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

[V1] [Hybrid] Support using float32 for state in Hybrid Models (Mamba2, Mamba1, Minimax) (#22928)

Browse Source 
Signed-off-by: Daniel Afrimi <danielafrimi8@gmail.com>
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: Daniel Afrimi <danielafrimi8@gmail.com>
Co-authored-by: Burkhard Ringlein <ngl@zurich.ibm.com>
Co-authored-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
Thomas Parnell
2025-08-15 14:57:06 +02:00
committed by GitHub
parent 22341b996e
commit 75531a6c13
23 changed files with 467 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
vllm/model_executor/models/mamba2.py
View File

@@ -11,7 +11,7 @@ from transformers import MambaConfig
from vllm import envs
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.config import CacheConfig, ModelConfig, VllmConfig
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
@@ -20,7 +20,7 @@ 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 (
MambaStateShapeCalculator)
MambaStateDtypeCalculator, MambaStateShapeCalculator)
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -45,6 +45,8 @@ class Mamba2DecoderLayer(nn.Module):
def __init__(self,
config: MambaConfig,
model_config: Optional[ModelConfig] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "") -> None:
super().__init__()
@@ -62,6 +64,8 @@ class Mamba2DecoderLayer(nn.Module):
head_dim=config.head_dim,
rms_norm_eps=config.layer_norm_epsilon,
activation=config.hidden_act,
model_config=model_config,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.mixer")
@@ -93,6 +97,8 @@ class Mamba2Model(nn.Module):
super().__init__()
config = vllm_config.model_config.hf_config
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
is_lora_enabled = bool(lora_config)
@@ -112,8 +118,11 @@ class Mamba2Model(nn.Module):
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: Mamba2DecoderLayer(
config, quant_config=quant_config, prefix=prefix),
lambda prefix: Mamba2DecoderLayer(config,
model_config=model_config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix),
prefix=f"{prefix}.layers")
self.norm_f = RMSNorm(config.hidden_size,
@@ -200,6 +209,18 @@ class Mamba2Model(nn.Module):
class Mamba2ForCausalLM(nn.Module, HasInnerState, IsAttentionFree):
@classmethod
def get_mamba_state_dtype_from_config(
cls,
vllm_config: "VllmConfig",
) -> tuple[torch.dtype, torch.dtype]:
return MambaStateDtypeCalculator.mamba2_state_dtype(
vllm_config.model_config.dtype,
vllm_config.cache_config.mamba_cache_dtype,
vllm_config.cache_config.mamba_ssm_cache_dtype,
)
@classmethod
def get_mamba_state_shape_from_config(
cls,
@@ -290,10 +311,13 @@ class Mamba2ForCausalLM(nn.Module, HasInnerState, IsAttentionFree):
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,
self.lm_head.weight.dtype,
num_mamba_layers,
*mamba_state_shape)
*mamba_state_shape,
*mamba_state_dtype)
mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)
else: