A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 03:15 UTC
[Bugfix] Move the _touch(computed_blocks) call in the allocate_slots method to after the check for allocating new blocks. (#11565)
sakunkun committed
Dec 31, 2024 at 14:29 UTC
2c5718809bb5f4bce2ae8e05041d613215dac1aa
2 files changed
+74
-8
tests/v1/core/test_prefix_caching.py
+61
-2
@@ -98,9 +98,9 @@ def test_prefill():
98
# Incomplete 1 block (6 tokens)
99
unique_token_ids = [3] * 6
100
req2 = make_request("2", common_token_ids + unique_token_ids)
101
- computed_block = manager.get_computed_blocks(req2)
101
+ computed_blocks = manager.get_computed_blocks(req2)
102
assert len(req2.kv_block_hashes) == 3
103
- assert [b.block_id for b in computed_block] == [0, 1, 2]
103
+ assert [b.block_id for b in computed_blocks] == [0, 1, 2]
104
num_new_tokens = 53 - 3 * 16
105
blocks = manager.allocate_slots(req2, num_new_tokens, computed_blocks)
106
assert [b.block_id for b in blocks] == [7, 8]
@@ -500,3 +500,62 @@ def test_mm_prefix_caching():
500
mm_hashes=mm_hashes)
501
computed_blocks = manager.get_computed_blocks(req1)
502
assert len(computed_blocks) == 3
503
+
504
+
505
+def test_prefill_not_enough_free_blocks_with_computed_blocks():
506
+ """
507
+ This is a unit test that tests the correctness of the allocate_slots
508
+ when there is not enough free blocks. Specifically, when a request
509
+ has computed blocks but cannot be allocated due to not enough free blocks,
510
+ the computed blocks should not be touched.
511
+ """
512
+ block_size = 16
513
+ manager = KVCacheManager(
514
+ block_size=block_size,
515
+ num_gpu_blocks=10,
516
+ max_model_len=8192,
517
+ sliding_window=None,
518
+ enable_caching=True,
519
+ num_preallocate_tokens=0,
520
+ )
521
+ # Complete 3 blocks (48 tokens)
522
+ # | Common-0 | Common-1 | Common-2 | ... |
523
+ common_token_ids = [i for i in range(3) for _ in range(16)]
524
+ req0 = make_request("0", common_token_ids)
525
+ computed_blocks = manager.get_computed_blocks(req0)
526
+ assert not computed_blocks
527
+ manager.allocate_slots(req0, 48, computed_blocks)
528
+ block_part0 = manager.req_to_blocks[req0.request_id]
529
+
530
+ # | Common-0 | Common-1 | Common-2 | Req1-3 | Req1-4 | Req1-5 | ... |
531
+ req1 = make_request("1", common_token_ids * 2)
532
+ computed_blocks = manager.get_computed_blocks(req1)
533
+ assert computed_blocks == block_part0
534
+ manager.allocate_slots(req1, 48, computed_blocks)
535
+ block_part1 = manager.req_to_blocks[req1.request_id]
536
+ # | Common-0 | Common-1 | Common-2 | Req1-3 (F) | Req1-4 (F) |
537
+ # | Req1-5(F)| ... |
538
+ manager.free(req1)
539
+ assert {block.ref_cnt for block in block_part1[:3]} == {1}
540
+ assert {block.ref_cnt for block in block_part1[3:]} == {0}
541
+
542
+ # | Common-0 | Common-1 | Common-2 | Req1-3 (F) | Req1-4 (F) |
543
+ # | Req1-5(F)| Req2-0 | Req2-1 | ... |
544
+ req2 = make_request("2", [7] * block_size * 2)
545
+ computed_blocks = manager.get_computed_blocks(req2)
546
+ assert not computed_blocks
547
+ manager.allocate_slots(req2, block_size * 2, computed_blocks)
548
+
549
+ # Req3 is Req2 + 3 new blocks, so the first 6 blocks are computed,
550
+ # but it cannot be allocated due to insufficient free blocks (2).
551
+ # In this case, the ref_cnt of the computed blocks should not be changed.
552
+ assert manager.free_block_queue.num_free_blocks == 5
553
+ req3 = make_request("3", common_token_ids * 3)
554
+ computed_blocks = manager.get_computed_blocks(req3)
555
+ assert computed_blocks == block_part1
556
+ # Req3 cannot be allocated.
557
+ assert manager.allocate_slots(req3, 48, computed_blocks) is None
558
+ # Block 0-2 are used by Req 1.
559
+ assert {block.ref_cnt for block in block_part1[:3]} == {1}
560
+ # Block 3-5 are free.
561
+ assert {block.ref_cnt for block in block_part1[3:]} == {0}
vllm/v1/core/kv_cache_manager.py
+13
-6
@@ -191,7 +191,7 @@ class KVCacheManager:
191
request: The request to allocate slots.
192
num_tokens: The number of tokens to allocate. Note that this does
193
not include the tokens that have already been computed.
194
- computed_blocks: The blocks that have already been computed.
194
+ computed_blocks: A list of computed blocks.
195
196
Returns:
197
A list of new allocated blocks.
@@ -200,6 +200,18 @@ class KVCacheManager:
200
raise ValueError(
201
f"num_tokens must be greater than 0, got {num_tokens}")
202
203
+ # If a computed block of a request is an eviction candidate (in the
204
+ # free queue and ref_cnt == 0), it cannot be counted as a free block
205
+ # when allocating this request.
206
+ num_evictable_computed_blocks = sum(1 for blk in computed_blocks
207
+ if blk.ref_cnt == 0)
208
+
209
+ num_required_blocks = cdiv(num_tokens, self.block_size)
210
+ if (num_required_blocks > self.free_block_queue.num_free_blocks -
211
+ num_evictable_computed_blocks):
212
+ # Cannot allocate new blocks.
213
+ return None
214
+
215
# Touch the computed blocks to make sure they won't be evicted.
216
if self.enable_caching:
217
self._touch(computed_blocks)
@@ -208,11 +220,6 @@ class KVCacheManager:
220
"Computed blocks should be empty when "
221
"prefix caching is disabled")
222
211
- num_required_blocks = cdiv(num_tokens, self.block_size)
212
- if (num_required_blocks > self.free_block_queue.num_free_blocks):
213
- # Cannot allocate new blocks.
214
- return None
215
-
223
# Determine the number of new blocks to allocate considering
224
# preallocated blocks.
225
num_new_blocks = min(