[Bugfix][Models] Register sleep-managed runtime buffers (#53507)

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
This commit is contained in:
Ronald
2026-08-29 10:32:24 +08:00
committed by GitHub
co-authored by aoshen02
parent 7fd9cc036e
commit 16d6c376bc
6 changed files with 427 additions and 19 deletions
@@ -68,7 +68,9 @@ class RelPositionalEncoding(nn.Module):
pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
pe_negative = pe_negative[1:].unsqueeze(0)
self.pe = torch.cat([pe_positive, pe_negative], dim=1)
self.register_buffer(
"pe", torch.cat([pe_positive, pe_negative], dim=1), persistent=False
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# Tmax = 2 * max_len - 1
+13 -4
View File
@@ -343,8 +343,12 @@ class Ernie4_5_VisionPatchEmbed(nn.Module):
class Ernie4_5_VisionRotaryEmbedding(nn.Module):
def __init__(self, dim: int, theta: float = 10000.0) -> None:
super().__init__()
self.inv_freq = 1.0 / theta ** (
torch.arange(start=0, end=dim, step=2, dtype=torch.float32) / dim
self.register_buffer(
"inv_freq",
1.0
/ theta
** (torch.arange(start=0, end=dim, step=2, dtype=torch.float32) / dim),
persistent=False,
)
def forward(self, seqlen: int) -> torch.Tensor:
@@ -1452,11 +1456,16 @@ class Ernie4_5_VLMoeForConditionalGeneration(
]
if token_id is not None
]
self._visual_token_ids_tensor_cache = torch.tensor(
visual_token_ids_tensor_cache = torch.tensor(
visual_token_ids, dtype=torch.long
)
else:
self._visual_token_ids_tensor_cache = None
visual_token_ids_tensor_cache = None
self.register_buffer(
"_visual_token_ids_tensor_cache",
visual_token_ids_tensor_cache,
persistent=False,
)
def compute_logits(
self,
+29 -13
View File
@@ -114,8 +114,10 @@ class Gemma3nAltUp(nn.Module):
hidden_size=hidden_size,
eps=rms_norm_eps,
)
self.router_input_scale = torch.tensor(
hidden_size**-1.0, dtype=self.modality_router.weight.dtype
self.register_buffer(
"router_input_scale",
torch.tensor(hidden_size**-1.0, dtype=self.modality_router.weight.dtype),
persistent=False,
)
self.correct_output_scale = nn.Parameter(
torch.zeros(hidden_size, dtype=torch.float32)
@@ -597,9 +599,13 @@ class Gemma3nSelfDecoder(nn.Module):
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens",
)
self.embed_scale = torch.tensor(
config.hidden_size**0.5,
dtype=self.embed_tokens.weight.dtype,
self.register_buffer(
"embed_scale",
torch.tensor(
config.hidden_size**0.5,
dtype=self.embed_tokens.weight.dtype,
),
persistent=False,
)
# Additional per-layer embeddings (PLE)
self.embed_tokens_per_layer = VocabParallelEmbedding(
@@ -608,9 +614,13 @@ class Gemma3nSelfDecoder(nn.Module):
quant_config=quant_config,
prefix=f"{prefix}.per_layer_embed_tokens",
)
self.embed_scale_per_layer = torch.tensor(
config.hidden_size_per_layer_input**0.5,
dtype=self.embed_tokens.weight.dtype,
self.register_buffer(
"embed_scale_per_layer",
torch.tensor(
config.hidden_size_per_layer_input**0.5,
dtype=self.embed_tokens.weight.dtype,
),
persistent=False,
)
self.per_layer_model_projection = ColumnParallelLinear(
config.hidden_size,
@@ -625,12 +635,18 @@ class Gemma3nSelfDecoder(nn.Module):
hidden_size=config.hidden_size_per_layer_input,
eps=config.rms_norm_eps,
)
self.per_layer_input_scale = torch.rsqrt(torch.tensor(2.0)).to(
self.embed_tokens.weight.dtype
self.register_buffer(
"per_layer_input_scale",
torch.rsqrt(torch.tensor(2.0)).to(self.embed_tokens.weight.dtype),
persistent=False,
)
self.per_layer_projection_scale = torch.tensor(
config.hidden_size**0.5,
dtype=self.embed_tokens.weight.dtype,
self.register_buffer(
"per_layer_projection_scale",
torch.tensor(
config.hidden_size**0.5,
dtype=self.embed_tokens.weight.dtype,
),
persistent=False,
)
self.altup_projections = nn.ModuleList(
[
+5 -1
View File
@@ -765,7 +765,11 @@ class VoxtralEncoderModel(nn.Module):
max_frequency=8000.0,
sampling_rate=self.config.sampling_rate,
)
self.mel_filters = torch.tensor(mel_filters, dtype=torch.float32)
self.register_buffer(
"mel_filters",
torch.tensor(mel_filters, dtype=torch.float32),
persistent=False,
)
def compute_whisper_melspec(
self,