,
        timeout=360,
        request_payloads=[b64_payload],
    )

    config = dataclasses.replace(
        config, frontend_port=dynamo_dynamic_ports.frontend_port
    )
    run_serve_deployment(config, request, ports=dynamo_dynamic_ports)


@pytest.mark.vllm
@pytest.mark.e2e
@pytest.mark.gpu_1
@pytest.mark.pre_merge
@pytest.mark.timeout(220)
def test_multimodal_b64_frontend_decoding(
    request,
    runtime_services_dynamic_ports,
    dynamo_dynamic_ports,
    predownload_models,
):
    """
    Test multimodal inference with base64 images through frontend decoding path.

    This exercises the Rust frontend image decode + NIXL RDMA transfer path
    with inline base64 data: URIs (not HTTP URLs). Verifies that the
    strip_inline_data_urls optimization does not break correctness.
    """
    b64_img = base64.b64encode(get_multimodal_test_image_bytes()).decode()

    b64_payload = chat_payload(
        [
            {
                "type": "text",
                "text": "What colors are in the following image? Respond only with the colors.",
            },
            {
                "type": "image_url",
                "image_url": {"url": f"data:image/png;base64,{b64_img}"},
            },
        ],
        repeat_count=1,
        expected_response=["green"],
        temperature=0.0,
        max_tokens=100,
    )

    config = VLLMConfig(
        name="test_multimodal_b64_frontend_decoding",
        directory=vllm_dir,
        script_name="agg_multimodal.sh",
        marks=[],
        model="Qwen/Qwen3-VL-2B-Instruct",
        script_args=[
            "--model",
            "Qwen/Qwen3-VL-2B-Instruct",
            "--frontend-decoding",
        ],
        delayed_start=0,
        timeout=220,
        request_payloads=[b64_payload],
    )

    config = dataclasses.replace(
        config, frontend_port=dynamo_dynamic_ports.frontend_port
    )
    run_serve_deployment(config, request, ports=dynamo_dynamic_ports)


# LoRA Test Directory
lora_dir = os.path.join(vllm_dir, "launch/lora")


def lora_chat_payload(
    lora_name: str,
    s3_uri: str,
    system_port: int = DefaultPort.SYSTEM1.value,
    repeat_count: int = 2,
    expected_response: Optional[list] = None,
    expected_log: Optional[list] = None,
    max_tokens: int = 100,
    temperature: float = 0.0,
) -> LoraTestChatPayload:
    """Create a LoRA-enabled chat payload for testing"""
    return LoraTestChatPayload(
        body={
            "model": lora_name,
            "messages": [
                {
                    "role": "user",
                    "content": "What is deep learning? Answer in one sentence.",
                }
            ],
            "max_tokens": max_tokens,
            "temperature": temperature,
            "stream": False,
        },
        lora_name=lora_name,
        s3_uri=s3_uri,
        system_port=system_port,
        repeat_count=repeat_count,
        expected_response=expected_response
        or ["learning", "neural", "network", "AI", "model"],
        expected_log=expected_log or [],
    )


@pytest.mark.vllm
@pytest.mark.e2e
@pytest.mark.gpu_1
@pytest.mark.model("Qwen/Qwen3-0.6B")
@pytest.mark.timeout(600)
@pytest.mark.post_merge
@pytest.mark.skip(reason="DYN-2260")
def test_lora_aggregated(
    request,
    runtime_services_dynamic_ports,
    predownload_models,
    minio_lora_service,
    dynamo_dynamic_ports,
):
    """
    Test LoRA inference with aggregated vLLM deployment.

    This test:
    1. Uses MinIO fixture to provide S3-compatible storage with uploaded LoRA
    2. Starts vLLM with LoRA support enabled
    3. Loads the LoRA adapter via system API
    4. Runs inference with the LoRA model
    """
    minio_config: MinioLoraConfig = minio_lora_service

    # Create payload that loads LoRA and tests inference
    lora_payload = lora_chat_payload(
        lora_name=minio_config.lora_name,
        s3_uri=minio_config.get_s3_uri(),
        system_port=DefaultPort.SYSTEM1.value,
        repeat_count=2,
    )

    # Create test config with MinIO environment variables
    config = VLLMConfig(
        name="test_lora_aggregated",
        directory=vllm_dir,
        script_name="lora/agg_lora.sh",
        marks=[],  # markers at function-level
        model="Qwen/Qwen3-0.6B",
        timeout=600,
        env=minio_config.get_env_vars(),
        request_payloads=[lora_payload],
    )

    config = dataclasses.replace(
        config, frontend_port=dynamo_dynamic_ports.frontend_port
    )
    run_serve_deployment(
        config,
        request,
        ports=dynamo_dynamic_ports,
        extra_env=minio_config.get_env_vars(),
    )


@pytest.mark.vllm
@pytest.mark.e2e
@pytest.mark.gpu_2
@pytest.mark.model("Qwen/Qwen3-0.6B")
@pytest.mark.timeout(600)
@pytest.mark.pre_merge
@pytest.mark.parametrize("num_system_ports", [2], indirect=True)
def test_lora_aggregated_router(
    request,
    runtime_services_dynamic_ports,
    predownload_models,
    minio_lora_service,
    dynamo_dynamic_ports,
    num_system_ports,
):
    """
    Test LoRA inference with aggregated vLLM deployment using KV router.

    This test:
    1. Uses MinIO fixture to provide S3-compatible storage with uploaded LoRA
    2. Starts multiple vLLM workers with LoRA support and KV router
    3. Loads the LoRA adapter on both workers via system API
    4. Runs inference with the LoRA model, verifying KV cache routing
    """
    assert (
        num_system_ports >= 2
    ), "serve tests require at least SYSTEM_PORT1 + SYSTEM_PORT2"
    minio_config: MinioLoraConfig = minio_lora_service

    # Create payloads that load LoRA on both workers and test inference
    # Worker 1 (DefaultPort.SYSTEM1)
    lora_payload_worker1 = lora_chat_payload(
        lora_name=minio_config.lora_name,
        s3_uri=minio_config.get_s3_uri(),
        system_port=DefaultPort.SYSTEM1.value,
        repeat_count=1,
    )

    # Worker 2 (DefaultPort.SYSTEM2)
    lora_payload_worker2 = lora_chat_payload(
        lora_name=minio_config.lora_name,
        s3_uri=minio_config.get_s3_uri(),
        system_port=DefaultPort.SYSTEM2.value,
        repeat_count=1,
    )

    # Additional inference payload to test routing (LoRA already loaded)
    inference_payload = chat_payload(
        content="Explain machine learning in simple terms.",
        repeat_count=2,
        expected_response=["learn", "data", "algorithm", "model", "pattern"],
        max_tokens=150,
        temperature=0.0,
    ).with_model(minio_config.lora_name)

    # Add env vars including PYTHONHASHSEED for deterministic KV event IDs
    env_vars = minio_config.get_env_vars()
    env_vars["PYTHONHASHSEED"] = "0"

    # Create test config with MinIO environment variables
    config = VLLMConfig(
        name="test_lora_aggregated_router",
        directory=vllm_dir,
        script_name="lora/agg_lora_router.sh",
        marks=[],  # markers at function-level
        model="Qwen/Qwen3-0.6B",
        timeout=600,
        env=env_vars,
        request_payloads=[
            lora_payload_worker1,
            lora_payload_worker2,
            inference_payload,
        ],
    )

    config = dataclasses.replace(
        config, frontend_port=dynamo_dynamic_ports.frontend_port
    )
    run_serve_deployment(
        config, request, ports=dynamo_dynamic_ports, extra_env=env_vars
    )