Skip to content

vllm.distributed.kv_transfer.kv_connector.v1.hf3fs.utils.gather_scatter_helper

Classes:

  • CopyBufferAllocator –

    Memory pool for tensor buffers to avoid frequent allocation/deallocation.

Functions:

  • gather_kv_caches –

    Gather KV cache data from KV cache storage to destination tensor.

  • scatter_kv_caches –

    Scatter KV cache data from source tensor to KV cache storage.

CopyBufferAllocator

Memory pool for tensor buffers to avoid frequent allocation/deallocation.

Methods:

Source code in vllm/distributed/kv_transfer/kv_connector/v1/hf3fs/utils/gather_scatter_helper.py 
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
class CopyBufferAllocator:
    """Memory pool for tensor buffers to avoid frequent allocation/deallocation."""

    def __init__(
        self, device: torch.device, dtype: torch.dtype, shape: list, max_count: int
    ):
        self._shape = shape
        self._max_count = max_count
        self._device = device
        self._free_buffers = [
            torch.empty(shape, dtype=dtype, device=device) for _ in range(max_count)
        ]
        self._inuse_count = 0

    def alloc_buffer(self, count: int) -> list[torch.Tensor] | None:
        """Allocate buffers from the pool."""
        if count == 0:
            return []

        if self._inuse_count + count <= self._max_count:
            self._inuse_count += count
            result = self._free_buffers[-count:]
            del self._free_buffers[-count:]
            return result
        return None

    def free_buffer(self, buffers: list[torch.Tensor]) -> None:
        """Return buffers to the pool."""
        if not buffers:
            return

        if self._inuse_count >= len(buffers):
            self._inuse_count -= len(buffers)
            self._free_buffers.extend(buffers)
        else:
            raise RuntimeError("Attempted to free more buffers than allocated")

alloc_buffer(count)

Allocate buffers from the pool.

Source code in vllm/distributed/kv_transfer/kv_connector/v1/hf3fs/utils/gather_scatter_helper.py 
238
239
240
241
242
243
244
245
246
247
248
def alloc_buffer(self, count: int) -> list[torch.Tensor] | None:
    """Allocate buffers from the pool."""
    if count == 0:
        return []

    if self._inuse_count + count <= self._max_count:
        self._inuse_count += count
        result = self._free_buffers[-count:]
        del self._free_buffers[-count:]
        return result
    return None

free_buffer(buffers)

Return buffers to the pool.

Source code in vllm/distributed/kv_transfer/kv_connector/v1/hf3fs/utils/gather_scatter_helper.py 
250
251
252
253
254
255
256
257
258
259
def free_buffer(self, buffers: list[torch.Tensor]) -> None:
    """Return buffers to the pool."""
    if not buffers:
        return

    if self._inuse_count >= len(buffers):
        self._inuse_count -= len(buffers)
        self._free_buffers.extend(buffers)
    else:
        raise RuntimeError("Attempted to free more buffers than allocated")

gather_kv_caches(kv_caches_ptrs, total_token_in_kvcache, dst_tensor, token_indices, tokens_per_block, num_heads, content_size, kv_cache_strides)

Gather KV cache data from KV cache storage to destination tensor.

Parameters:

  • kv_caches_ptrs

    (Tensor) –

    Tensor of KV cache pointers (one per layer)

  • total_token_in_kvcache

    (int) –

    Total number of tokens in KV cache

  • dst_tensor

    (Tensor) –

    Destination [L, H, N, C] tensor

  • token_indices

    (list[int]) –

    List of token positions to gather

  • tokens_per_block

    (int) –

    Number of stored states in each cache block

  • num_heads

    (int) –

    Size of the H axis

  • content_size

    (int) –

    Size of the C axis

  • kv_cache_strides

    (tuple[int, ...]) –

    Element strides of each [B, H, N, C] layer view

Source code in vllm/distributed/kv_transfer/kv_connector/v1/hf3fs/utils/gather_scatter_helper.py 
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
def gather_kv_caches(
    kv_caches_ptrs: torch.Tensor,
    total_token_in_kvcache: int,
    dst_tensor: torch.Tensor,
    token_indices: list[int],
    tokens_per_block: int,
    num_heads: int,
    content_size: int,
    kv_cache_strides: tuple[int, ...],
) -> None:
    """Gather KV cache data from KV cache storage to destination tensor.

