biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core] Deprecating block manager v1 and make block manager v2 default (#8704)

Browse Source 
Removing the block manager v1. This is the initial piece of prefix-caching-centric design. In order to achieve prefix-caching-centric design, we need to simplify the code path so that we only use v2 block manager (which has much higher performance on prefix caching).
This commit is contained in:
Kuntai Du
2024-10-17 11:38:15 -05:00
committed by GitHub
parent 5eda21e773
commit 81ede99ca4
45 changed files with 206 additions and 2109 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
vllm/config.py
View File

@@ -949,7 +949,6 @@ class SchedulerConfig:
iteration.
max_model_len: Maximum length of a sequence (including prompt
and generated text).
use_v2_block_manager: Whether to use the BlockSpaceManagerV2 or not.
num_lookahead_slots: The number of slots to allocate per sequence per
step, beyond the known token ids. This is used in speculative
decoding to store KV activations of tokens which may or may not be
@@ -976,7 +975,6 @@ class SchedulerConfig:
max_num_batched_tokens: Optional[int],
max_num_seqs: int,
max_model_len: int,
use_v2_block_manager: bool = True,
num_lookahead_slots: int = 0,
delay_factor: float = 0.0,
enable_chunked_prefill: bool = False,
@@ -1026,7 +1024,6 @@ class SchedulerConfig:
self.max_num_seqs = max_num_seqs
self.max_model_len = max_model_len
self.use_v2_block_manager = use_v2_block_manager
self.num_lookahead_slots = num_lookahead_slots
self.delay_factor = delay_factor
self.chunked_prefill_enabled = enable_chunked_prefill
@@ -1067,18 +1064,6 @@ class SchedulerConfig:
f"({self.num_scheduler_steps}) must be greater than or "
"equal to 1.")
if (not self.use_v2_block_manager \
and not envs.VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1):
raise ValueError(
"The use of BlockSpaceManagerV1 is deprecated and will "
"be removed in a future release. Please switch to "
"BlockSpaceManagerV2 by setting --use-v2-block-manager to "
"True. If you wish to suppress this error temporarily, "
"you can set the environment variable "
"`VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1. If your use "
"case is not supported in BlockSpaceManagerV2, please "
"file an issue with detailed information.")
@property
def is_multi_step(self) -> bool:
return self.num_scheduler_steps > 1
@@ -1137,7 +1122,6 @@ class SpeculativeConfig:
speculative_disable_mqa_scorer: Optional[bool],
speculative_max_model_len: Optional[int],
enable_chunked_prefill: bool,
use_v2_block_manager: bool,
disable_log_stats: bool,
speculative_disable_by_batch_size: Optional[int],
ngram_prompt_lookup_max: Optional[int],
@@ -1178,9 +1162,6 @@ class SpeculativeConfig:
enable_chunked_prefill (bool): Whether vLLM is configured to use
chunked prefill or not. Used for raising an error since its not
yet compatible with spec decode.
use_v2_block_manager (bool): Whether vLLM is configured to use the
v2 block manager or not. Used for raising an error since the v2
block manager is required with spec decode.
speculative_disable_by_batch_size (Optional[int]): Disable
speculative decoding for new incoming requests when the number
of enqueue requests is larger than this value, if provided.
@@ -1231,11 +1212,6 @@ class SpeculativeConfig:
"Speculative decoding and chunked prefill are "
f"currently mutually exclusive ({enable_chunked_prefill=}).")
if not use_v2_block_manager:
raise ValueError(
"Speculative decoding requires usage of the V2 "
"block manager. Enable it with --use-v2-block-manager.")
# TODO: The user should be able to specify revision/max model len
# for the draft model. It is not currently supported.
draft_revision = None