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

156
vllm/model_executor/models/phimoe.py
View File

@@ -23,6 +23,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only PhiMoE model."""
from collections.abc import Iterable
from itertools import islice
from typing import Optional, Union
@@ -36,26 +37,36 @@ 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_world_size
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import (QKVParallelLinear,
ReplicatedLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
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, 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, 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 PhiMoEConfig(PretrainedConfig):
model_type = "phimoe"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -128,7 +139,6 @@ class PhiMoEConfig(PretrainedConfig):
class mp(torch.autograd.Function):
@staticmethod
def forward(
ctx,
@@ -173,8 +183,9 @@ def sparsemixer(scores, jitter_eps=0.01):
# compute mask for sparsity
mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True)
factor = scores.abs().clamp(min=mask_logits_threshold)
mask_logits_threshold = ((mask_logits_threshold - scores) /
factor) > (2 * jitter_eps)
mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (
2 * jitter_eps
)
# apply mask
masked_gates = scores.masked_fill(mask_logits_threshold, float("-inf"))
@@ -195,24 +206,21 @@ def sparsemixer(scores, jitter_eps=0.01):
)
with torch.no_grad():
# compute mask for sparsity
mask_logits_threshold, max_ind = masked_scores.max(dim=-1,
keepdim=True)
mask_logits_threshold, max_ind = masked_scores.max(dim=-1, keepdim=True)
factor = scores.abs().clamp(min=mask_logits_threshold)
mask_logits_threshold = ((mask_logits_threshold - scores) /
factor) > (2 * jitter_eps)
mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (
2 * jitter_eps
)
# apply mask
masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold,
float("-inf"))
masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold, float("-inf"))
selected_experts_top2 = max_ind
# compute scores for gradients
masked_gates_top2 = torch.softmax(masked_gates_top2, dim=-1)
multiplier_top2 = masked_gates_top2.gather(dim=-1,
index=selected_experts_top2)
multiplier_top2 = masked_gates_top2.gather(dim=-1, index=selected_experts_top2)
multiplier = torch.concat((multiplier, multiplier_top2), dim=-1)
selected_experts = torch.concat((selected_experts, selected_experts_top2),
dim=-1)
selected_experts = torch.concat((selected_experts, selected_experts_top2), dim=-1)
return (
multiplier,
@@ -226,8 +234,7 @@ def phimoe_routing_function(
topk: int,
renormalize: bool,
):
assert hidden_states.shape[0] == gating_output.shape[0], (
"Number of tokens mismatch")
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
assert topk == 2, "Only top-2 routing is supported"
assert renormalize is False, "Renormalization is not supported"
@@ -278,7 +285,8 @@ class PhiMoE(nn.Module):
quant_config=quant_config,
tp_size=tp_size,
custom_routing_function=phimoe_routing_function,
prefix=f"{prefix}.experts")
prefix=f"{prefix}.experts",
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
# NOTE: hidden_states can have either 1D or 2D shape.
@@ -291,7 +299,6 @@ class PhiMoE(nn.Module):
class PhiMoEAttention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -376,7 +383,6 @@ class PhiMoEAttention(nn.Module):
class PhiMoEDecoderLayer(nn.Module):
def __init__(
self,
config: PhiMoEConfig,
@@ -393,8 +399,9 @@ class PhiMoEDecoderLayer(nn.Module):
num_heads=config.num_attention_heads,
max_position=config.max_position_embeddings,
num_kv_heads=config.num_key_value_heads,
head_dim=getattr(config, "head_dim",
self.hidden_size // config.num_attention_heads),
head_dim=getattr(
config, "head_dim", self.hidden_size // config.num_attention_heads
),
rope_theta=rope_theta,
cache_config=cache_config,
quant_config=quant_config,
@@ -409,12 +416,12 @@ class PhiMoEDecoderLayer(nn.Module):
quant_config=quant_config,
prefix=f"{prefix}.block_sparse_moe",
)
self.input_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.rms_norm_eps,
elementwise_affine=True)
self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.rms_norm_eps,
elementwise_affine=True)
self.input_layernorm = nn.LayerNorm(
config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True
)
self.post_attention_layernorm = nn.LayerNorm(
config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True
)
def forward(
self,
@@ -444,7 +451,6 @@ class PhiMoEDecoderLayer(nn.Module):
@support_torch_compile
class PhiMoEModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -453,8 +459,11 @@ class PhiMoEModel(nn.Module):
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_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.config = config
@@ -468,15 +477,17 @@ class PhiMoEModel(nn.Module):
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: PhiMoEDecoderLayer(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers")
self.norm = nn.LayerNorm(config.hidden_size,
eps=config.rms_norm_eps,
elementwise_affine=True)
config, cache_config, quant_config, prefix=prefix
),
prefix=f"{prefix}.layers",
)
self.norm = nn.LayerNorm(
config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True
)
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)
@@ -507,10 +518,9 @@ class PhiMoEModel(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)
return hidden_states
@@ -523,8 +533,7 @@ class PhiMoEModel(nn.Module):
num_experts=self.config.num_local_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 = [
# (param_name, shard_name, shard_id)
("qkv_proj", "q_proj", "q"),
@@ -536,14 +545,15 @@ class PhiMoEModel(nn.Module):
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
for name, loaded_weight in weights:
if (self.quant_config is not None and
(scale_name := self.quant_config.get_cache_scale(name))):
if self.quant_config is not None and (
scale_name := self.quant_config.get_cache_scale(name)
):
# Loading kv cache quantization scales
param = params_dict[scale_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
loaded_weight[0])
weight_loader = getattr(param, "weight_loader", default_weight_loader)
loaded_weight = (
loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
)
weight_loader(param, loaded_weight)
loaded_params.add(scale_name)
continue
@@ -594,8 +604,9 @@ class PhiMoEModel(nn.Module):
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
@@ -627,8 +638,9 @@ class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
self.lora_config = lora_config
self.quant_config = vllm_config.quant_config
self.model = PhiMoEModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = PhiMoEModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
self.unpadded_vocab_size = config.vocab_size
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
@@ -640,16 +652,20 @@ class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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=None,
bias=True,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, 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)
@@ -661,16 +677,16 @@ class PhiMoEForCausalLM(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(self, hidden_states: torch.Tensor) -> torch.Tensor:
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)