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

228
vllm/model_executor/models/dots1.py
View File

@@ -24,6 +24,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only dots1 model."""
from collections.abc import Iterable
from itertools import islice
from typing import Any, Optional, Union
@@ -35,33 +36,45 @@ from transformers import Dots1Config
from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import (get_pp_group,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce)
from vllm.distributed import (
get_pp_group,
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 FusedMoE
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, maybe_remap_kv_scale_name)
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class Dots1MLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -73,19 +86,24 @@ class Dots1MLP(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,
prefix=f"{prefix}.gate_up_proj")
self.down_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
prefix=f"{prefix}.down_proj")
prefix=f"{prefix}.gate_up_proj",
)
self.down_proj = RowParallelLinear(
intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
prefix=f"{prefix}.down_proj",
)
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):
@@ -96,7 +114,6 @@ class Dots1MLP(nn.Module):
class Dots1MoE(nn.Module):
def __init__(
self,
config: Dots1Config,
@@ -109,17 +126,22 @@ class Dots1MoE(nn.Module):
self.n_shared_experts = config.n_shared_experts
if config.hidden_act != "silu":
raise ValueError(f"Unsupported activation: {config.hidden_act}. "
"Only silu is supported for now.")
raise ValueError(
f"Unsupported activation: {config.hidden_act}. "
"Only silu is supported for now."
)
self.gate = ReplicatedLinear(config.hidden_size,
config.n_routed_experts,
bias=False,
quant_config=None,
prefix=f"{prefix}.gate")
self.gate = ReplicatedLinear(
config.hidden_size,
config.n_routed_experts,
bias=False,
quant_config=None,
prefix=f"{prefix}.gate",
)
if config.topk_method == "noaux_tc":
self.gate.e_score_correction_bias = (nn.Parameter(
torch.empty(config.n_routed_experts)))
self.gate.e_score_correction_bias = nn.Parameter(
torch.empty(config.n_routed_experts)
)
else:
self.gate.e_score_correction_bias = None
@@ -138,11 +160,11 @@ class Dots1MoE(nn.Module):
scoring_func=config.scoring_func,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias)
e_score_correction_bias=self.gate.e_score_correction_bias,
)
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 = Dots1MLP(
hidden_size=config.hidden_size,
intermediate_size=intermediate_size,
@@ -158,19 +180,18 @@ class Dots1MoE(nn.Module):
if self.n_shared_experts is not None:
shared_output = self.shared_experts(hidden_states)
router_logits, _ = self.gate(hidden_states)
final_hidden_states = self.experts(
hidden_states=hidden_states,
router_logits=router_logits) * self.routed_scaling_factor
final_hidden_states = (
self.experts(hidden_states=hidden_states, router_logits=router_logits)
* self.routed_scaling_factor
)
if shared_output is not None:
final_hidden_states = final_hidden_states + shared_output
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_dim)
class Dots1Attention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -200,8 +221,7 @@ class Dots1Attention(nn.Module):
# the KV heads across multiple tensor parallel GPUs.
assert tp_size % self.total_num_kv_heads == 0
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
self.head_dim = getattr(config, "head_dim",
hidden_size // self.total_num_heads)
self.head_dim = getattr(config, "head_dim", hidden_size // self.total_num_heads)
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
@@ -244,14 +264,15 @@ class Dots1Attention(nn.Module):
self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps)
def forward(self, positions: torch.Tensor,
hidden_states: torch.Tensor) -> torch.Tensor:
def forward(
self, positions: torch.Tensor, hidden_states: torch.Tensor
) -> torch.Tensor:
qkv, _ = self.qkv_proj(hidden_states)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
q = self.q_norm(q.reshape(-1, self.num_heads,
self.head_dim)).reshape(q.shape)
k = self.k_norm(k.reshape(-1, self.num_kv_heads,
self.head_dim)).reshape(k.shape)
q = self.q_norm(q.reshape(-1, self.num_heads, self.head_dim)).reshape(q.shape)
k = self.k_norm(k.reshape(-1, self.num_kv_heads, self.head_dim)).reshape(
k.shape
)
q, k = self.rotary_emb(positions, q, k)
attn_output = self.attn(q, k, v)
output, _ = self.o_proj(attn_output)
@@ -259,7 +280,6 @@ class Dots1Attention(nn.Module):
class Dots1DecoderLayer(nn.Module):
def __init__(
self,
config: Dots1Config,
@@ -272,9 +292,8 @@ class Dots1DecoderLayer(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)
layer_idx = int(prefix.split(sep='.')[-1])
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
layer_idx = int(prefix.split(sep=".")[-1])
self.layer_idx = layer_idx
self.self_attn = Dots1Attention(
@@ -289,12 +308,14 @@ class Dots1DecoderLayer(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 = Dots1MoE(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 = Dots1MoE(
config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
)
else:
self.mlp = Dots1MLP(
hidden_size=config.hidden_size,
@@ -303,10 +324,10 @@ class Dots1DecoderLayer(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
)
self.routed_scaling_factor = config.routed_scaling_factor
def forward(
@@ -319,19 +340,15 @@ class Dots1DecoderLayer(nn.Module):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
else:
hidden_states, residual = self.input_layernorm(
hidden_states, residual)
hidden_states = self.self_attn(positions=positions,
hidden_states=hidden_states)
hidden_states, residual = self.post_attention_layernorm(
hidden_states, residual)
hidden_states, residual = self.input_layernorm(hidden_states, residual)
hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
hidden_states = self.mlp(hidden_states)
return hidden_states, residual
@support_torch_compile
class Dots1Model(nn.Module):
fall_back_to_pt_during_load = False
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
@@ -350,7 +367,8 @@ class Dots1Model(nn.Module):
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens")
prefix=f"{prefix}.embed_tokens",
)
else:
self.embed_tokens = PPMissingLayer()
@@ -363,15 +381,16 @@ class Dots1Model(nn.Module):
cache_config=cache_config,
quant_config=quant_config,
),
prefix=f"{prefix}.layers")
prefix=f"{prefix}.layers",
)
if get_pp_group().is_last_rank:
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
else:
self.norm = PPMissingLayer()
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)
@@ -400,10 +419,9 @@ class Dots1Model(nn.Module):
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
@@ -412,10 +430,10 @@ class Dots1Model(nn.Module):
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts)
num_experts=self.config.n_routed_experts,
)
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 = [
("qkv_proj", "q_proj", "q"),
("qkv_proj", "k_proj", "k"),
@@ -430,10 +448,10 @@ class Dots1Model(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
if (("mlp.experts." in name) and name not in params_dict):
if ("mlp.experts." in name) and name not in params_dict:
continue
name = name.replace(weight_name, param_name)
if name.endswith(".bias") and name not in params_dict:
@@ -456,11 +474,13 @@ class Dots1Model(nn.Module):
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param,
loaded_weight,
name,
shard_id=shard_id,
expert_id=expert_id)
weight_loader(
param,
loaded_weight,
name,
shard_id=shard_id,
expert_id=expert_id,
)
break
else:
if name.endswith(".bias") and name not in params_dict:
@@ -471,15 +491,15 @@ class Dots1Model(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
class Dots1ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -498,19 +518,22 @@ class Dots1ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = Dots1Model(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = Dots1Model(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
if get_pp_group().is_last_rank:
self.lm_head = ParallelLMHead(config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(
prefix, "lm_head"))
self.lm_head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
else:
self.lm_head = PPMissingLayer()
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)
@@ -537,8 +560,7 @@ class Dots1ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
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)