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