Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -23,6 +23,7 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Inference-only Deepseek model."""
|
||||
|
||||
from collections.abc import Iterable
|
||||
from itertools import islice
|
||||
from typing import Any, Optional, Union
|
||||
@@ -33,33 +34,43 @@ from transformers import PretrainedConfig
|
||||
|
||||
from vllm.attention import Attention
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.distributed import (
|
||||
get_pp_group,
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce,
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.linear import (
|
||||
MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear,
|
||||
)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
ParallelLMHead, VocabParallelEmbedding)
|
||||
ParallelLMHead,
|
||||
VocabParallelEmbedding,
|
||||
)
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from .interfaces import SupportsLoRA, SupportsPP
|
||||
from .utils import (AutoWeightsLoader, extract_layer_index,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory, make_layers,
|
||||
maybe_prefix)
|
||||
from .utils import (
|
||||
AutoWeightsLoader,
|
||||
extract_layer_index,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory,
|
||||
make_layers,
|
||||
maybe_prefix,
|
||||
)
|
||||
|
||||
|
||||
class DeepseekMLP(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
@@ -71,17 +82,19 @@ class DeepseekMLP(nn.Module):
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.gate_up_proj = MergedColumnParallelLinear(
|
||||
hidden_size, [intermediate_size] * 2,
|
||||
hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
|
||||
)
|
||||
self.down_proj = RowParallelLinear(
|
||||
intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
quant_config=quant_config)
|
||||
self.down_proj = RowParallelLinear(intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
quant_config=quant_config,
|
||||
reduce_results=reduce_results)
|
||||
quant_config=quant_config,
|
||||
reduce_results=reduce_results,
|
||||
)
|
||||
if hidden_act != "silu":
|
||||
raise ValueError(f"Unsupported activation: {hidden_act}. "
|
||||
"Only silu is supported for now.")
|
||||
raise ValueError(
|
||||
f"Unsupported activation: {hidden_act}. Only silu is supported for now."
|
||||
)
|
||||
self.act_fn = SiluAndMul()
|
||||
|
||||
def forward(self, x):
|
||||
@@ -92,7 +105,6 @@ class DeepseekMLP(nn.Module):
|
||||
|
||||
|
||||
class DeepseekMoE(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
@@ -108,26 +120,29 @@ class DeepseekMoE(nn.Module):
|
||||
if self.tp_size > self.n_routed_experts:
|
||||
raise ValueError(
|
||||
f"Tensor parallel size {self.tp_size} is greater than "
|
||||
f"the number of experts {self.n_routed_experts}.")
|
||||
f"the number of experts {self.n_routed_experts}."
|
||||
)
|
||||
|
||||
self.experts = nn.ModuleList([
|
||||
DeepseekMLP(hidden_size=config.hidden_size,
|
||||
intermediate_size=config.moe_intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=quant_config,
|
||||
reduce_results=False)
|
||||
for idx in range(self.n_routed_experts)
|
||||
])
|
||||
self.experts = nn.ModuleList(
|
||||
[
|
||||
DeepseekMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.moe_intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=quant_config,
|
||||
reduce_results=False,
|
||||
)
|
||||
for idx in range(self.n_routed_experts)
|
||||
]
|
||||
)
|
||||
self.pack_params()
|
||||
|
||||
self.gate = ReplicatedLinear(config.hidden_size,
|
||||
self.n_routed_experts,
|
||||
bias=False,
|
||||
quant_config=None)
|
||||
self.gate = ReplicatedLinear(
|
||||
config.hidden_size, self.n_routed_experts, bias=False, quant_config=None
|
||||
)
|
||||
|
||||
if config.n_shared_experts is not None:
|
||||
intermediate_size = (config.moe_intermediate_size *
|
||||
config.n_shared_experts)
|
||||
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
|
||||
self.shared_experts = DeepseekMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=intermediate_size,
|
||||
@@ -167,25 +182,21 @@ class DeepseekMoE(nn.Module):
|
||||
hidden_states,
|
||||
router_logits,
|
||||
self.top_k,
|
||||
renormalize=self.config.norm_topk_prob)
|
||||
renormalize=self.config.norm_topk_prob,
|
||||
)
|
||||
|
||||
final_hidden_states = fused_experts(hidden_states,
|
||||
self.w1,
|
||||
self.w2,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
inplace=True)
|
||||
final_hidden_states = fused_experts(
|
||||
hidden_states, self.w1, self.w2, topk_weights, topk_ids, inplace=True
|
||||
)
|
||||
|
||||
if self.config.n_shared_experts is not None:
|
||||
final_hidden_states = final_hidden_states + shared_output
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(
|
||||
final_hidden_states)
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
|
||||
|
||||
return final_hidden_states.view(num_tokens, hidden_dim)
|
||||
|
||||
|
||||
class DeepseekAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
@@ -244,13 +255,15 @@ class DeepseekAttention(nn.Module):
|
||||
base=rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
)
|
||||
self.attn = Attention(self.num_heads,
|
||||
self.head_dim,
|
||||
self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn")
|
||||
self.attn = Attention(
|
||||
self.num_heads,
|
||||
self.head_dim,
|
||||
self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn",
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -266,7 +279,6 @@ class DeepseekAttention(nn.Module):
|
||||
|
||||
|
||||
class DeepseekDecoderLayer(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
@@ -279,8 +291,7 @@ class DeepseekDecoderLayer(nn.Module):
|
||||
self.hidden_size = config.hidden_size
|
||||
rope_theta = getattr(config, "rope_theta", 10000)
|
||||
rope_scaling = getattr(config, "rope_scaling", None)
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings",
|
||||
8192)
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
||||
