[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:
@@ -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):
|
||||
|
||||
Reference in New Issue
Block a user