biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

285
vllm/model_executor/models/minicpm.py
View File

@@ -23,6 +23,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only MiniCPM model compatible with HuggingFace weights."""
import math
from collections.abc import Iterable
from itertools import islice
@@ -35,30 +36,42 @@ from transformers import PretrainedConfig
from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
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 FatreluAndMul, 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 (
DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class MiniCPMMoE(nn.Module):
@@ -90,34 +103,53 @@ class MiniCPMMoE(nn.Module):
params_dtype = torch.get_default_dtype()
self.params_dtype = params_dtype
self.gate = ReplicatedLinear(self.hidden_size,
self.num_total_experts,
bias=False,
params_dtype=self.params_dtype,
quant_config=None)
self.gate = ReplicatedLinear(
self.hidden_size,
self.num_total_experts,
bias=False,
params_dtype=self.params_dtype,
quant_config=None,
)
self.ws = nn.Parameter(
torch.empty(self.num_total_experts,
2 * self.intermediate_size,
self.hidden_size,
device=current_platform.device_type,
dtype=self.params_dtype))
torch.empty(
self.num_total_experts,
2 * self.intermediate_size,
self.hidden_size,
device=current_platform.device_type,
dtype=self.params_dtype,
)
)
self.w2s = nn.Parameter(
torch.empty(self.num_total_experts,
self.hidden_size,
self.intermediate_size,
device=current_platform.device_type,
dtype=self.params_dtype))
torch.empty(
self.num_total_experts,
self.hidden_size,
self.intermediate_size,
device=current_platform.device_type,
dtype=self.params_dtype,
)
)
set_weight_attrs(self.ws, {
"weight_loader": self.weight_loader,
})
set_weight_attrs(self.w2s, {
"weight_loader": self.weight_loader,
})
set_weight_attrs(
self.ws,
{
"weight_loader": self.weight_loader,
},
)
set_weight_attrs(
self.w2s,
{
"weight_loader": self.weight_loader,
},
)
def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
weight_name: str, expert_id: int):
def weight_loader(
self,
param: nn.Parameter,
loaded_weight: torch.Tensor,
weight_name: str,
expert_id: int,
):
tp_rank = get_tensor_model_parallel_rank()
param_data = param.data
shard_size = self.intermediate_size
@@ -125,8 +157,9 @@ class MiniCPMMoE(nn.Module):
if weight_name.endswith("w1.weight"):
param_data[expert_id, 0:shard_size, :] = loaded_weight[shard, :]
if weight_name.endswith("w3.weight"):
param_data[expert_id,
shard_size:2 * shard_size, :] = loaded_weight[shard, :]
param_data[expert_id, shard_size : 2 * shard_size, :] = loaded_weight[
shard, :
]
if weight_name.endswith("w2.weight"):
param_data[expert_id, :, :] = loaded_weight[:, shard]
@@ -136,27 +169,21 @@ class MiniCPMMoE(nn.Module):
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
topk_weights, topk_ids, _ = fused_topk(hidden_states,
router_logits,
self.top_k,
renormalize=True)
topk_weights, topk_ids, _ = fused_topk(
hidden_states, router_logits, self.top_k, renormalize=True
)
final_hidden_states = fused_experts(hidden_states,
self.ws,
self.w2s,
topk_weights,
topk_ids,
inplace=True)
final_hidden_states = fused_experts(
hidden_states, self.ws, self.w2s, topk_weights, topk_ids, inplace=True
)
if self.tp_size > 1:
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_size)
class MiniCPMMLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -167,20 +194,20 @@ class MiniCPMMLP(nn.Module):
) -> None:
super().__init__()
self.gate_up_proj = MergedColumnParallelLinear(
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)
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
)
if hidden_act == "silu":
self.act_fn = SiluAndMul()
elif hidden_act == "fatrelu":
self.act_fn = FatreluAndMul(threshold=hidden_act_param)
else:
raise ValueError(f"Unsupported activation: {hidden_act}. "
"Only silu and fatrelu are supported for now.")
raise ValueError(
f"Unsupported activation: {hidden_act}. "
"Only silu and fatrelu are supported for now."
