[Bugfix][RL] Load attention sinks through weight_loader for layerwise reload (#57797)

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
This commit is contained in:
Ronald
2026-10-14 10:17:58 +00:00
committed by GitHub
co-authored by aoshen02
parent f9b65c30c7
commit 19aba3e4d5
30 changed files with 403 additions and 476 deletions
+2 -44
View File
@@ -43,6 +43,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
)
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
make_attention_sink,
maybe_remap_kv_scale_name,
remap_moe_expert_weights,
)
@@ -127,9 +128,7 @@ class OAIAttention(nn.Module):
tp_size = get_tensor_model_parallel_world_size()
self.sinks = torch.nn.Parameter(
torch.empty(config.num_attention_heads // tp_size, requires_grad=False)
)
self.sinks = make_attention_sink(config.num_attention_heads // tp_size)
self.q_size = self.num_attention_heads * self.head_dim // tp_size
self.kv_size = self.num_key_value_heads * self.head_dim // tp_size
@@ -556,8 +555,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
self,
ep_rank_end: int,
ep_rank_start: int,
heads_per_rank: int,
head_start: int,
weights: Iterable[tuple[str, torch.Tensor]],
stacked_params_mapping: list[tuple[str, str, str]],
) -> set[str]:
@@ -720,13 +717,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
)
loaded_params.add(name)
continue
elif "sinks" in name:
# Handle attention sinks (distributed across ranks)
param = params_dict[name]
narrow_weight = weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
continue
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
@@ -752,8 +742,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
self,
ep_rank_end: int,
ep_rank_start: int,
heads_per_rank: int,
head_start: int,
weights: Iterable[tuple[str, torch.Tensor]],
stacked_params_mapping: list[tuple[str, str, str]],
) -> set[str]:
@@ -1082,14 +1070,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
loaded_params.add(fused_name)
continue
elif "sinks" in name:
# Handle attention sinks (distributed across ranks)
param = params_dict[name]
narrow_weight = loaded_weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
continue
for param_name, weight_name, shard_id in stacked_params_mapping:
# Skip non-stacked layers and experts (experts handled below).
if weight_name not in name:
@@ -1168,8 +1148,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
self,
ep_rank_end: int,
ep_rank_start: int,
heads_per_rank: int,
head_start: int,
weights: Iterable[tuple[str, torch.Tensor]],
stacked_params_mapping: list[tuple[str, str, str]],
) -> set[str]:
@@ -1256,13 +1234,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
param.copy_(weight)
loaded_params.add(name)
continue
elif "sinks" in name:
# Handle attention sinks (distributed across ranks)
param = params_dict[name]
narrow_weight = weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
continue
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
@@ -1345,13 +1316,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
(".qkv_proj", ".v_proj", "v"),
]
tp_rank = get_tensor_model_parallel_rank()
tp_size = get_tensor_model_parallel_world_size()
# Attention heads per rank
heads_per_rank = self.config.num_attention_heads // tp_size
head_start = tp_rank * heads_per_rank
ep_size = get_ep_group().world_size
ep_rank = get_ep_group().rank_in_group
num_experts = self.config.num_local_experts
@@ -1383,8 +1347,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
return self._load_weights_mxfp4(
ep_rank_end,
ep_rank_start,
heads_per_rank,
head_start,
weights,
stacked_params_mapping,
)
@@ -1392,8 +1354,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
return self._load_weights_quark(
ep_rank_end,
ep_rank_start,
heads_per_rank,
head_start,
weights,
stacked_params_mapping,
)
@@ -1401,8 +1361,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
return self._load_weights_other(
ep_rank_end,
ep_rank_start,
heads_per_rank,
head_start,
weights,
stacked_params_mapping,
)
+3 -23
View File
@@ -14,7 +14,6 @@ from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
get_ep_group,
get_pp_group,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.logger import init_logger
@@ -36,6 +35,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import make_attention_sink
from vllm.model_executor.models.interfaces import (
EagleModelMixin,
SupportsEagle3,
@@ -354,9 +354,7 @@ class LagunaAttention(nn.Module):
# Attention sinks (learnable per-head bias for SWA layers)
sinks = None
if attention_sink:
self.sink = torch.nn.Parameter(
torch.empty(self.total_num_heads // tp_size, requires_grad=False)
)
self.sink = make_attention_sink(self.total_num_heads // tp_size)
sinks = self.sink
# Resolve rope params per-layer-type. ``config.rope_parameters`` is
@@ -677,22 +675,6 @@ class LagunaModel(nn.Module, EagleModelMixin):
return hidden_states, aux_hidden_states
return hidden_states
def _slice_sink(
self, weights: Iterable[tuple[str, torch.Tensor]]
) -> Iterable[tuple[str, torch.Tensor]]:
tp_size = get_tensor_model_parallel_world_size()
tp_rank = get_tensor_model_parallel_rank()
for name, loaded_weight in weights:
# Handle attention sinks (distributed across ranks). Derive the
# per-rank slice from the parameter's own shape so per-layer
# variations in head count are handled correctly.
if "sink" in name:
heads_per_rank = loaded_weight.shape[0] // tp_size
loaded_weight = loaded_weight.narrow(
0, tp_rank * heads_per_rank, heads_per_rank
)
yield name, loaded_weight
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(
self,
@@ -709,9 +691,7 @@ class LagunaModel(nn.Module, EagleModelMixin):
"_input_scale",
],
)
return loader.load_weights(
self._slice_sink(weights), mapper=self.hf_to_vllm_mapper
)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
class LagunaForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SupportsEagle3):
+5 -15
View File
@@ -47,6 +47,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
)
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
make_attention_sink,
maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.utils import sequence_parallel_chunk
@@ -290,9 +291,7 @@ class MiMoV2Attention(nn.Module):
)
self.attention_sink_bias = (
torch.nn.Parameter(torch.empty(self.num_heads), requires_grad=False)
if add_swa_attention_sink_bias
else None
make_attention_sink(self.num_heads) if add_swa_attention_sink_bias else None
)
sliding_window = sliding_window_size if sliding_window_size > -1 else None
@@ -848,18 +847,9 @@ class MiMoV2Model(nn.Module, EagleModelMixin):
param = params_dict[name]
if "attention_sink_bias" in name:
total_heads = loaded_weight.shape[0]
heads_per_rank = total_heads // tp_size
head_start = tp_rank * heads_per_rank
narrow_weight = loaded_weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
else:
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
@@ -377,14 +377,6 @@ class MiMoV2MTP(nn.Module):
continue
param = params_dict[name]
# attention_sink_bias is head-parallel; slice by tp
if "attention_sink_bias" in name:
total_heads = loaded_weight.shape[0]
heads_per_rank = total_heads // tp_size
loaded_weight = loaded_weight.narrow(
0, tp_rank * heads_per_rank, heads_per_rank
)
direct_weight_loader = getattr(
param, "weight_loader", default_weight_loader
)
+3 -22
View File
@@ -14,7 +14,6 @@ from vllm import _custom_ops as ops
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.logger import init_logger
@@ -33,6 +32,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import make_attention_sink
from vllm.multimodal.inputs import NestedTensors
from vllm.platforms import current_platform
from vllm.transformers_utils.config import set_default_rope_theta
@@ -251,9 +251,7 @@ class DFlashQwen3Attention(nn.Module):
)
self.attention_sink_bias = (
torch.nn.Parameter(torch.empty(self.num_heads), requires_grad=False)
if add_swa_attention_sink_bias
else None
make_attention_sink(self.num_heads) if add_swa_attention_sink_bias else None
)
self.sliding_window = sliding_window
@@ -746,26 +744,9 @@ class DFlashQwen3Model(nn.Module):
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
def _preprocess(
self, weights: Iterable[tuple[str, torch.Tensor]]
) -> Iterable[tuple[str, torch.Tensor]]:
tp_size = get_tensor_model_parallel_world_size()
tp_rank = get_tensor_model_parallel_rank()
for name, loaded_weight in weights:
if "attention_sink_bias" in name:
# Sink bias is per-head; shard it across TP ranks like the
# attention heads themselves.
heads_per_rank = loaded_weight.shape[0] // tp_size
loaded_weight = loaded_weight.narrow(
0, tp_rank * heads_per_rank, heads_per_rank
)
yield name, loaded_weight
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(self)
return loader.load_weights(
self._preprocess(weights), mapper=self.hf_to_vllm_mapper
)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
class DFlashQwen3ForCausalLM(Qwen3ForCausalLM):