[KV offload][3/N] Add worker-side CPU support (#21448)

Signed-off-by: Or Ozeri <oro@il.ibm.com>

tests/v1/kv_offload/test_cpu_gpu.py

@@ -0,0 +1,177 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import random
+import time
+
+import pytest
+import torch
+
+from vllm.platforms import current_platform
+from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
+from vllm.v1.attention.backends.flashinfer import FlashInferBackend
+from vllm.v1.attention.backends.mla.flashattn_mla import FlashAttnMLABackend
+from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
+from vllm.v1.kv_offload.worker.cpu_gpu import CpuGpuOffloadingHandler
+
+NUM_GPU_BLOCKS = [64]
+NUM_CPU_BLOCKS = [256]
+GPU_BLOCK_SIZES = [16]
+GPU_BLOCKS_PER_CPU_BLOCK = [1, 3]
+HEAD_SIZES = [64]
+NUM_HEADS = [8]
+NUM_LAYERS = [4]
+DTYPES = [torch.bfloat16]
+SEEDS = [0]
+CUDA_DEVICES = ['cuda:0']
+NUM_MAPPINGS = [3]
+
+
+@pytest.mark.parametrize("gpu_to_cpu", [True, False])
+@pytest.mark.parametrize("num_mappings", NUM_MAPPINGS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("gpu_block_size", GPU_BLOCK_SIZES)
+@pytest.mark.parametrize("gpu_blocks_per_cpu_block", GPU_BLOCKS_PER_CPU_BLOCK)
+@pytest.mark.parametrize("num_gpu_blocks", NUM_GPU_BLOCKS)
+@pytest.mark.parametrize("num_cpu_blocks", NUM_CPU_BLOCKS)
+@pytest.mark.parametrize("num_layers", NUM_LAYERS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@pytest.mark.parametrize("device", CUDA_DEVICES)
+@torch.inference_mode()
+def test_transfer(
+    gpu_to_cpu: bool,
+    num_mappings: int,
+    head_size: int,
+    num_heads: int,
+    gpu_block_size: int,
+    gpu_blocks_per_cpu_block: int,
+    num_gpu_blocks: int,
+    num_cpu_blocks: int,
+    num_layers: int,
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+) -> None:
+    current_platform.seed_everything(seed)
+
+    # create per-layer GPU KV caches
+    attn_backends_list = [
+        FlashAttentionBackend, FlashInferBackend, FlashAttnMLABackend
+    ]
+
+    gpu_caches = {}
+    attn_backends = {}
+    for i in range(num_layers):
+        layer_name = f'layer {i}'
+
+        attn_backend = attn_backends_list[i % len(attn_backends_list)]
+        attn_backends[layer_name] = attn_backend
+
+        gpu_cache_shape = attn_backend.get_kv_cache_shape(
+            num_gpu_blocks, gpu_block_size, num_heads, head_size)
+        gpu_caches[layer_name] = torch.rand(gpu_cache_shape,
+                                            dtype=dtype,
+                                            device=device)
+
+    # create handler
+    cpu_block_size = gpu_blocks_per_cpu_block * gpu_block_size
+    handler = CpuGpuOffloadingHandler(attn_backends=attn_backends,
+                                      gpu_block_size=gpu_block_size,
+                                      cpu_block_size=cpu_block_size,
+                                      num_cpu_blocks=num_cpu_blocks,
+                                      gpu_caches=gpu_caches)
+
+    # select block mappings
+    gpu_blocks = random.sample(range(num_gpu_blocks),
+                               num_mappings * gpu_blocks_per_cpu_block)
+    cpu_blocks = random.sample(range(num_cpu_blocks), num_mappings)
+
+    # convert cpu blocks to gpu block size
+    cpu_blocks_in_gpu_block_size = []
+    for cpu_block in cpu_blocks:
+        base_block_id = cpu_block * gpu_blocks_per_cpu_block
+        for i in range(gpu_blocks_per_cpu_block):
+            cpu_blocks_in_gpu_block_size.append(i + base_block_id)
