Add memory analyzer & utomatically configure KV cache size (#6)

README.md

@@ -3,7 +3,7 @@
 ## Installation
 
 ```bash
-pip install cmake torch transformers
+pip install psutil numpy torch transformers
 pip install flash-attn # This may take up to 10 mins.
 pip install -e .
 ```

cacheflow/master/scheduler.py

@@ -9,8 +9,6 @@ from cacheflow.sequence import SequenceGroupInputs
 from cacheflow.sequence import SequenceOutputs
 from cacheflow.sequence import SequenceStatus
 
-_MAX_NUM_BATCHED_TOKENS = 2048
-
 
 class Scheduler:
 
@@ -21,12 +19,14 @@ class Scheduler:
         block_size: int,
         num_gpu_blocks: int,
         num_cpu_blocks: int,
+        max_num_batched_tokens: int,
     ) -> None:
         self.frontend = frontend
         self.controllers = controllers
         self.block_size = block_size
         self.num_gpu_blocks = num_gpu_blocks
         self.num_cpu_blocks = num_cpu_blocks
+        self.max_num_batched_tokens = max_num_batched_tokens
 
         # Create the block space manager.
         self.block_manager = BlockSpaceManager(
@@ -164,7 +164,7 @@ class Scheduler:
                 num_prompt_tokens = seq_group.seqs[0].get_len()
                 if self.block_manager.can_allocate(seq_group):
                     if (num_batched_tokens + num_prompt_tokens
-                        <= _MAX_NUM_BATCHED_TOKENS):
+                        <= self.max_num_batched_tokens):
                         self._allocate(seq_group)
                         num_batched_tokens += num_prompt_tokens
                         continue

cacheflow/models/__init__.py

@@ -1,10 +1,12 @@
 from cacheflow.models.input_metadata import InputMetadata
+from cacheflow.models.model_utils import get_memory_analyzer
 from cacheflow.models.model_utils import get_model
-from cacheflow.models.model_utils import set_seed
+from cacheflow.models.utils import set_seed
 
