Skip to content

[Model] Muse Glimmer: reuse in-tree RMSNorm / GemmaRMSNorm #64920

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jeejeelee merged 3 commits into from
Nov 2, 2026
Merged

[Model] Muse Glimmer: reuse in-tree RMSNorm / GemmaRMSNorm #64920

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
2717556
Rework MuseGlimmerRMSNorm with CustomOps
bohnstingl Oct 31, 2026
f4fe8c8
Merge branch 'main' into muse_glimmerrmsnorm
vllm-agent Nov 2, 2026
59c7677
Merge branch 'main' of github.com:vllm-project/vllm into muse_glimmer…
bohnstingl Nov 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
70 changes: 20 additions & 50 deletions vllm/model_executor/models/muse_glimmer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention, MMEncoderAttention
from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
MergedColumnParallelLinear,
Expand Down Expand Up @@ -343,41 +344,6 @@ def video_replacement(item_idx: int) -> PromptUpdateDetails:
]


class MuseGlimmerRMSNorm(nn.Module):
"""RMSNorm mirroring HF MuseGlimmer exactly (fp32 compute, cast at the end).

``normed = _norm(x.float()) * (w.float() + weight_offset)`` cast back to the
input dtype. When ``with_scale`` is False the layer is weightless (used for
QK-norm and the token-embedding norm).
"""

def __init__(
self,
dim: int | None = None,
eps: float = 1e-6,
with_scale: bool = True,
weight_offset: int = 0,
) -> None:
super().__init__()
self.eps = eps
self.with_scale = with_scale
self.weight_offset = weight_offset
if with_scale:
assert dim is not None
self.weight = nn.Parameter(torch.zeros(dim))
else:
self.register_parameter("weight", None)

def _norm(self, x: torch.Tensor) -> torch.Tensor:
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
out = self._norm(hidden_states.float())
if self.with_scale:
out = out * (self.weight.float() + self.weight_offset)
return out.type_as(hidden_states)


class MuseGlimmerVisionAttention(nn.Module):
def __init__(
self,
Expand Down Expand Up @@ -919,7 +885,7 @@ def __init__(
self.use_rope = bool(config.layer_rope_theta[self.layer_idx])

# Weightless, computed in fp32, applied per head over head_dim.
self.qk_norm = MuseGlimmerRMSNorm(eps=config.rms_norm_eps, with_scale=False)
self.qk_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps, has_weight=False)
self.scale_query_by = config.qk_scale_factor

self.qkv_proj = QKVParallelLinear(
Expand Down Expand Up @@ -1028,17 +994,15 @@ def __init__(
)
# Sandwich norms with baked +1 offset. Pre-norms use rms_norm_eps; the
# post-norms use the (typically smaller) post_norm_eps.
self.input_layernorm = MuseGlimmerRMSNorm(
config.hidden_size, eps=config.rms_norm_eps, weight_offset=1
self.input_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = GemmaRMSNorm(
config.hidden_size, eps=config.post_norm_eps
)
self.post_attention_layernorm = MuseGlimmerRMSNorm(
config.hidden_size, eps=config.post_norm_eps, weight_offset=1
self.pre_feedforward_layernorm = GemmaRMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
self.pre_feedforward_layernorm = MuseGlimmerRMSNorm(
config.hidden_size, eps=config.rms_norm_eps, weight_offset=1
)
self.post_feedforward_layernorm = MuseGlimmerRMSNorm(
config.hidden_size, eps=config.post_norm_eps, weight_offset=1
self.post_feedforward_layernorm = GemmaRMSNorm(
config.hidden_size, eps=config.post_norm_eps
)

def forward(
Expand Down Expand Up @@ -1078,7 +1042,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
)
# MuseGlimmer normalizes token embeddings with a weightless RMSNorm instead of
# Gemma's sqrt(hidden_size) multiplier.
self.embed_norm = MuseGlimmerRMSNorm(eps=config.rms_norm_eps, with_scale=False)
self.embed_norm = RMSNorm(
config.hidden_size, eps=config.rms_norm_eps, has_weight=False
)

self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
Expand All @@ -1088,7 +1054,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
prefix=f"{prefix}.layers",
)
# Final norm: weight-as-scale, no offset.
self.norm = MuseGlimmerRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.norm = RMSNorm(
config.hidden_size, eps=config.rms_norm_eps, dtype=torch.float32
)
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size
)
Expand Down Expand Up @@ -1250,7 +1218,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):

self.vision_encoder: MuseGlimmerVisionEncoder | None
self.vision_adapter: MuseGlimmerVisionAdapter | None
self.perception_emb_norm: MuseGlimmerRMSNorm | None
self.perception_emb_norm: RMSNorm | None
if self.has_vision:
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.vision_encoder = MuseGlimmerVisionEncoder(
Expand All @@ -1263,8 +1231,10 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
text_config.hidden_size,
bias=False,
)
self.perception_emb_norm = MuseGlimmerRMSNorm(
eps=text_config.rms_norm_eps, with_scale=False
self.perception_emb_norm = RMSNorm(
text_config.hidden_size,
eps=text_config.rms_norm_eps,
has_weight=False,
)
else:
self.vision_encoder = None
Expand Down