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

209
vllm/model_executor/models/bamba.py
View File

@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Inference-only Bamba model."""
# Added by the IBM Team, 2024
from collections.abc import Iterable
from typing import Optional
@@ -16,29 +17,38 @@ from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateDtypeCalculator, MambaStateShapeCalculator)
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
)
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.sequence import IntermediateTensors
from .interfaces import (HasInnerState, IsHybrid, SupportsLoRA, SupportsPP,
SupportsQuant)
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .interfaces import HasInnerState, IsHybrid, SupportsLoRA, SupportsPP, SupportsQuant
from .utils import (
AutoWeightsLoader,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class BambaMLP(nn.Module):
def __init__(
self,
config: BambaConfig,
@@ -59,8 +69,10 @@ class BambaMLP(nn.Module):
quant_config=quant_config,
)
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.act_fn = SiluAndMul()
def forward(self, x):
@@ -71,38 +83,38 @@ class BambaMLP(nn.Module):
class BambaMixerDecoderLayer(nn.Module):
def __init__(self,
config: BambaConfig,
layer_idx: int,
model_config: Optional[ModelConfig] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "") -> None:
def __init__(
self,
config: BambaConfig,
layer_idx: int,
model_config: Optional[ModelConfig] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.config = config
self.mamba = MambaMixer2(hidden_size= config.hidden_size,
ssm_state_size = config.mamba_d_state,
conv_kernel_size = config.mamba_d_conv,
intermediate_size = config.mamba_expand *\
config.hidden_size,
use_conv_bias = config.mamba_conv_bias,
use_bias = config.mamba_proj_bias,
n_groups=config.mamba_n_groups,
num_heads=config.mamba_n_heads,
head_dim=config.mamba_d_head,
rms_norm_eps=config.rms_norm_eps,
activation=config.hidden_act,
model_config=model_config,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.mixer")
self.mamba = MambaMixer2(
hidden_size=config.hidden_size,
ssm_state_size=config.mamba_d_state,
conv_kernel_size=config.mamba_d_conv,
intermediate_size=config.mamba_expand * config.hidden_size,
use_conv_bias=config.mamba_conv_bias,
use_bias=config.mamba_proj_bias,
n_groups=config.mamba_n_groups,
num_heads=config.mamba_n_heads,
head_dim=config.mamba_d_head,
rms_norm_eps=config.rms_norm_eps,
activation=config.hidden_act,
model_config=model_config,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.mixer",
)
self.feed_forward = BambaMLP(config, quant_config=quant_config)
self.input_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.pre_ff_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.pre_ff_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def forward(
self,
@@ -114,8 +126,7 @@ class BambaMixerDecoderLayer(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)
output = torch.empty_like(hidden_states)
self.mamba(hidden_states, output)
@@ -126,7 +137,6 @@ class BambaMixerDecoderLayer(nn.Module):
class BambaAttentionDecoderLayer(nn.Module):
def __init__(
self,
config: BambaConfig,
@@ -139,8 +149,7 @@ class BambaAttentionDecoderLayer(nn.Module):
super().__init__()
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.hidden_size = config.hidden_size
tp_size = get_tensor_model_parallel_world_size()
self.total_num_heads = config.num_attention_heads
@@ -188,10 +197,12 @@ class BambaAttentionDecoderLayer(nn.Module):
bias=False,
quant_config=quant_config,
)
self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
config.hidden_size,
bias=False,
quant_config=quant_config)
self.o_proj = RowParallelLinear(
self.total_num_heads * self.head_dim,
config.hidden_size,
bias=False,
quant_config=quant_config,
)
self.attn = Attention(
self.num_heads,
@@ -203,10 +214,8 @@ class BambaAttentionDecoderLayer(nn.Module):
)
self.feed_forward = BambaMLP(config, quant_config=quant_config)
self.input_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.pre_ff_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.pre_ff_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def self_attention(
self,
@@ -233,29 +242,26 @@ class BambaAttentionDecoderLayer(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_attention(
positions=positions,
hidden_states=hidden_states,
)
# Fully Connected
hidden_states, residual = self.pre_ff_layernorm(
hidden_states, residual)
hidden_states, residual = self.pre_ff_layernorm(hidden_states, residual)
hidden_states = self.feed_forward(hidden_states)
return hidden_states, residual
ALL_DECODER_LAYER_TYPES = {
"attention": BambaAttentionDecoderLayer,
"mamba": BambaMixerDecoderLayer
"mamba": BambaMixerDecoderLayer,
}
@support_torch_compile
class BambaModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -266,8 +272,11 @@ class BambaModel(nn.Module):
lora_config = vllm_config.lora_config
self.config = 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
@@ -279,8 +288,7 @@ class BambaModel(nn.Module):
def get_layer(prefix: str):
layer_idx = int(prefix.rsplit(".", 1)[1])
layer_class = ALL_DECODER_LAYER_TYPES[
config.layers_block_type[layer_idx]]
layer_class = ALL_DECODER_LAYER_TYPES[config.layers_block_type[layer_idx]]
return layer_class(
config,
layer_idx,
@@ -291,13 +299,13 @@ class BambaModel(nn.Module):
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers")
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers"
)
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size
)
self.final_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
@@ -309,7 +317,6 @@ class BambaModel(nn.Module):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
@@ -330,15 +337,13 @@ class BambaModel(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.final_layernorm(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"),
@@ -383,22 +388,22 @@ class BambaModel(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
class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
IsHybrid, SupportsQuant):
class BambaForCausalLM(
nn.Module, HasInnerState, SupportsLoRA, SupportsPP, IsHybrid, SupportsQuant
):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
"k_proj",
"v_proj",
],
"gate_up_proj": ["up_proj", "down_proj"]
"gate_up_proj": ["up_proj", "down_proj"],
}
# LoRA specific attributes
@@ -413,7 +418,6 @@ class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
cls,
vllm_config: "VllmConfig",
) -> tuple[torch.dtype, torch.dtype]:
return MambaStateDtypeCalculator.mamba2_state_dtype(
vllm_config.model_config.dtype,
vllm_config.cache_config.mamba_cache_dtype,
@@ -460,8 +464,9 @@ class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
super().__init__()
self.config = config
self.scheduler_config = scheduler_config
self.model = BambaModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = BambaModel(
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
@@ -472,28 +477,33 @@ class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
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,
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)
def forward(self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
**kwargs):
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
**kwargs,
):
hidden_states = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states
@@ -504,7 +514,6 @@ class BambaForCausalLM(nn.Module, HasInnerState, 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)