 
 __all__ = [
     'InputMetadata',
+    'get_memory_analyzer',
     'get_model',
-    'set_seed'
+    'set_seed',
 ]

cacheflow/models/memory_analyzer.py

@@ -0,0 +1,125 @@
+import torch
+from transformers import AutoConfig
+
+from cacheflow.models.utils import get_cpu_memory
+from cacheflow.models.utils import get_dtype_size
+from cacheflow.models.utils import get_gpu_memory
+
+_GiB = 1 << 30    
+
+
+class CacheFlowMemoryAnalyzer:
+
+    def get_max_num_gpu_blocks(
+        self,
+        max_num_batched_tokens: int,
+        memory_utilization: float,
+    ) -> int:
+        raise NotImplementedError()
+
+    def get_max_num_cpu_blocks(
+        self,
+        memory_utilization: float,
+    ) -> int:
+        raise NotImplementedError()
+
+
+class OPTMemoryAnalyzer(CacheFlowMemoryAnalyzer):
+
+    def __init__(
+        self,
+        model_name: str,
+        block_size: int,
+        dtype: torch.dtype,
+    ) -> None:
+        self.model_name = model_name
+        self.block_size = block_size
+        self.dtype = dtype
+
+        # TODO(woosuk): Support tensor parallelism.
+        config = AutoConfig.from_pretrained(model_name)
+        self.num_layers = config.num_hidden_layers
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_size = config.hidden_size // self.num_heads
+        self.ffn_size = config.ffn_dim
+        self.embedding_size = config.word_embed_proj_dim
+        self.vocab_size = config.vocab_size
+        self.max_position = config.max_position_embeddings
+
+    def _get_param_size(self) -> int:
+        # TODO(woosuk): Support tensor parallelism.
+        word_embedding = self.vocab_size * self.embedding_size
+        if self.embedding_size != self.vocab_size:
+            # Project in/out.
+            word_embedding += 2 * self.embedding_size * self.vocab_size
+        position_embedding = self.max_position * self.hidden_size
+
+        ln1 = 2 * self.hidden_size
+        q = self.hidden_size * self.hidden_size + self.hidden_size
+        k = self.hidden_size * self.hidden_size + self.hidden_size
+        v = self.hidden_size * self.hidden_size + self.hidden_size
+        out = self.hidden_size * self.hidden_size + self.hidden_size
+        mha = ln1 + q + k + v + out
+
+        ln2 = 2 * self.hidden_size
+        ffn1 = self.hidden_size * self.ffn_size + self.ffn_size
+        ffn2 = self.ffn_size * self.hidden_size + self.hidden_size
+        ffn = ln2 + ffn1 + ffn2
+
+        total = (word_embedding + position_embedding + 
+                 self.num_layers * (mha + ffn))
+        dtype_size = get_dtype_size(self.dtype)
+        return dtype_size * total
+
+    def _get_max_act_size(
+        self,
+        max_num_batched_tokens: int,
+    ) -> int:
+        # TODO(woosuk): Support tensor parallelism.
+        # NOTE: We approxmiately calculate the maximum activation size by
+        # 1) estimating the maximum activation tensor size during inference, and
+        # 2) multiplying it by 4.
+        # Here, we assume that FlashAttention is used and
+        # thus the attention maps are never materialized in GPU DRAM.
+        qkv = 3 * (max_num_batched_tokens * self.hidden_size)
+        ffn = max_num_batched_tokens * self.ffn_size
+        max_act = 4 * max(qkv, ffn)
+        dtype_size = get_dtype_size(self.dtype)
+        return dtype_size * max_act
+
+    def _get_workspace_size(self) -> int:
+        return 1 * _GiB
+
+    def _get_cache_block_size(self) -> int:
+        key_cache_block = self.block_size * self.num_heads * self.head_size
+        value_cache_block = self.block_size * self.num_heads * self.head_size
+        total = self.num_layers * (key_cache_block + value_cache_block)
+        dtype_size = get_dtype_size(self.dtype)
+        return dtype_size * total
+
+    def get_max_num_gpu_blocks(
+        self,
+        max_num_batched_tokens: int,
+        memory_utilization: float = 0.95,
+    ) -> int:
+        # NOTE(woosuk): This assumes that the machine has homogeneous GPUs.
+        gpu_memory = get_gpu_memory()
+        usable_memory = int(memory_utilization * gpu_memory)
+
+        param_size = self._get_param_size()
+        act_size = self._get_max_act_size(max_num_batched_tokens)
+        workspace_size = self._get_workspace_size()
+
+        max_cache_size = usable_memory - (param_size + act_size + workspace_size)
+        max_num_blocks = max_cache_size // self._get_cache_block_size()
+        return max_num_blocks
+
+    def get_max_num_cpu_blocks(
+        self,
+        memory_utilization: float = 0.25,
+    ) -> int:
+        cpu_memory = get_cpu_memory()
+        usable_memory = int(memory_utilization * cpu_memory)
+        max_num_blocks = usable_memory // self._get_cache_block_size()
+        return max_num_blocks

cacheflow/models/model_utils.py

@@ -1,21 +1,20 @@
-import random
 from typing import Union
 
-import numpy as np
 import torch
 import torch.nn as nn
 
+from cacheflow.models.memory_analyzer import CacheFlowMemoryAnalyzer
+from cacheflow.models.memory_analyzer import OPTMemoryAnalyzer
 from cacheflow.models.opt import OPTForCausalLM
+from cacheflow.models.utils import get_torch_dtype
 
-MODEL_CLASSES = {
+
+_MODELS = {
     'opt': OPTForCausalLM,
 }
 
-STR_DTYPE_TO_TORCH_DTYPE = {
-    'half': torch.half,
-    'float': torch.float,
-    'float16': torch.float16,
-    'float32': torch.float32,
+_MEMORY_ANALYZERS = {
+    'opt': OPTMemoryAnalyzer,
 }
 
 
@@ -23,20 +22,23 @@ def get_model(
     model_name: str,
     dtype: Union[torch.dtype, str],
 ) -> nn.Module:
-    if isinstance(dtype, str):
-        torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[dtype.lower()]
-    else:
-        torch_dtype = dtype
-    for model_class, hf_model in MODEL_CLASSES.items():
+    torch_dtype = get_torch_dtype(dtype)
+    for model_class, hf_model in _MODELS.items():
         if model_class in model_name:
-            model = hf_model.from_pretrained(model_name, torch_dtype=torch_dtype)
+            model = hf_model.from_pretrained(
+                model_name, torch_dtype=torch_dtype)
             return model.eval()
-    raise ValueError(f'Invalid model name: {model_name}')
+    raise ValueError(f'Unsupported model name: {model_name}')
 
