[WIP] [V1] VLM hashing and mapper caching

examples/offline_inference_vision_language.py

@@ -5,6 +5,7 @@
 For most models, the prompt format should follow corresponding examples
 on HuggingFace model repository.
 """
+import random
 from transformers import AutoTokenizer
 
 from vllm import LLM, SamplingParams
@@ -23,7 +24,9 @@
 
     prompt = f"USER: <image>\n{question}\nASSISTANT:"
 
-    llm = LLM(model="llava-hf/llava-1.5-7b-hf", max_model_len=4096)
+    llm = LLM(model="llava-hf/llava-1.5-7b-hf",
+              max_model_len=4096,
+              mm_cache_preprocessor=args.mm_cache_preprocessor)
     stop_token_ids = None
     return llm, prompt, stop_token_ids
 
@@ -507,14 +510,35 @@
 
     else:
         # Batch inference
-        inputs = [{
-            "prompt": prompt,
-            "multi_modal_data": {
-                modality: data
-            },
-        } for _ in range(args.num_prompts)]
-
+        if args.image_repeat_ratio is not None:
+            assert (args.image_repeat_ratio <= 1.0
+                    and args.image_repeat_ratio >= 0)
+            no_yes = [0, 1]
+            probs = [1.0 - args.image_repeat_ratio, args.image_repeat_ratio]
+
+        inputs = []
+        cur_image = data
+        for i in range(args.num_prompts):
+            if args.image_repeat_ratio is not None:
+                res = random.choices(no_yes, probs)[0]
+                if res == 0:
+                    # No repeat => Modify one pixel
+                    cur_image = cur_image.copy()
+                    new_val = (i // 256 // 256, i // 256, i % 256)
+                    cur_image.putpixel((0, 0), new_val)
+
+            inputs.append({
+                "prompt": prompt,
+                "multi_modal_data": {
+                    modality: cur_image
+                }
+            })
+
+    import time
+    start_time = time.time()
     outputs = llm.generate(inputs, sampling_params=sampling_params)
+    elapsed_time = time.time() - start_time
+    print("-- generate time = {}".format(elapsed_time))
 
     for o in outputs:
         generated_text = o.outputs[0].text
@@ -544,5 +568,19 @@
                         type=int,
                         default=16,
                         help='Number of frames to extract from the video.')
+
+    parser.add_argument(
+        '--image-repeat-ratio',
+        type=float,
+        default=None,
+        help=
+        'Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)'
+    )
+
+    parser.add_argument(
+        '--mm-cache-preprocessor',
+        action='store_true',
+        help='If True, enable caching of multi-modal preprocessor/mapper.')
+
     args = parser.parse_args()
     main(args)

vllm/config.py

@@ -131,6 +131,8 @@ class ModelConfig:
             HuggingFace config.
         mm_processor_kwargs: Arguments to be forwarded to the model's processor
             for multi-modal data, e.g., image processor.
+        mm_cache_preprocessor: If True, enable caching of multi-modal 
+            preprocessor/mapper.
         override_neuron_config: Initialize non default neuron config or
             override default neuron config that are specific to Neuron devices,
             this argument will be used to configure the neuron config that
@@ -169,6 +171,7 @@ def __init__(
             config_format: ConfigFormat = ConfigFormat.AUTO,
             hf_overrides: Optional[HfOverrides] = None,
             mm_processor_kwargs: Optional[Dict[str, Any]] = None,
+            mm_cache_preprocessor: bool = False,
             override_neuron_config: Optional[Dict[str, Any]] = None,
             override_pooler_config: Optional["PoolerConfig"] = None) -> None:
         self.model = model
@@ -235,6 +238,7 @@ def __init__(
         self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
         self.use_async_output_proc = use_async_output_proc
         self.mm_processor_kwargs = mm_processor_kwargs
+        self.mm_cache_preprocessor = mm_cache_preprocessor
 
         # Set enforce_eager to False if the value is unset.
         if self.enforce_eager is None:
@@ -2593,7 +2597,8 @@ def __str__(self):
         "decoding_config=%r, observability_config=%r, "
         "seed=%d, served_model_name=%s, "
         "num_scheduler_steps=%d, enable_prefix_caching=%s, "
-        "use_async_output_proc=%s, mm_processor_kwargs=%s") % \
+        "use_async_output_proc=%s, mm_processor_kwargs=%s, "
+        "mm_cache_preprocessor=%s") % \
         (self.model_config.model, self.speculative_config,
         self.model_config.tokenizer,
         self.model_config.skip_tokenizer_init,
@@ -2619,7 +2624,8 @@ def __str__(self):
         self.scheduler_config.num_scheduler_steps,
         self.cache_config.enable_prefix_caching,
         self.model_config.use_async_output_proc,
-        self.model_config.mm_processor_kwargs)
+        self.model_config.mm_processor_kwargs,
+        self.model_config.mm_cache_preprocessor)
 
 
 _current_vllm_config: Optional[VllmConfig] = None

vllm/engine/arg_utils.py

@@ -143,6 +143,7 @@ class EngineArgs:
     tokenizer_pool_extra_config: Optional[Dict[str, Any]] = None
     limit_mm_per_prompt: Optional[Mapping[str, int]] = None
     mm_processor_kwargs: Optional[Dict[str, Any]] = None
+    mm_cache_preprocessor: bool = False
     enable_lora: bool = False
     enable_lora_bias: bool = False
     max_loras: int = 1
@@ -592,6 +593,10 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             type=json.loads,
             help=('Overrides for the multimodal input mapping/processing, '
                   'e.g., image processor. For example: {"num_crops": 4}.'))
+        parser.add_argument(
+            '--mm-cache-preprocessor',
+            action='store_true',
+            help='If True, enable caching of multi-modal preprocessor/mapper.')
 
