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

79
vllm/model_executor/models/granitemoehybrid.py
View File

@@ -9,19 +9,15 @@ import torch
from torch import nn
from transformers import GraniteMoeHybridConfig
from vllm import envs
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.mamba2_metadata import (
Mamba2Metadata, prepare_mamba2_metadata)
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateDtypeCalculator, MambaStateShapeCalculator)
@@ -30,10 +26,7 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.mamba_cache import (MambaCacheManager,
MambaCacheParams)
from vllm.sequence import IntermediateTensors
from vllm.utils import LayerBlockType
from .granitemoe import GraniteMoeMoE
from .granitemoeshared import GraniteMoeSharedMLP
@@ -102,14 +95,12 @@ class GraniteMoeHybridMambaDecoderLayer(nn.Module):
self,
hidden_states: torch.Tensor,
residual: Optional[torch.Tensor],
mamba_cache_params: MambaCacheParams,
mamba2_metadata: Mamba2Metadata,
**kwargs,
):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
output = torch.empty_like(hidden_states)
self.mamba(hidden_states, output, mamba_cache_params, mamba2_metadata)
self.mamba(hidden_states, output)
hidden_states = residual + output * self.residual_multiplier
residual = hidden_states
@@ -182,8 +173,6 @@ class GraniteMoeHybridAttentionDecoderLayer(nn.Module):
positions: torch.Tensor,
hidden_states: torch.Tensor,
residual: Optional[torch.Tensor],
mamba_cache_params: MambaCacheParams,
mamba2_metadata: Mamba2Metadata,
) -> torch.Tensor:
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
@@ -366,22 +355,10 @@ class GraniteMoeHybridModel(nn.Module):
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
mamba_cache_params: MambaCacheParams,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
attn_metadata = get_forward_context().attn_metadata
if not envs.VLLM_USE_V1:
mamba2_metadata = prepare_mamba2_metadata(
chunk_size=self.config.mamba_chunk_size,
attn_metadata=attn_metadata,
)
else:
# v1 get mamba2_metadata from forward_context
mamba2_metadata = None
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
@@ -399,20 +376,9 @@ class GraniteMoeHybridModel(nn.Module):
for i, layer in enumerate(self.layers):
if isinstance(layer, GraniteMoeHybridAttentionDecoderLayer):
num_attn += 1
layer_mamba_cache_params = None
if isinstance(
layer,
GraniteMoeHybridMambaDecoderLayer) and mamba_cache_params:
layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
i - num_attn)
hidden_states, residual = layer(
positions=positions,
hidden_states=hidden_states,
residual=residual,
mamba_cache_params=layer_mamba_cache_params,
mamba2_metadata=mamba2_metadata)
hidden_states, residual = layer(positions=positions,
hidden_states=hidden_states,
residual=residual)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
@@ -552,13 +518,11 @@ class GraniteMoeHybridForCausalLM(nn.Module, HasInnerState, SupportsLoRA,
def get_mamba_state_shape_from_config(
cls,
vllm_config: "VllmConfig",
use_v1: bool = True,
) -> tuple[tuple[int, int], tuple[int, int, int]]:
"""Calculate shapes for Mamba's convolutional and state caches.
Args:
vllm_config: vLLM config
use_v1: Get shapes for V1 (or V0)
Returns:
Tuple containing:
@@ -577,7 +541,6 @@ class GraniteMoeHybridForCausalLM(nn.Module, HasInnerState, SupportsLoRA,
head_dim=hf_config.mamba_d_head,
state_size=hf_config.mamba_d_state,
conv_kernel=hf_config.mamba_d_conv,
use_v1=use_v1,
)
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
@@ -620,9 +583,6 @@ class GraniteMoeHybridForCausalLM(nn.Module, HasInnerState, SupportsLoRA,
scale=1 /
self.config.logits_scaling)
# Used to track and store by the Mamba cache between steps.
self.mamba_cache: Optional[MambaCacheManager] = None
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
@@ -636,38 +596,11 @@ class GraniteMoeHybridForCausalLM(nn.Module, HasInnerState, SupportsLoRA,
inputs_embeds: Optional[torch.Tensor] = None,
**kwargs):
mamba_cache_params = None
if not envs.VLLM_USE_V1:
if self.mamba_cache is None:
num_mamba_layers = (
self.model_config.get_num_layers_by_block_type(
self.vllm_config.parallel_config,
LayerBlockType.mamba))
mamba_state_shape = \
self.get_mamba_state_shape_from_config(
self.vllm_config, use_v1=False)
mamba_state_dtype = \
self.get_mamba_state_dtype_from_config(
self.vllm_config)
self.mamba_cache = MambaCacheManager(self.vllm_config,
num_mamba_layers,
*mamba_state_shape,
*mamba_state_dtype)
mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)
hidden_states = self.model(input_ids, positions, mamba_cache_params,
intermediate_tensors, inputs_embeds)
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
return hidden_states
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
return self.mamba_cache.copy_inputs_before_cuda_graphs(
input_buffers, **kwargs)
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)
def compute_logits(
self,
hidden_states: torch.Tensor,