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

[core] clean up cudagraph batchsize padding logic (#10996)

Browse Source 
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-12-12 22:57:50 -08:00
committed by GitHub
parent 34f1a806d5
commit be39e3cd18
11 changed files with 150 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm/model_executor/models/mamba.py
View File

@@ -6,7 +6,7 @@ from torch import nn
from transformers import MambaConfig
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.config import _BATCH_SIZES_TO_CAPTURE, CacheConfig, VllmConfig
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.layernorm import RMSNorm
@@ -195,6 +195,17 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree, SupportsPP):
self.make_empty_intermediate_tensors = (
self.backbone.make_empty_intermediate_tensors)
if self.scheduler_config is not None and \
not self.model_config.enforce_eager:
if self.scheduler_config.max_num_seqs > \
vllm_config.compilation_config.max_capture_size:
self.max_batch_size = \
vllm_config.compilation_config.max_capture_size
else:
self.max_batch_size = vllm_config.pad_for_cudagraph(
self.scheduler_config.max_num_seqs)
else:
self.max_batch_size = 8192 + 2
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.backbone.get_input_embeddings(input_ids)
@@ -208,15 +219,11 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree, SupportsPP):
inputs_embeds: Optional[torch.Tensor] = None,
**kwargs):
if self.mamba_cache is None:
max_batch_size = (VllmConfig.get_graph_batch_size(
self.scheduler_config.max_num_seqs) if self.scheduler_config
else max(_BATCH_SIZES_TO_CAPTURE) + 2)
num_mamba_layers = self.model_config.get_num_layers_by_block_type(
self.vllm_config.parallel_config, LayerBlockType.mamba)
self.mamba_cache = MambaCacheManager(
self.lm_head.weight.dtype, num_mamba_layers, max_batch_size,
*self._get_mamba_cache_shape())
self.lm_head.weight.dtype, num_mamba_layers,
self.max_batch_size, *self._get_mamba_cache_shape())
(
mamba_cache_tensors,