TP/quantization/weight loading refactor part 2 - Refactor quantized linear logic and extend quantization support to all models (#1622)
Refactor the tensor parallelism, quantization, and weight-loading codes. Summary of the new features enabled by this PR: - **All models** are able to be quantized with AWQ and SqueezeLLM, and [soon GPTQ](https://github.com/vllm-project/vllm/pull/1580). - Model loading code became much simpler. - Support model parallelism for all MQA/GQA models when the number of key/value heads is smaller than the tensor parallel size.
This commit is contained in:
@@ -33,17 +33,19 @@ from vllm.transformers_utils.configs.yi import YiConfig
|
||||
|
||||
from vllm.model_executor.input_metadata import InputMetadata
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.attention import PagedAttentionWithRoPE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (LinearMethodBase,
|
||||
MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.quantized_linear import ParallelLinear
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
VocabParallelEmbedding, ParallelLMHead)
|
||||
from vllm.model_executor.parallel_utils.parallel_state import (
|
||||
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
|
||||
from vllm.model_executor.parallel_utils.layers import VocabParallelEmbedding
|
||||
from vllm.model_executor.quantization_utils import QuantizationConfig
|
||||
from vllm.model_executor.weight_utils import (
|
||||
convert_pyslice_to_tensor, hf_model_weights_iterator,
|
||||
load_tensor_parallel_weights, load_padded_tensor_parallel_vocab)
|
||||
get_tensor_model_parallel_world_size)
|
||||
from vllm.model_executor.weight_utils import (default_weight_loader,
|
||||
hf_model_weights_iterator)
|
||||
from vllm.sequence import SamplerOutput
|
||||
|
||||
KVCache = Tuple[torch.Tensor, torch.Tensor]
|
||||
@@ -56,19 +58,17 @@ class YiMLP(nn.Module):
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.gate_up_proj = ParallelLinear.column(hidden_size,
|
||||
2 * intermediate_size,
|
||||
bias=False,
|
||||
gather_output=False,
|
||||
quant_config=quant_config)
|
||||
self.down_proj = ParallelLinear.row(intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
input_is_parallel=True,
|
||||
quant_config=quant_config)
|
||||
self.gate_up_proj = MergedColumnParallelLinear(
|
||||
hidden_size, [intermediate_size] * 2,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
self.down_proj = RowParallelLinear(intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
if hidden_act != "silu":
|
||||
raise ValueError(f"Unsupported activation: {hidden_act}. "
|
||||
"Only silu is supported for now.")
|
||||
@@ -91,7 +91,7 @@ class YiAttention(nn.Module):
|
||||
rope_theta: float = 10000,
|
||||
rope_scaling: Optional[Dict[str, Any]] = None,
|
||||
max_position_embeddings: int = 8192,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
@@ -109,7 +109,6 @@ class YiAttention(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)
|
||||
num_kv_heads_replicas = max(1, tp_size // self.total_num_kv_heads)
|
||||
self.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
|
||||
@@ -117,21 +116,19 @@ class YiAttention(nn.Module):
|
||||
self.rope_theta = rope_theta
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
|
||||
self.qkv_proj = ParallelLinear.column(
|
||||
self.qkv_proj = QKVParallelLinear(
|
||||
hidden_size,
|
||||
(self.total_num_heads +
|
||||
2 * self.total_num_kv_heads * num_kv_heads_replicas) *
|
||||
self.head_dim,
|
||||
self.total_num_heads,
|
||||
self.total_num_kv_heads,
|
||||
bias=False,
|
||||
gather_output=False,
|
||||
quant_config=quant_config,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.o_proj = ParallelLinear.row(
|
||||
self.o_proj = RowParallelLinear(
|
||||
self.total_num_heads * self.head_dim,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
input_is_parallel=True,
|
||||
quant_config=quant_config,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.attn = PagedAttentionWithRoPE(
|
||||
self.num_heads,
|
||||
@@ -165,11 +162,10 @@ class YiDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: YiConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.hidden_size = config.hidden_size
