[V1] Enable multi-input by default

docs/source/models/supported_models.md

@@ -759,7 +759,7 @@ On the other hand, modalities separated by `/` are mutually exclusive.
 See [this page](#multimodal-inputs) on how to pass multi-modal inputs to the model.
 
 :::{important}
-To enable multiple multi-modal items per text prompt, you have to set `limit_mm_per_prompt` (offline inference)
+**To enable multiple multi-modal items per text prompt in vLLM V0**, you have to set `limit_mm_per_prompt` (offline inference)
 or `--limit-mm-per-prompt` (online serving). For example, to enable passing up to 4 images per text prompt:
 
 Offline inference:
@@ -777,6 +777,8 @@ Online serving:
 vllm serve Qwen/Qwen2-VL-7B-Instruct --limit-mm-per-prompt image=4
 ```
 
+**This is no longer required if you are using vLLM V1.**
+
 :::
 
 :::{note}

docs/source/serving/offline_inference.md

@@ -110,6 +110,30 @@ If you run out of CPU RAM, try the following options:
 - (Multi-modal models only) you can set the size of multi-modal input cache using `VLLM_MM_INPUT_CACHE_GIB` environment variable (default 4 GiB).
 - (CPU backend only) you can set the size of KV cache using `VLLM_CPU_KVCACHE_SPACE` environment variable (default 4 GiB).
 
+#### Disable unused modalities
+
+You can disable unused modalities (except for text) by setting its limit to zero.
+
+For example, if your application only accepts image input, there is no need to allocate any memory for videos.
+
+```python
+from vllm import LLM
+
+# Accept images but not videos
+llm = LLM(model="Qwen/Qwen2.5-VL-3B-Instruct",
+          limit_mm_per_prompt={"video": 0})
+```
+
+You can even run a multi-modal model for text-only inference:
+
+```python
+from vllm import LLM
+
+# Don't accept images. Just text.
+llm = LLM(model="google/gemma-3-27b-it",
+          limit_mm_per_prompt={"image": 0})
+```
+
 ### Performance optimization and tuning
 
 You can potentially improve the performance of vLLM by finetuning various options.

examples/offline_inference/audio_language.py

@@ -196,6 +196,11 @@ def main(args):
     req_data = model_example_map[model](question_per_audio_count[audio_count],
                                         audio_count)
 
+    # Disable other modalities to save memory
+    default_limits = {"image": 0, "video": 0, "audio": 0}
+    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
+        req_data.engine_args.limit_mm_per_prompt or {})
+
     engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
     llm = LLM(**engine_args)
 

examples/offline_inference/encoder_decoder_multimodal.py

@@ -133,6 +133,11 @@ def main(args):
 
     req_data = model_example_map[model]()
 
+    # Disable other modalities to save memory
+    default_limits = {"image": 0, "video": 0, "audio": 0}
+    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
+        req_data.engine_args.limit_mm_per_prompt or {})
+
     engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
     llm = LLM(**engine_args)
 

examples/offline_inference/vision_language.py

@@ -45,7 +45,7 @@ def run_aria(questions: list[str], modality: str) -> ModelRequestData:
         max_model_len=4096,
         max_num_seqs=2,
         dtype="bfloat16",
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [(f"<|im_start|>user\n<fim_prefix><|img|><fim_suffix>{question}"
@@ -71,7 +71,7 @@ def run_aya_vision(questions: list[str], modality: str) -> ModelRequestData:
         max_model_len=2048,
         max_num_seqs=2,
         mm_processor_kwargs={"crop_to_patches": True},
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
     prompts = [
         f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|><image>{question}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
@@ -92,7 +92,7 @@ def run_blip2(questions: list[str], modality: str) -> ModelRequestData:
     prompts = [f"Question: {question} Answer:" for question in questions]
     engine_args = EngineArgs(
         model="Salesforce/blip2-opt-6.7b",
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -110,7 +110,7 @@ def run_chameleon(questions: list[str], modality: str) -> ModelRequestData:
         model="facebook/chameleon-7b",
         max_model_len=4096,
         max_num_seqs=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -129,8 +129,8 @@ def run_deepseek_vl2(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         max_model_len=4096,
         max_num_seqs=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
         hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]},
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [
@@ -155,7 +155,7 @@ def run_florence2(questions: list[str], modality: str) -> ModelRequestData:
         max_num_seqs=2,
         trust_remote_code=True,
         dtype="bfloat16",
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = ["<MORE_DETAILED_CAPTION>" for _ in questions]
@@ -175,7 +175,7 @@ def run_fuyu(questions: list[str], modality: str) -> ModelRequestData:
         model="adept/fuyu-8b",
         max_model_len=2048,
         max_num_seqs=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -194,7 +194,7 @@ def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
         max_model_len=2048,
         max_num_seqs=2,
         mm_processor_kwargs={"do_pan_and_scan": True},
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [("<bos><start_of_turn>user\n"
@@ -219,7 +219,7 @@ def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
         trust_remote_code=True,
         enforce_eager=True,
         hf_overrides={"architectures": ["GLM4VForCausalLM"]},
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [
@@ -246,7 +246,7 @@ def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         trust_remote_code=True,
         max_model_len=8192,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name,
@@ -287,7 +287,7 @@ def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
                 "longest_edge": 3 * 364
             },
         },
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
     prompts = [(
         f"<|begin_of_text|>User:<image>{question}<end_of_utterance>\nAssistant:"
@@ -314,7 +314,7 @@ def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
                 "longest_edge": 384
             },
         },
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
     prompts = [
         (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
@@ -337,7 +337,7 @@ def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         trust_remote_code=True,
