xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 05:15:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

[PP] Correct cache size check (#13873)

Browse Source 
Signed-off-by: Yang Zheng <zhengy.gator@gmail.com>
This commit is contained in:
Yang Zheng 2025-02-27 17:47:29 +08:00 committed by GitHub
parent 10c3b8c1cf
commit 4b1d141f49
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 14 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/worker/hpu_worker.py
View File

@ -258,9 +258,10 @@ class HPUWorker(LocalOrDistributedWorkerBase):
This also warms up the model, which may record CUDA graphs.
"""
raise_if_cache_size_invalid(num_gpu_blocks,
self.cache_config.block_size,
self.model_config.max_model_len)
raise_if_cache_size_invalid(
num_gpu_blocks, self.cache_config.block_size,
self.model_config.max_model_len,
self.parallel_config.pipeline_parallel_size)
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks
@ -442,13 +443,13 @@ def init_worker_distributed_environment(
parallel_config.pipeline_parallel_size)
def raise_if_cache_size_invalid(num_gpu_blocks, block_size,
max_model_len) -> None:
def raise_if_cache_size_invalid(num_gpu_blocks, block_size, max_model_len,
pipeline_parallel_size) -> None:
if num_gpu_blocks <= 0:
raise ValueError("No available memory for the cache blocks. "
"Try increasing `gpu_memory_utilization` when "
"initializing the engine.")
max_seq_len = block_size * num_gpu_blocks
max_seq_len = block_size * (num_gpu_blocks // pipeline_parallel_size)
if max_model_len > max_seq_len:
raise ValueError(
f"The model's max seq len ({max_model_len}) "

13
vllm/worker/worker.py
View File

@ -288,10 +288,11 @@ class Worker(LocalOrDistributedWorkerBase):
This also warms up the model, which may record CUDA graphs.
"""
raise_if_cache_size_invalid(num_gpu_blocks,
self.cache_config.block_size,
self.cache_config.is_attention_free,
self.model_config.max_model_len)
raise_if_cache_size_invalid(
num_gpu_blocks, self.cache_config.block_size,
self.cache_config.is_attention_free,
self.model_config.max_model_len,
self.parallel_config.pipeline_parallel_size)
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks
@ -530,7 +531,7 @@ def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):
def raise_if_cache_size_invalid(num_gpu_blocks, block_size, is_attention_free,
max_model_len) -> None:
max_model_len, pipeline_parallel_size) -> None:
if is_attention_free and num_gpu_blocks != 0:
raise ValueError("No memory should be allocated for the cache blocks "
f"for an attention-free model, but {num_gpu_blocks} "
@ -539,7 +540,7 @@ def raise_if_cache_size_invalid(num_gpu_blocks, block_size, is_attention_free,
raise ValueError("No available memory for the cache blocks. "
"Try increasing `gpu_memory_utilization` when "
"initializing the engine.")
max_seq_len = block_size * num_gpu_blocks
max_seq_len = block_size * (num_gpu_blocks // pipeline_parallel_size)
if not is_attention_free and max_model_len > max_seq_len:
raise ValueError(
f"The model's max seq len ({max_model_len}) "