)
def forward(self, x):
gate_up, _ = self.gate_up_proj(x)
@@ -190,7 +217,6 @@ class MiniCPMMLP(nn.Module):
class MiniCPMAttention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -249,13 +275,15 @@ class MiniCPMAttention(nn.Module):
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,
@@ -274,7 +302,6 @@ class MiniCPMAttention(nn.Module):
class MiniCPMDecoderLayer(nn.Module):
def __init__(
self,
config: PretrainedConfig,
@@ -289,15 +316,15 @@ class MiniCPMDecoderLayer(nn.Module):
self.hidden_size = config.hidden_size
self.rope_theta = getattr(config, "rope_theta", 10000)
self.rope_scaling = getattr(config, "rope_scaling", None)
self.max_position_embeddings = getattr(config,
"max_position_embeddings", 8192)
self.max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.prefix = prefix
self._init_attn_block()
self._init_ffn_block()
def _init_attn_block(self):
self.input_layernorm = RMSNorm(self.config.hidden_size,
eps=self.config.rms_norm_eps)
self.input_layernorm = RMSNorm(
self.config.hidden_size, eps=self.config.rms_norm_eps
)
self.self_attn = MiniCPMAttention(
hidden_size=self.hidden_size,
num_heads=self.config.num_attention_heads,
@@ -311,15 +338,16 @@ class MiniCPMDecoderLayer(nn.Module):
)
def _init_ffn_block(self):
self.post_attention_layernorm = RMSNorm(self.config.hidden_size,
eps=self.config.rms_norm_eps)
self.post_attention_layernorm = RMSNorm(
self.config.hidden_size, eps=self.config.rms_norm_eps
)
self.num_experts = getattr(self.config, "num_experts", 0)
if self.num_experts == 0:
self.mlp = MiniCPMMLP(
hidden_size=self.hidden_size,
intermediate_size=self.config.intermediate_size,
hidden_act=self.config.hidden_act,
hidden_act_param=getattr(self.config, "hidden_act_param", 0.),
hidden_act_param=getattr(self.config, "hidden_act_param", 0.0),
quant_config=self.quant_config,
)
else:
@@ -327,7 +355,8 @@ class MiniCPMDecoderLayer(nn.Module):
num_experts=self.config.num_experts,
top_k=self.config.num_experts_per_tok,
hidden_size=self.config.hidden_size,
intermediate_size=self.config.intermediate_size)
intermediate_size=self.config.intermediate_size,
)
def forward(
self,
@@ -342,22 +371,23 @@ class MiniCPMDecoderLayer(nn.Module):
positions=positions,
hidden_states=hidden_states,
)
hidden_states = residual + hidden_states * \
(self.config.scale_depth / math.sqrt(self.config.num_hidden_layers))
hidden_states = residual + hidden_states * (
self.config.scale_depth / math.sqrt(self.config.num_hidden_layers)
)
# Fully Connected
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states * \
(self.config.scale_depth / math.sqrt(self.config.num_hidden_layers))
hidden_states = residual + hidden_states * (
self.config.scale_depth / math.sqrt(self.config.num_hidden_layers)
)
return hidden_states, None
@support_torch_compile
class MiniCPMModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -369,8 +399,11 @@ class MiniCPMModel(nn.Module):
self.config = config
self.cache_config = cache_config
self.quant_config = quant_config
lora_vocab = (lora_config.lora_extra_vocab_size *
(lora_config.max_loras or 1)) if lora_config else 0
lora_vocab = (
(lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
if lora_config
else 0
)
self.vocab_size = config.vocab_size + lora_vocab
self.org_vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
@@ -384,9 +417,9 @@ class MiniCPMModel(nn.Module):
self.aux_hidden_state_layers = tuple[int, ...]()
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], self.config.hidden_size))
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], self.config.hidden_size
)
def _init_layers(
self,
@@ -398,8 +431,10 @@ class MiniCPMModel(nn.Module):
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: MiniCPMDecoderLayer(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers")
config, cache_config, quant_config, prefix=prefix
),
prefix=f"{prefix}.layers",
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
embedding = self.embed_tokens(input_ids)
@@ -411,8 +446,9 @@ class MiniCPMModel(nn.Module):
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors, tuple[torch.Tensor,