         max_model_len=4096,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name,
@@ -375,7 +375,7 @@ def run_llava(questions: list[str], modality: str) -> ModelRequestData:
     engine_args = EngineArgs(
         model="llava-hf/llava-1.5-7b-hf",
         max_model_len=4096,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -392,7 +392,7 @@ def run_llava_next(questions: list[str], modality: str) -> ModelRequestData:
     engine_args = EngineArgs(
         model="llava-hf/llava-v1.6-mistral-7b-hf",
         max_model_len=8192,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -414,7 +414,7 @@ def run_llava_next_video(questions: list[str],
         model="llava-hf/LLaVA-NeXT-Video-7B-hf",
         max_model_len=8192,
         max_num_seqs=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -442,7 +442,7 @@ def run_llava_onevision(questions: list[str],
     engine_args = EngineArgs(
         model="llava-hf/llava-onevision-qwen2-7b-ov-hf",
         max_model_len=16384,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -465,7 +465,7 @@ def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
         model="TIGER-Lab/Mantis-8B-siglip-llama3",
         max_model_len=4096,
         hf_overrides={"architectures": ["MantisForConditionalGeneration"]},
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
     stop_token_ids = [128009]
 
@@ -506,7 +506,7 @@ def run_minicpmv_base(questions: list[str], modality: str, model_name):
         max_model_len=4096,
         max_num_seqs=2,
         trust_remote_code=True,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
     # NOTE The stop_token_ids are different for various versions of MiniCPM-V
     # 2.0
@@ -561,7 +561,7 @@ def run_mistral3(questions: list[str], modality: str) -> ModelRequestData:
         max_model_len=8192,
         max_num_seqs=2,
         tensor_parallel_size=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
@@ -587,7 +587,7 @@ def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         max_model_len=8192,
         max_num_seqs=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name)
@@ -611,7 +611,7 @@ def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
-def run_llama4(questions: list[str], modality: str):
+def run_llama4(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
 
     model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
@@ -621,8 +621,8 @@ def run_llama4(questions: list[str], modality: str):
         max_model_len=8192,
         max_num_seqs=4,
         tensor_parallel_size=8,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
         gpu_memory_utilization=0.4,
+        limit_mm_per_prompt={"image": 1},
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name)
@@ -657,7 +657,7 @@ def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         trust_remote_code=True,
         dtype="bfloat16",
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [
@@ -683,7 +683,7 @@ def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
         trust_remote_code=True,
         max_model_len=4096,
         tensor_parallel_size=4,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name,
@@ -710,7 +710,8 @@ def run_paligemma(questions: list[str], modality: str) -> ModelRequestData:
     prompts = ["caption en" for _ in questions]
     engine_args = EngineArgs(
         model="google/paligemma-3b-mix-224",
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache)
+        limit_mm_per_prompt={"image": 1},
+    )
 
     return ModelRequestData(
         engine_args=engine_args,
@@ -726,7 +727,8 @@ def run_paligemma2(questions: list[str], modality: str) -> ModelRequestData:
     prompts = ["caption en" for _ in questions]
     engine_args = EngineArgs(
         model="google/paligemma2-3b-ft-docci-448",
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache)
+        limit_mm_per_prompt={"image": 1},
+    )
 
     return ModelRequestData(
         engine_args=engine_args,
@@ -762,7 +764,7 @@ def run_phi3v(questions: list[str], modality: str) -> ModelRequestData:
         max_num_seqs=2,
         # Note - mm_processor_kwargs can also be passed to generate/chat calls
         mm_processor_kwargs={"num_crops": 16},
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -793,6 +795,7 @@ def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
         max_num_seqs=2,
         enable_lora=True,
         max_lora_rank=320,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(
@@ -813,7 +816,7 @@ def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         max_model_len=6144,
         max_num_seqs=2,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
@@ -834,7 +837,7 @@ def run_qwen_vl(questions: list[str], modality: str) -> ModelRequestData:
         max_model_len=1024,
         max_num_seqs=2,
         hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]},
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     prompts = [f"{question}Picture 1: <img></img>\n" for question in questions]
@@ -859,7 +862,7 @@ def run_qwen2_vl(questions: list[str], modality: str) -> ModelRequestData:
             "min_pixels": 28 * 28,
             "max_pixels": 1280 * 28 * 28,
         },
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     if modality == "image":
@@ -894,7 +897,7 @@ def run_qwen2_5_vl(questions: list[str], modality: str) -> ModelRequestData:
             "max_pixels": 1280 * 28 * 28,
             "fps": 1,
         },
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     if modality == "image":
@@ -925,7 +928,7 @@ def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
         model=model_name,
         trust_remote_code=True,
         max_model_len=4096,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        limit_mm_per_prompt={"image": 1},
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name,
@@ -1082,7 +1085,15 @@ def main(args):
 
     req_data = model_example_map[model](questions, modality)
 
-    engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
+    # Disable other modalities to save memory
+    default_limits = {"image": 0, "video": 0, "audio": 0}
+    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
+        req_data.engine_args.limit_mm_per_prompt or {})
+
+    engine_args = asdict(req_data.engine_args) | {
+        "seed": args.seed,
+        "disable_mm_preprocessor_cache": args.disable_mm_preprocessor_cache,
+    }
     llm = LLM(**engine_args)
 
     # To maintain code compatibility in this script, we add LoRA here.