+
+    # maybe skip a GPU block to test writing to the middle of a CPU block
+    if gpu_to_cpu:
+        gpu_blocks = gpu_blocks[gpu_blocks_per_cpu_block - 1:]
+        cpu_blocks_in_gpu_block_size = cpu_blocks_in_gpu_block_size[
+            gpu_blocks_per_cpu_block - 1:]
+
+    # set transfer direction
+    if gpu_to_cpu:
+        src_kv_caches = handler.gpu_tensors
+        dst_kv_caches = handler.cpu_tensors
+        src_spec_class = GPULoadStoreSpec
+        dst_spec_class = CPULoadStoreSpec
+        src_blocks = gpu_blocks
+        dst_blocks = cpu_blocks
+        src_blocks_in_gpu_block_size = gpu_blocks
+        dst_blocks_in_gpu_block_size = cpu_blocks_in_gpu_block_size
+        dst_size_in_gpu_blocks = num_cpu_blocks * gpu_blocks_per_cpu_block
+    else:
+        src_kv_caches = handler.cpu_tensors
+        dst_kv_caches = handler.gpu_tensors
+        src_spec_class = CPULoadStoreSpec
+        dst_spec_class = GPULoadStoreSpec
+        src_blocks = cpu_blocks
+        dst_blocks = gpu_blocks
+        src_blocks_in_gpu_block_size = cpu_blocks_in_gpu_block_size
+        dst_blocks_in_gpu_block_size = gpu_blocks
+        dst_size_in_gpu_blocks = num_gpu_blocks
+
+    # build dst -> src mapping
+    dst_to_src = {}
+    for src_block, dst_block in zip(src_blocks_in_gpu_block_size,
+                                    dst_blocks_in_gpu_block_size):
+        dst_to_src[dst_block] = src_block
+
+    # build transfer specs
+    src_spec = src_spec_class(src_blocks)
+    dst_spec = dst_spec_class(dst_blocks)
+
+    # clone src and dst tensors before transfer
+    orig_src_caches = [x.clone() for x in src_kv_caches]
+    orig_dst_caches = [x.clone() for x in dst_kv_caches]
+
+    # call transfer function
+    assert handler.transfer_async(1, (src_spec, dst_spec))
+    assert set(handler.transfer_events.keys()) == {1}
+
+    # wait for transfer to complete
+    end_time = time.time() + 10
+    while time.time() < end_time:
+        finished = handler.get_finished()
+        if finished:
+            assert finished == [(1, True)]
+            break
+        time.sleep(0.1)
+
+    # verify src tensors did not change
+    for orig_tensor, tensor in zip(orig_src_caches, src_kv_caches):
+        assert torch.equal(orig_tensor, tensor)
+
+    # verify dst tensors
+    for dst_block in range(dst_size_in_gpu_blocks):
+        src_block_candidate = dst_to_src.get(dst_block)
+        for src_cache, dst_cache, orig_dst_cache, kv_dim in zip(
+                src_kv_caches, dst_kv_caches, orig_dst_caches,
+                handler.kv_dim_before_num_blocks):
+            if kv_dim:
+                # iterate over key, value
+                for i in range(2):
+                    if src_block_candidate is not None:
+                        expected_value = src_cache[i][src_block_candidate]
+                    else:
+                        expected_value = orig_dst_cache[i][dst_block]
+                    torch.testing.assert_close(dst_cache[i][dst_block].cpu(),
+                                               expected_value.cpu())
+            else:
+                if src_block_candidate is not None:
+                    expected_value = src_cache[src_block_candidate]
+                else:
+                    expected_value = orig_dst_cache[dst_block]
+                torch.testing.assert_close(dst_cache[dst_block].cpu(),
+                                           expected_value.cpu())

