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

4
tests/multi_step/test_correctness_llm.py
View File

@@ -76,7 +76,6 @@ def test_multi_step_llm(
enforce_eager=enforce_eager,
gpu_memory_utilization=0.7,
tensor_parallel_size=tp_size,
use_v2_block_manager=True,
enable_chunked_prefill=enable_chunked_prefill,
num_scheduler_steps=num_scheduler_steps,
) as vllm_model:
@@ -169,7 +168,6 @@ def test_multi_step_llm_w_prompt_logprobs(
enforce_eager=enforce_eager,
gpu_memory_utilization=0.7,
tensor_parallel_size=tp_size,
use_v2_block_manager=True,
num_scheduler_steps=num_scheduler_steps,
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
@@ -305,7 +303,6 @@ def test_multi_step_llm_chunked_prefill_prefix_cache(
enforce_eager=enforce_eager,
gpu_memory_utilization=0.7,
tensor_parallel_size=tp_size,
use_v2_block_manager=True,
num_scheduler_steps=num_scheduler_steps,
max_model_len=48,
max_num_batched_tokens=48,
@@ -324,7 +321,6 @@ def test_multi_step_llm_chunked_prefill_prefix_cache(
enforce_eager=enforce_eager,
gpu_memory_utilization=0.7,
tensor_parallel_size=tp_size,
use_v2_block_manager=True,
enable_chunked_prefill=True,
enable_prefix_caching=True,
num_scheduler_steps=num_scheduler_steps,