    Args:
        kv_caches_ptrs: Tensor of KV cache pointers (one per layer)
        total_token_in_kvcache: Total number of tokens in KV cache
        dst_tensor: Destination ``[L, H, N, C]`` tensor
        token_indices: List of token positions to gather
        tokens_per_block: Number of stored states in each cache block
        num_heads: Size of the H axis
        content_size: Size of the C axis
        kv_cache_strides: Element strides of each ``[B, H, N, C]`` layer view

    """
    num_layers = kv_caches_ptrs.shape[0]
    num_tokens_in_block = len(token_indices)

    assert dst_tensor.shape == (
        num_layers,
        num_heads,
        num_tokens_in_block,
        content_size,
    )
    assert len(kv_cache_strides) == 4

    device = dst_tensor.device
    token_indices_tensor = async_tensor_h2d(
        token_indices, device=device, dtype=torch.int32
    )

    grid = (num_layers, num_tokens_in_block)
    BLOCK_SIZE = 128

    kv_cache_gather_kernel[grid](
        kv_caches_ptrs,
        dst_tensor,
        token_indices_tensor,
        num_tokens_in_block,
        total_token_in_kvcache,
        num_layers,
        tokens_per_block,
        *kv_cache_strides,
        num_heads=num_heads,
        content_size=content_size,
        BLOCK_SIZE=BLOCK_SIZE,
    )

scatter_kv_caches(kv_caches_ptrs, total_token_in_kvcache, src_tensor, token_indices, tokens_per_block, num_heads, content_size, kv_cache_strides)

Scatter KV cache data from source tensor to KV cache storage.

Parameters:

  • kv_caches_ptrs

    (Tensor) –

    Tensor of KV cache pointers (one per layer)

  • total_token_in_kvcache

    (int) –

    Total number of tokens in KV cache

  • src_tensor

    (Tensor) –

    Source [L, H, N, C] tensor containing data to scatter

  • token_indices

    (list[int]) –

    List of token positions to update

  • tokens_per_block

    (int) –

    Number of stored states in each cache block

  • num_heads

    (int) –

    Size of the H axis

  • content_size

    (int) –

    Size of the C axis

  • kv_cache_strides

    (tuple[int, ...]) –

    Element strides of each [B, H, N, C] layer view

Source code in vllm/distributed/kv_transfer/kv_connector/v1/hf3fs/utils/gather_scatter_helper.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
def scatter_kv_caches(
    kv_caches_ptrs: torch.Tensor,
    total_token_in_kvcache: int,
    src_tensor: torch.Tensor,
    token_indices: list[int],
    tokens_per_block: int,
    num_heads: int,
    content_size: int,
    kv_cache_strides: tuple[int, ...],
) -> None:
    """Scatter KV cache data from source tensor to KV cache storage.

    Args:
        kv_caches_ptrs: Tensor of KV cache pointers (one per layer)
        total_token_in_kvcache: Total number of tokens in KV cache
        src_tensor: Source ``[L, H, N, C]`` tensor containing data to scatter
        token_indices: List of token positions to update
        tokens_per_block: Number of stored states in each cache block
        num_heads: Size of the H axis
        content_size: Size of the C axis
        kv_cache_strides: Element strides of each ``[B, H, N, C]`` layer view

    """
    num_layers = len(kv_caches_ptrs)
    num_tokens_in_block = len(token_indices)

    assert src_tensor.shape == (
        num_layers,
        num_heads,
        num_tokens_in_block,
        content_size,
    )
    assert len(kv_cache_strides) == 4

    device = src_tensor.device
    token_indices_tensor = async_tensor_h2d(
        token_indices, device=device, dtype=torch.int32
    )

    grid = (num_layers, num_tokens_in_block)
    BLOCK_SIZE = 128

    kv_cache_scatter_kernel[grid](
        kv_caches_ptrs,
        src_tensor,
        token_indices_tensor,
        num_tokens_in_block,
        total_token_in_kvcache,
        num_layers,
        tokens_per_block,
        *kv_cache_strides,
        num_heads=num_heads,
        content_size=content_size,
        BLOCK_SIZE=BLOCK_SIZE,
    )