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

[V1] Remove V0 code paths for Hybrid models (#25400)

Browse Source 
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2025-09-23 17:26:13 +02:00
committed by GitHub
parent 2c58742dff
commit a903669e10
31 changed files with 352 additions and 2296 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
vllm/model_executor/models/mamba.py
View File

@@ -8,7 +8,6 @@ import torch
from torch import nn
from transformers import MambaConfig
from vllm import envs
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed.parallel_state import get_pp_group
@@ -24,10 +23,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.interfaces import (HasInnerState,
IsAttentionFree, SupportsPP)
from vllm.model_executor.models.mamba_cache import (MambaCacheManager,
MambaCacheParams)
from vllm.sequence import IntermediateTensors
from vllm.utils import LayerBlockType
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
@@ -72,7 +68,6 @@ class MambaDecoderLayer(nn.Module):
self,
hidden_states: torch.Tensor,
residual: Optional[torch.Tensor],
mamba_cache_params: MambaCacheParams,
**kwargs,
):
if residual is None:
@@ -82,7 +77,7 @@ class MambaDecoderLayer(nn.Module):
hidden_states, residual = self.norm(hidden_states, residual)
output = torch.empty_like(hidden_states)
self.mixer(hidden_states, output, mamba_cache_params)
self.mixer(hidden_states, output)
return output, residual
@@ -134,7 +129,6 @@ class MambaModel(nn.Module):
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
mamba_cache_params: Optional[MambaCacheParams] = None,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
@@ -151,17 +145,9 @@ class MambaModel(nn.Module):
for i in range(self.start_layer, self.end_layer):
layer = self.layers[i]
layer_cache_params = None
if mamba_cache_params is not None:
layer_cache_params = mamba_cache_params.at_layer_idx(
i - self.start_layer)
hidden_states, residual = layer(
positions=positions,
hidden_states=hidden_states,
residual=residual,
mamba_cache_params=layer_cache_params)
hidden_states, residual = layer(positions=positions,
hidden_states=hidden_states,
residual=residual)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
"hidden_states": hidden_states,
@@ -225,9 +211,6 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree, SupportsPP):
prefix=maybe_prefix(prefix, "lm_head"),
)
# Used to track and store by the Mamba cache between steps.
self.mamba_cache: Optional[MambaCacheManager] = None
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size)
@@ -244,22 +227,7 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree, SupportsPP):
inputs_embeds: Optional[torch.Tensor] = None,
**kwargs):
mamba_cache_params = None
if not envs.VLLM_USE_V1:
if self.mamba_cache is None:
num_layers = self.model_config.get_num_layers_by_block_type(
self.vllm_config.parallel_config, LayerBlockType.mamba)
state_shape = self.get_mamba_state_shape_from_config(
self.vllm_config)
state_dtype = self.get_mamba_state_dtype_from_config(
self.vllm_config)
self.mamba_cache = MambaCacheManager(self.vllm_config,
num_layers, *state_shape,
*state_dtype)
mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)
hidden_states = self.backbone(input_ids, positions, mamba_cache_params,
hidden_states = self.backbone(input_ids, positions,
intermediate_tensors, inputs_embeds)
return hidden_states
@@ -288,8 +256,7 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree, SupportsPP):
tp_world_size=parallel_config.tensor_parallel_size,
intermediate_size=hf_config.intermediate_size,
state_size=hf_config.state_size,
conv_kernel=hf_config.conv_kernel,
use_v1=envs.VLLM_USE_V1)
conv_kernel=hf_config.conv_kernel)
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
return self.mamba_cache.copy_inputs_before_cuda_graphs(