         # LoRA related configs
         parser.add_argument('--enable-lora',
@@ -964,6 +969,7 @@ def create_model_config(self) -> ModelConfig:
             use_async_output_proc=not self.disable_async_output_proc,
             config_format=self.config_format,
             mm_processor_kwargs=self.mm_processor_kwargs,
+            mm_cache_preprocessor=self.mm_cache_preprocessor,
             override_neuron_config=self.override_neuron_config,
             override_pooler_config=self.override_pooler_config,
         )

vllm/v1/core/scheduler.py

@@ -510,6 +510,7 @@ class NewRequestData:
     prompt_token_ids: List[int]
     prompt: Optional[str]
     mm_inputs: List["MultiModalKwargs"]
+    mm_hash: List[str]
     mm_positions: List["PlaceholderRange"]
     sampling_params: SamplingParams
     block_ids: List[int]
@@ -527,6 +528,7 @@ def from_request(
             prompt_token_ids=request.prompt_token_ids,
             prompt=request.prompt,
             mm_inputs=request.mm_inputs,
+            mm_hash=request.mm_hash,
             mm_positions=request.mm_positions,
             sampling_params=request.sampling_params,
             block_ids=block_ids,

vllm/v1/engine/__init__.py

@@ -36,6 +36,7 @@ class EngineCoreRequest:
     prompt: Optional[str]
     prompt_token_ids: List[int]
     mm_inputs: Optional[List[MultiModalKwargs]]
+    mm_hash: Optional[List[str]]
     mm_placeholders: Optional[MultiModalPlaceholderDict]
     sampling_params: SamplingParams
     eos_token_id: Optional[int]

vllm/v1/engine/core.py

@@ -19,7 +19,6 @@
 from vllm.v1.engine import (EngineCoreOutput, EngineCoreOutputs,
                             EngineCoreProfile, EngineCoreRequest,
                             EngineCoreRequestType)
-from vllm.v1.engine.mm_input_mapper import MMInputMapper
 from vllm.v1.executor.gpu_executor import GPUExecutor
 from vllm.v1.request import Request, RequestStatus
 from vllm.v1.serial_utils import PickleEncoder
@@ -55,9 +54,6 @@ def __init__(
         vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
         vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks
 
-        # Set up multimodal input mapper (e.g., convert PIL images to tensors).
-        self.mm_input_mapper = MMInputMapper(vllm_config.model_config)
-
         # Setup scheduler.
         self.scheduler = Scheduler(vllm_config.scheduler_config,
                                    vllm_config.cache_config,

vllm/v1/engine/mm_input_mapper.py

@@ -1,8 +1,13 @@
+import time
+import PIL
+
+from blake3 import blake3
 from typing import Any, Dict, List, Optional
 
 from vllm.config import ModelConfig
 from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalDataDict,
                              MultiModalKwargs, MultiModalRegistry)
+from vllm.v1.utils import LRUDictCache
 
 
 class MMInputMapper:
@@ -11,29 +16,97 @@
         self,
         model_config: ModelConfig,
         mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
+        mm_cache_size: int = 128,
     ):
         self.mm_registry = mm_registry
         self.multi_modal_input_mapper = mm_registry.create_input_mapper(
             model_config)
         self.mm_registry.init_mm_limits_per_prompt(model_config)
 
+        self.mm_cache = LRUDictCache(size=mm_cache_size)
+        self.mm_cache_hits = 0
+        self.mm_cache_misses = 0
+
+        # Set to None to disable (TODO: Disable!)
+        self.mm_debug_cache_hit_ratio_steps = 32
+
+    def cache_hit_ratio(self, steps) -> float:
+        total_steps = self.mm_cache_hits + self.mm_cache_misses
+
+        if total_steps > 0 and total_steps % steps == 0:
+            print("[debug] MMInputMapper: cache_hit_ratio = {}".format(
+                self.mm_cache_hits / total_steps))
+
     def process_inputs(
         self,
         mm_data: MultiModalDataDict,
+        mm_hash: Optional[List[str]],
         mm_processor_kwargs: Optional[Dict[str, Any]],
     ) -> List[MultiModalKwargs]:
         image_inputs = mm_data["image"]
         if not isinstance(image_inputs, list):
             image_inputs = [image_inputs]
 
