A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 05:45 UTC
[Model] Add has_weight to RMSNorm and re-enable weights loading tracker for Mamba (#10739)
Signed-off-by: Isotr0py <2037008807@qq.com>
Isotr0py committed
Dec 10, 2024 at 10:23 UTC
d1f6d1c8af892c7269f113711783374eebb52511
3 files changed
+34
-12
vllm/model_executor/layers/layernorm.py
+9
-2
@@ -20,6 +20,7 @@ class RMSNorm(CustomOp):
20
hidden_size: int,
21
eps: float = 1e-6,
22
var_hidden_size: Optional[int] = None,
23
+ has_weight: bool = True,
24
) -> None:
25
super().__init__()
26
@@ -27,7 +28,11 @@ class RMSNorm(CustomOp):
28
self.variance_epsilon = eps
29
self.variance_size_override = (None if var_hidden_size == hidden_size
30
else var_hidden_size)
30
- self.weight = nn.Parameter(torch.ones(hidden_size))
31
+ self.has_weight = has_weight
32
+
33
+ self.weight = torch.ones(hidden_size)
34
+ if self.has_weight:
35
+ self.weight = nn.Parameter(self.weight)
36
37
def forward_native(
38
self,
@@ -59,7 +64,9 @@ class RMSNorm(CustomOp):
64
variance = x_var.pow(2).mean(dim=-1, keepdim=True)
65
66
x = x * torch.rsqrt(variance + self.variance_epsilon)
62
- x = x.to(orig_dtype) * self.weight
67
+ x = x.to(orig_dtype)
68
+ if self.has_weight:
69
+ x = x * self.weight
70
if residual is None:
71
return x
72
else:
vllm/model_executor/layers/mamba/mamba_mixer.py
+18
-8
@@ -40,6 +40,7 @@ class MambaMixer(CustomOp):
40
use_conv_bias: bool,
41
use_bias: bool,
42
use_rms_norm: bool,
43
+ rms_norm_has_weight: bool = True,
44
rms_norm_eps: float = 1e-5,
45
activation="silu"):
46
super().__init__()
@@ -105,14 +106,23 @@ class MambaMixer(CustomOp):
106
input_is_parallel=True,
107
)
108
108
- self.dt_layernorm = RMSNorm(time_step_rank,
109
- eps=rms_norm_eps) if use_rms_norm else None
110
-
111
- self.b_layernorm = RMSNorm(ssm_state_size,
112
- eps=rms_norm_eps) if use_rms_norm else None
113
-
114
- self.c_layernorm = RMSNorm(ssm_state_size,
115
- eps=rms_norm_eps) if use_rms_norm else None
109
+ self.dt_layernorm = RMSNorm(
110
+ time_step_rank,
111
+ eps=rms_norm_eps,
112
+ has_weight=rms_norm_has_weight,
113
+ ) if use_rms_norm else None
114
+
115
+ self.b_layernorm = RMSNorm(
116
+ ssm_state_size,
117
+ eps=rms_norm_eps,
118
+ has_weight=rms_norm_has_weight,
119
+ ) if use_rms_norm else None
120
+
121
+ self.c_layernorm = RMSNorm(
122
+ ssm_state_size,
123
+ eps=rms_norm_eps,
124
+ has_weight=rms_norm_has_weight,
125
+ ) if use_rms_norm else None
126
127
def forward_native(self, hidden_states: torch.Tensor,
128
attn_metadata: AttentionMetadata,
vllm/model_executor/models/mamba.py
+7
-2
@@ -1,5 +1,5 @@
1
"""PyTorch MAMBA model."""
2
-from typing import Iterable, List, Optional, Tuple
2
+from typing import Iterable, List, Optional, Set, Tuple
3
4
import torch
5
from torch import nn
@@ -47,6 +47,7 @@ class MambaDecoderLayer(nn.Module):
47
use_conv_bias=config.use_conv_bias,
48
use_bias=config.use_bias,
49
use_rms_norm=self.is_falcon_mamba,
50
+ rms_norm_has_weight=not self.is_falcon_mamba,
51
rms_norm_eps=mixer_rms_eps,
52
activation=config.hidden_act)
53
@@ -241,8 +242,10 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree):
242
next_tokens = self.sampler(logits, sampling_metadata)
243
return next_tokens
244
244
- def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
245
+ def load_weights(self, weights: Iterable[Tuple[str,
246
+ torch.Tensor]]) -> Set[str]:
247
params_dict = dict(self.named_parameters())
248
+ loaded_params: Set[str] = set()
249
for name, loaded_weight in weights:
250
if "A_log" in name:
251
name = name.replace("A_log", "A")
@@ -254,3 +257,5 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree):
257
weight_loader = getattr(param, "weight_loader",
258
default_weight_loader)
259
weight_loader(param, loaded_weight)
260
+ loaded_params.add(name)
261
+ return loaded_params