[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,
)