self.self_attn = DeepseekAttention(
|
||||
hidden_size=self.hidden_size,
|
||||
num_heads=config.num_attention_heads,
|
||||
@@ -292,12 +303,14 @@ class DeepseekDecoderLayer(nn.Module):
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.self_attn",
|
||||
)
|
||||
if (config.n_routed_experts is not None
|
||||
and layer_idx >= config.first_k_dense_replace
|
||||
and layer_idx % config.moe_layer_freq == 0):
|
||||
self.mlp = DeepseekMoE(config=config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.mlp")
|
||||
if (
|
||||
config.n_routed_experts is not None
|
||||
and layer_idx >= config.first_k_dense_replace
|
||||
and layer_idx % config.moe_layer_freq == 0
|
||||
):
|
||||
self.mlp = DeepseekMoE(
|
||||
config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
|
||||
)
|
||||
else:
|
||||
self.mlp = DeepseekMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
@@ -306,10 +319,10 @@ class DeepseekDecoderLayer(nn.Module):
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.mlp",
|
||||
)
|
||||
self.input_layernorm = RMSNorm(config.hidden_size,
|
||||
eps=config.rms_norm_eps)
|
||||
self.post_attention_layernorm = RMSNorm(config.hidden_size,
|
||||
eps=config.rms_norm_eps)
|
||||
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.post_attention_layernorm = RMSNorm(
|
||||
config.hidden_size, eps=config.rms_norm_eps
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -322,22 +335,19 @@ class DeepseekDecoderLayer(nn.Module):
|
||||
residual = hidden_states
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
else:
|
||||
hidden_states, residual = self.input_layernorm(
|
||||
hidden_states, residual)
|
||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||
hidden_states = self.self_attn(
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
)
|
||||
|
||||
# Fully Connected
|
||||
hidden_states, residual = self.post_attention_layernorm(
|
||||
hidden_states, residual)
|
||||
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
|
||||
hidden_states = self.mlp(hidden_states)
|
||||
return hidden_states, residual
|
||||
|
||||
|
||||
class DeepseekModel(nn.Module):
|
||||
|
||||
fall_back_to_pt_during_load = False
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
@@ -358,11 +368,12 @@ class DeepseekModel(nn.Module):
|
||||
lambda prefix: DeepseekDecoderLayer(
|
||||
config, cache_config, quant_config=quant_config, prefix=prefix
|
||||
),
|
||||
prefix=f"{prefix}.layers")
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size))
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.embed_tokens(input_ids)
|
||||
@@ -386,15 +397,13 @@ class DeepseekModel(nn.Module):
|
||||
for layer in islice(self.layers, self.start_layer, self.end_layer):
|
||||
hidden_states, residual = layer(positions, hidden_states, residual)
|
||||
if not get_pp_group().is_last_rank:
|
||||
return IntermediateTensors({
|
||||
"hidden_states": hidden_states,
|
||||
"residual": residual
|
||||
})
|
||||
return IntermediateTensors(
|
||||
{"hidden_states": hidden_states, "residual": residual}
|
||||
)
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
return hidden_states
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("qkv_proj", "q_proj", "q"),
|
||||
@@ -409,7 +418,7 @@ class DeepseekModel(nn.Module):
|
||||
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:
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
if weight_name not in name:
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
@@ -417,8 +426,9 @@ class DeepseekModel(nn.Module):
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
# Skip experts that are not assigned to this worker.
|
||||
if (("mlp.experts." in name or "mlp.shared_experts." in name)
|
||||
and name not in params_dict):
|
||||
if (
|
||||
"mlp.experts." in name or "mlp.shared_experts." in name
|
||||
) and name not in params_dict:
|
||||
continue
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
@@ -431,14 +441,14 @@ class DeepseekModel(nn.Module):
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
# Skip experts that are not assigned to this worker.
|
||||
if (("mlp.experts." in name or "mlp.shared_experts." in name)
|
||||
and name not in params_dict):
|
||||
if (
|
||||
"mlp.experts." in name or "mlp.shared_experts." in name
|
||||
) and name not in params_dict:
|
||||
continue
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(param, loaded_weight)
|
||||
loaded_params.add(name)
|
||||
return loaded_params
|
||||
@@ -456,8 +466,9 @@ class DeepseekForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
quant_config = vllm_config.quant_config
|
||||
self.config = config
|
||||
self.quant_config = quant_config
|
||||
self.model = DeepseekModel(vllm_config=vllm_config,
|
||||
prefix=maybe_prefix(prefix, "model"))
|
||||
self.model = DeepseekModel(
|
||||
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
|
||||
)
|
||||
self.lm_head = ParallelLMHead(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
@@ -468,7 +479,8 @@ class DeepseekForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
self.lm_head.weight = self.model.embed_tokens.weight
|
||||
self.logits_processor = LogitsProcessor(config.vocab_size)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
self.model.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)
|
||||
@@ -480,8 +492,9 @@ class DeepseekForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
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)
|
||||
hidden_states = self.model(
|
||||
input_ids, positions, intermediate_tensors, inputs_embeds
|
||||
)
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(
|
||||
@@ -491,7 +504,6 @@ class DeepseekForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
logits = self.logits_processor(self.lm_head, hidden_states)
|
||||
return logits
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
loader = AutoWeightsLoader(self)
|
||||
return loader.load_weights(weights)
|
||||
|
||||
Reference in New Issue
Block a user