[Model] use AutoWeightsLoader for olmoe,opt,orion,persimmon,phi3_small (#16548)

Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
This commit is contained in:
rongfu.leng
2025-04-17 17:48:31 +00:00
committed by GitHub
parent a018e555fd
commit 5125d72f02
5 changed files with 216 additions and 193 deletions
+58 -54
View File
@@ -39,7 +39,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsPP
from .utils import (is_pp_missing_parameter,
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
@@ -255,7 +255,7 @@ class OlmoeModel(nn.Module):
quant_config = vllm_config.quant_config
self.vocab_size = config.vocab_size
self.config = config
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
@@ -308,56 +308,6 @@ class OlmoeModel(nn.Module):
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
class OlmoeForCausalLM(nn.Module, SupportsPP):
fall_back_to_pt_during_load = False
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = OlmoeModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.lm_head = ParallelLMHead(config.vocab_size,
config.hidden_size,
quant_config=quant_config)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.sampler = get_sampler()
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
return hidden_states
def compute_logits(self, hidden_states: torch.Tensor,
sampling_metadata: SamplingMetadata) -> torch.Tensor:
logits = self.logits_processor(self.lm_head, hidden_states,
sampling_metadata)
return logits
def sample(
self,
logits: Optional[torch.Tensor],
sampling_metadata: SamplingMetadata,
) -> Optional[SamplerOutput]:
next_tokens = self.sampler(logits, sampling_metadata)
return next_tokens
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
stacked_params_mapping = [
@@ -380,8 +330,6 @@ class OlmoeForCausalLM(nn.Module, SupportsPP):
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in 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:
@@ -453,3 +401,59 @@ class OlmoeForCausalLM(nn.Module, SupportsPP):
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
class OlmoeForCausalLM(nn.Module, SupportsPP):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = OlmoeModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.lm_head = ParallelLMHead(config.vocab_size,
config.hidden_size,
quant_config=quant_config)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.sampler = get_sampler()
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
return hidden_states
def compute_logits(self, hidden_states: torch.Tensor,
sampling_metadata: SamplingMetadata) -> torch.Tensor:
logits = self.logits_processor(self.lm_head, hidden_states,
sampling_metadata)
return logits
def sample(
self,
logits: Optional[torch.Tensor],
sampling_metadata: SamplingMetadata,
) -> Optional[SamplerOutput]:
next_tokens = self.sampler(logits, sampling_metadata)
return next_tokens
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
loader = AutoWeightsLoader(
self,
skip_prefixes=["rotary_emb.inv_freq"],
)
return loader.load_weights(weights)