list[torch.Tensor]]]:
) -> Union[
torch.Tensor, IntermediateTensors, tuple[torch.Tensor, list[torch.Tensor]]
]:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
@@ -425,11 +461,12 @@ class MiniCPMModel(nn.Module):
aux_hidden_states = []
for idx, layer in enumerate(
islice(self.layers, self.start_layer, self.end_layer)):
islice(self.layers, self.start_layer, self.end_layer)
):
if idx in self.aux_hidden_state_layers:
aux_hidden_states.append(
hidden_states +
residual if residual is not None else hidden_states)
hidden_states + residual if residual is not None else hidden_states
)
hidden_states, residual = layer(
positions,
hidden_states,
@@ -437,10 +474,9 @@ class MiniCPMModel(nn.Module):
)
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)
@@ -448,8 +484,7 @@ class MiniCPMModel(nn.Module):
return hidden_states, aux_hidden_states
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"),
@@ -460,8 +495,11 @@ class MiniCPMModel(nn.Module):
]
expert_params_mapping = [
# (param_name, weight_name, expert_id)
("ws" if weight_name in ["w1", "w3"] else "w2s",
f"experts.{expert_id}.{weight_name}.weight", expert_id)
(
"ws" if weight_name in ["w1", "w3"] else "w2s",
f"experts.{expert_id}.{weight_name}.weight",
expert_id,
)
for expert_id in range(self.num_experts)
for weight_name in ["w1", "w2", "w3"]
]
@@ -471,12 +509,11 @@ class MiniCPMModel(nn.Module):
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if ("rotary_emb.cos_cached" in name
or "rotary_emb.sin_cached" in name):
if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
# Models trained using ColossalAI may include these tensors in
# the checkpoint. Skip them.
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)
@@ -498,10 +535,9 @@ class MiniCPMModel(nn.Module):
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param,
loaded_weight,
weight_name,
expert_id=expert_id)
weight_loader(
param, loaded_weight, weight_name, expert_id=expert_id
)
break
else:
# Skip loading extra bias for GPTQ models.
@@ -510,8 +546,9 @@ class MiniCPMModel(nn.Module):
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
@@ -551,8 +588,9 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
self.cache_config = cache_config
self.quant_config = quant_config
self.model = self._init_model(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = self._init_model(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
unpadded_vocab_size = config.vocab_size
if lora_config:
@@ -564,7 +602,8 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
padding_size=DEFAULT_VOCAB_PADDING_SIZE
# We need bigger padding if using lora for kernel
# compatibility
if not lora_config else lora_config.lora_vocab_padding_size,
if not lora_config
else lora_config.lora_vocab_padding_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
@@ -572,10 +611,10 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
self.scale_width = self.config.hidden_size / self.config.dim_model_base
self.logits_processor = LogitsProcessor(unpadded_vocab_size,
config.vocab_size)
self.logits_processor = LogitsProcessor(unpadded_vocab_size, config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
self.model.make_empty_intermediate_tensors
)
def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""):
return MiniCPMModel(vllm_config=vllm_config, prefix=prefix)
@@ -596,10 +635,12 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors, tuple[torch.Tensor,
list[torch.Tensor]]]:
model_output = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
) -> Union[
torch.Tensor, IntermediateTensors, tuple[torch.Tensor, list[torch.Tensor]]
]:
model_output = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
if isinstance(model_output, tuple) and len(model_output) == 2:
# Aux hidden states are present.
@@ -621,11 +662,9 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
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,
skip_prefixes=(["lm_head."]
if self.config.tie_word_embeddings else None),
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
)
return loader.load_weights(weights)