+        use_hash = False #mm_hash is not None
+        if use_hash:
+            assert len(image_inputs) == len(mm_hash)  # Sanity
+
         # Process each image input separately so that later we can schedule
         # them in a fine-grained manner.
+        # Utilize caching (if enabled)
         mm_inputs: List[MultiModalKwargs] = []
-        num_images = len(image_inputs)
-        for i in range(num_images):
-            mm_input = self.multi_modal_input_mapper(
-                {"image": [image_inputs[i]]},
-                mm_processor_kwargs=mm_processor_kwargs,
-            )
+        for i in range(len(image_inputs)):
+            if self.mm_debug_cache_hit_ratio_steps is not None:
+                self.cache_hit_ratio(self.mm_debug_cache_hit_ratio_steps)
+
+            mm_input = self.mm_cache.get(mm_hash[i]) if use_hash else None
+            if mm_input is None:
+                self.mm_cache_misses += 1
+                mm_input = self.multi_modal_input_mapper(
+                    {"image": [image_inputs[i]]},
+                    mm_processor_kwargs=mm_processor_kwargs,
+                )
+                if use_hash:
+                    self.mm_cache.put(mm_hash[i], mm_input)
+            else:
+                self.mm_cache_hits += 1
+
             mm_inputs.append(mm_input)
+
         return mm_inputs
+
+
+class MMHasher:
+
+    def __init__(self):
+        pass
+
+    def hash(self, mm_data: MultiModalDataDict) -> List[str]:
+        image_inputs = mm_data["image"]
+        if not isinstance(image_inputs, list):
+            image_inputs = [image_inputs]
+
+        ret = []
+        for image in image_inputs:
+            assert isinstance(image, PIL.Image.Image)
+
+            # FIXME(alexm): Remove debug
+
+            # print("  type(data) = {}, data = {}".format(type(image), image))
+
+            # Convert image to bytes
+            start_time = time.time()
+            bytes = image.tobytes()
+            elapsed_time = time.time() - start_time
+            # print("    tobytes time = {}".format(elapsed_time))
+
+            # Hash image bytes
+            start_time = time.time()
+            hasher = blake3()
+            hasher.update(bytes)
+            ret.append(hasher.hexdigest())
+            elapsed_time = time.time() - start_time
+            # print("    hash time = {}".format(elapsed_time))
+            # print("    hash val = {}".format(ret[-1]))
+
+        return ret

vllm/v1/engine/processor.py

@@ -14,7 +14,7 @@
 from vllm.transformers_utils.config import try_get_generation_config
 from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
 from vllm.v1.engine import DetokenizerRequest, EngineCoreRequest
-from vllm.v1.engine.mm_input_mapper import MMInputMapper
+from vllm.v1.engine.mm_input_mapper import MMInputMapper, MMHasher
 
 
 class Processor:
@@ -43,6 +43,10 @@ def __init__(
         # Multi-modal (huggingface) input mapper
         self.mm_input_mapper = MMInputMapper(model_config)
 
+        # Multi-modal hasher (for images)
+        self.mm_hasher = MMHasher(
+        ) if model_config.mm_cache_preprocessor else None
+
     # TODO: run in an ThreadpoolExecutor or BackgroundProcess.
     # This ideally should releases the GIL, so we should not block the
     # asyncio loop while this is running.
@@ -101,8 +105,10 @@ def process_inputs(
             self.generation_config_fields, eos_token_id)
 
         # Preprocess multi-modal data
+        mm_hash = self.mm_hasher.hash(decoder_inputs.multi_modal_data
+                                      ) if self.mm_hasher is not None else None
         mm_inputs = self.mm_input_mapper.process_inputs(
-            decoder_inputs.multi_modal_data,
+            decoder_inputs.multi_modal_data, mm_hash,
             decoder_inputs.mm_processor_kwargs) if len(
                 decoder_inputs.multi_modal_data) > 0 else None
 
@@ -124,6 +130,7 @@ def process_inputs(
             decoder_inputs.prompt,
             decoder_inputs.prompt_token_ids,
             mm_inputs,
+            mm_hash,
             decoder_inputs.multi_modal_placeholders,
             sampling_params,
             eos_token_id,

vllm/v1/request.py

@@ -20,6 +20,7 @@ def __init__(
         eos_token_id: Optional[int],
         arrival_time: float,
         lora_request: Optional[LoRARequest] = None,
+        mm_hash: Optional[List[str]] = None,
     ) -> None:
         self.request_id = request_id
         self.inputs = SingletonInputsAdapter(inputs)
@@ -56,6 +57,7 @@ def __init__(
             self.mm_inputs = self.inputs.multi_modal_inputs
         else:
             self.mm_inputs: List[MultiModalKwargs] = []
+        self.mm_hash = mm_hash
 
     @classmethod
     def from_engine_core_request(cls, request: EngineCoreRequest) -> "Request":
@@ -73,6 +75,7 @@ def from_engine_core_request(cls, request: EngineCoreRequest) -> "Request":
             eos_token_id=request.eos_token_id,
             arrival_time=request.arrival_time,
             lora_request=request.lora_request,
+            mm_hash=request.mm_hash,
         )
 
     @property