 
-def set_seed(seed: int) -> None:
-    random.seed(seed)
-    np.random.seed(seed)
-    torch.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(seed)
+def get_memory_analyzer(
+    model_name: str,
+    block_size: int,
+    dtype: Union[torch.dtype, str],
+) -> CacheFlowMemoryAnalyzer:
+    torch_dtype = get_torch_dtype(dtype)
+    for model_class, memory_analyzer in _MEMORY_ANALYZERS.items():
+        if model_class in model_name:
+            return memory_analyzer(
+                model_name, block_size, torch_dtype)
+    raise ValueError(f'Unsupported model name: {model_name}')

cacheflow/models/utils.py

@@ -0,0 +1,43 @@
+from typing import Union
+
+import random
+
+import numpy as np
+import psutil
+import torch
+
+_STR_DTYPE_TO_TORCH_DTYPE = {
+    'half': torch.half,
+    'float': torch.float,
+    'float16': torch.float16,
+    'float32': torch.float32,
+}
+
+
+def get_torch_dtype(dtype: Union[torch.dtype, str]) -> torch.dtype:
+    if isinstance(dtype, str):
+        torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype.lower()]
+    else:
+        torch_dtype = dtype
+    return torch_dtype
+
+
+def get_dtype_size(dtype: Union[torch.dtype, str]) -> int:
+    torch_dtype = get_torch_dtype(dtype)
+    return torch.tensor([], dtype=torch_dtype).element_size()
+
+
+def set_seed(seed: int) -> None:
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+def get_gpu_memory(gpu: int = 0) -> int:
+    return torch.cuda.get_device_properties(gpu).total_memory
+
+
+def get_cpu_memory() -> int:
+    return psutil.virtual_memory().total

server.py

@@ -3,6 +3,7 @@ from typing import List
 
 from cacheflow.master.frontend import Frontend
 from cacheflow.master.scheduler import Scheduler
+from cacheflow.models import get_memory_analyzer
 from cacheflow.worker.controller import Controller
 
 parser = argparse.ArgumentParser(description='CacheFlow server')
@@ -10,17 +11,25 @@ parser.add_argument('--model', type=str, default='facebook/opt-125m', help='mode
 parser.add_argument('--num-nodes', type=int, default=1, help='number of nodes')
 parser.add_argument('--num-workers', type=int, default=1, help='number of workers per node')
 parser.add_argument('--block-size', type=int, default=8, choices=[8, 16], help='token block size')
-# TODO(woosuk): Add an analytical model to determine the maximum number of GPU/CPU blocks.
-parser.add_argument('--num-gpu-blocks', type=int, default=1024, help='number of GPU blocks (per GPU)')
-parser.add_argument('--num-cpu-blocks', type=int, default=32, help='number of CPU blocks (per GPU)')
 # NOTE(woosuk): If FlashAttention is used, the float data type is not supported.
 parser.add_argument('--dtype', type=str, default='half', choices=['half', 'float'], help='data type')
 # TODO(woosuk): Support fine-grained seeds (e.g., seed per request).
 parser.add_argument('--seed', type=int, default=0, help='random seed')
+parser.add_argument('--max-batch-size', type=int, default=2048, help='maximum number of batched tokens')
 args = parser.parse_args()
 
 
 def main():
+    memory_analyzer = get_memory_analyzer(
+        model_name=args.model,
+        block_size=args.block_size,
+        dtype=args.dtype,
+    )
+    num_gpu_blocks = memory_analyzer.get_max_num_gpu_blocks(
+        max_num_batched_tokens=args.max_batch_size)
+    num_cpu_blocks = memory_analyzer.get_max_num_cpu_blocks()
+    print(f'# GPU blocks: {num_gpu_blocks}, # CPU blocks: {num_cpu_blocks}')
+
     # Create a controller for each node.
     controllers: List[Controller] = []
     for i in range(args.num_nodes):
@@ -29,8 +38,8 @@ def main():
             num_workers=args.num_workers,
             model_name=args.model,
             block_size=args.block_size,
-            num_gpu_blocks=args.num_gpu_blocks,
-            num_cpu_blocks=args.num_cpu_blocks,
+            num_gpu_blocks=num_gpu_blocks,
+            num_cpu_blocks=num_cpu_blocks,
             dtype=args.dtype,
             seed=args.seed,
         )
@@ -47,8 +56,9 @@ def main():
         frontend=frontend,
         controllers=controllers,
         block_size=args.block_size,
-        num_gpu_blocks=args.num_gpu_blocks,
-        num_cpu_blocks=args.num_cpu_blocks,
+        num_gpu_blocks=num_gpu_blocks,
+        num_cpu_blocks=num_cpu_blocks,
+        max_num_batched_tokens=args.max_batch_size,
     )
     # Connect the controllers.
     for i in range(len(controllers) - 1):