vllm/v1/kv_offload/worker/cpu_gpu.py

@@ -0,0 +1,171 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import numpy as np
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.attention import AttentionBackend
+from vllm.logger import init_logger
+from vllm.utils import is_pin_memory_available
+from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
+from vllm.v1.kv_offload.worker.worker import (OffloadingHandler,
+                                              TransferResult, TransferSpec)
+
+logger = init_logger(__name__)
+
+
+def expand_block_ids(block_ids: np.ndarray,
+                     block_size_factor: int,
+                     output: np.ndarray,
+                     skip_count: int = 0):
+    """
+    Convert a list of block IDs to a list of matching block ids,
+    assuming each block is composed of actual block_size_factor blocks.
+    Outputs to output tensor.
+    The first skip_count blocks will be skipped.
+    Note that skip_count must be less than block_size_factor.
+
+    For example, if block_ids = [0, 1, 3] and block_size_factor =  4,
+    then it yields [0, 1, 2, 3, 4, 5, 6, 7, 12, 13, 14, 15]
+    since 0 maps to [0, 1, 2, 3]
+    1 maps to [4, 5, 6, 7]
+    and 3 maps to [12, 13, 14, 15]
+    """
+    assert skip_count < block_size_factor
+
+    first_range = np.arange(skip_count, block_size_factor)
+    full_range = np.arange(0, block_size_factor)
+
+    output_idx = 0
+    for i, block_id in enumerate(block_ids):
+        base_block_id = block_id * block_size_factor
+        indices = first_range if i == 0 else full_range
+        output_end_idx = output_idx + len(indices)
+        output[output_idx:output_end_idx] = base_block_id + indices
+        output_idx = output_end_idx
+
+
+class CpuGpuOffloadingHandler(OffloadingHandler):
+
+    def __init__(self, gpu_block_size: int, cpu_block_size: int,
+                 num_cpu_blocks: int, gpu_caches: dict[str, torch.Tensor],
+                 attn_backends: dict[str, type[AttentionBackend]]):
+        assert cpu_block_size % gpu_block_size == 0
+        self.block_size_factor = cpu_block_size // gpu_block_size
+
+        # cuda streams for gpu->cpu and cpu->gpu
+        self.d2h_stream = torch.cuda.Stream()
+        self.h2d_stream = torch.cuda.Stream()
+
+        # job_id -> transfer cuda event
+        self.transfer_events: dict[int, torch.cuda.Event] = {}
+        # list of cuda events available for re-use
+        self.events_pool: list[torch.cuda.Event] = []
+
+        pin_memory = is_pin_memory_available()
+
+        # allocate cpu tensors
+        logger.info("Allocating %d CPU tensors...", len(gpu_caches))
+        self.gpu_tensors: list[torch.Tensor] = []
+        self.cpu_tensors: list[torch.Tensor] = []
+        self.kv_dim_before_num_blocks: list[bool] = []
+        for layer_name, gpu_tensor in gpu_caches.items():
+            self.gpu_tensors.append(gpu_tensor)
+
+            gpu_shape = gpu_tensor.shape
+            test_shape = attn_backends[layer_name].get_kv_cache_shape(
+                num_blocks=1234, block_size=16, num_kv_heads=8, head_size=256)
+            if test_shape[0] == 1234:
+                # shape is (num_blocks, ...)
+                num_blocks_idx = 0
+                self.kv_dim_before_num_blocks.append(False)
+            else:
+                # shape should be (2, num_blocks, ...)
+                assert test_shape[0] == 2
+                assert test_shape[1] == 1234
+                assert gpu_shape[0] == 2
+
+                num_blocks_idx = 1
+                self.kv_dim_before_num_blocks.append(True)
+
+            cpu_shape = list(gpu_shape)
+            cpu_shape[num_blocks_idx] = num_cpu_blocks * self.block_size_factor
+
+            logger.debug("Allocating CPU tensor of shape %r", cpu_shape)
+            self.cpu_tensors.append(
+                torch.zeros(cpu_shape,
+                            dtype=gpu_tensor.dtype,
+                            device="cpu",
+                            pin_memory=pin_memory))
+
+    def transfer_async(self, job_id: int, spec: TransferSpec) -> bool:
+        src_spec, dst_spec = spec
+        if isinstance(src_spec, CPULoadStoreSpec):
+            assert isinstance(dst_spec, GPULoadStoreSpec)
+            stream = self.h2d_stream
+            src_tensors = self.cpu_tensors
+            dst_tensors = self.gpu_tensors
+            src_block_size_factor = self.block_size_factor
+            dst_block_size_factor = 1
+        else:
+            assert isinstance(src_spec, GPULoadStoreSpec)
+            assert isinstance(dst_spec, CPULoadStoreSpec)
+            stream = self.d2h_stream
+            src_tensors = self.gpu_tensors
+            dst_tensors = self.cpu_tensors
+            src_block_size_factor = 1
+            dst_block_size_factor = self.block_size_factor
+
+        src_blocks = src_spec.block_ids
+        dst_blocks = dst_spec.block_ids
+        assert src_blocks.ndim == 1
+        assert dst_blocks.ndim == 1
+
+        dst_sub_blocks_to_skip = (-src_blocks.size % dst_block_size_factor)
+        src_sub_block_count = src_blocks.size * src_block_size_factor
+
+        assert (
+            src_sub_block_count == dst_blocks.size * dst_block_size_factor -
+            dst_sub_blocks_to_skip)
+
+        src_to_dst = np.empty((src_sub_block_count, 2), dtype=np.int64)
+        expand_block_ids(src_blocks, src_block_size_factor, src_to_dst[:, 0])
+        expand_block_ids(dst_blocks,
+                         dst_block_size_factor,
+                         src_to_dst[:, 1],
+                         skip_count=dst_sub_blocks_to_skip)
+        src_to_dst_tensor = torch.from_numpy(src_to_dst)
+
+        event = self.events_pool.pop() if self.events_pool \
+            else torch.cuda.Event()
+        with torch.cuda.stream(stream):
+            for src_tensor, dst_tensor, kv_dim in zip(
+                    src_tensors, dst_tensors, self.kv_dim_before_num_blocks):
+                if kv_dim:
+                    src_key_cache = src_tensor[0]
+                    dst_key_cache = dst_tensor[0]
+                    ops.swap_blocks(src_key_cache, dst_key_cache,
+                                    src_to_dst_tensor)
+                    src_value_cache = src_tensor[1]
+                    dst_value_cache = dst_tensor[1]
+                    ops.swap_blocks(src_value_cache, dst_value_cache,
+                                    src_to_dst_tensor)
+                else:
+                    ops.swap_blocks(src_tensor, dst_tensor, src_to_dst_tensor)
+            event.record(stream)
+
+        self.transfer_events[job_id] = event
+
+        # success
+        return True
+
+    def get_finished(self) -> list[TransferResult]:
+        results: list[TransferResult] = []
+        for job_id, event in self.transfer_events.items():
+            if event.query():
+                results.append((job_id, True))
+                self.events_pool.append(event)
+        for job_id, _ in results:
+            del self.transfer_events[job_id]
+        return results