|
||||
# Requires transformers > 4.32.0
|
||||
rope_theta = getattr(config, "rope_theta", 10000)
|
||||
rope_scaling = getattr(config, "rope_scaling", None)
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings",
|
||||
@@ -181,13 +177,13 @@ class YiDecoderLayer(nn.Module):
|
||||
rope_theta=rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
max_position_embeddings=max_position_embeddings,
|
||||
quant_config=quant_config,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.mlp = YiMLP(
|
||||
hidden_size=self.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=quant_config,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.ln1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.ln2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
@@ -225,20 +221,18 @@ class YiModel(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: YiConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.padding_idx = config.pad_token_id
|
||||
self.vocab_size = config.vocab_size
|
||||
|
||||
vocab_size = ((config.vocab_size + 63) // 64) * 64
|
||||
self.embed_tokens = VocabParallelEmbedding(
|
||||
vocab_size,
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
)
|
||||
self.layers = nn.ModuleList([
|
||||
YiDecoderLayer(config, quant_config)
|
||||
YiDecoderLayer(config, linear_method)
|
||||
for _ in range(config.num_hidden_layers)
|
||||
])
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
@@ -274,19 +268,13 @@ class YiForCausalLM(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: YiConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
linear_method: Optional[LinearMethodBase] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.quant_config = quant_config
|
||||
self.model = YiModel(config, quant_config)
|
||||
vocab_size = ((config.vocab_size + 63) // 64) * 64
|
||||
# NOTE: The LM head is not quantized.
|
||||
self.lm_head = ParallelLinear.column(config.hidden_size,
|
||||
vocab_size,
|
||||
bias=False,
|
||||
gather_output=False,
|
||||
quant_config=None)
|
||||
self.linear_method = linear_method
|
||||
self.model = YiModel(config, linear_method)
|
||||
self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
|
||||
self.sampler = Sampler(config.vocab_size)
|
||||
|
||||
def forward(
|
||||
@@ -303,124 +291,33 @@ class YiForCausalLM(nn.Module):
|
||||
input_metadata)
|
||||
return next_tokens
|
||||
|
||||
_column_parallel_layers = []
|
||||
_row_parallel_layers = ["o_proj", "down_proj"]
|
||||
|
||||
def load_weights(self,
|
||||
model_name_or_path: str,
|
||||
cache_dir: Optional[str] = None,
|
||||
load_format: str = "auto",
|
||||
revision: Optional[str] = None):
|
||||
if self.quant_config is None:
|
||||
col_weight_suffixes = ["weight"]
|
||||
row_weight_suffixes = ["weight"]
|
||||
else:
|
||||
col_weight_suffixes = (
|
||||
self.quant_config.get_col_parallel_tensor_names())
|
||||
row_weight_suffixes = (
|
||||
self.quant_config.get_row_parallel_tensor_names())
|
||||
|
||||
column_parallel_weights: List[str] = []
|
||||
for layer in self._column_parallel_layers:
|
||||
for suffix in col_weight_suffixes:
|
||||
column_parallel_weights.append(f"{layer}.{suffix}")
|
||||
row_parallel_weights: List[str] = []
|
||||
for layer in self._row_parallel_layers:
|
||||
for suffix in row_weight_suffixes:
|
||||
row_parallel_weights.append(f"{layer}.{suffix}")
|
||||
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
q_proj_shard_size = (self.config.hidden_size // tp_size)
|
||||
num_kv_heads_replicas = max(1,
|
||||
tp_size // self.config.num_key_value_heads)
|
||||
num_kv_heads_per_gpu = max(1,
|
||||
self.config.num_key_value_heads // tp_size)
|
||||
kv_proj_shard_size = (self.config.hidden_size //
|
||||
self.config.num_attention_heads *
|
||||
num_kv_heads_per_gpu)
|
||||
attention_weight_specs = [
|
||||
# (weight_name, shard_size, offset)
|
||||
("q_proj", q_proj_shard_size, 0),
|
||||
("k_proj", kv_proj_shard_size, q_proj_shard_size),
|
||||
("v_proj", kv_proj_shard_size,
|
||||
q_proj_shard_size + kv_proj_shard_size),
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("qkv_proj", "q_proj", "q"),
|
||||
("qkv_proj", "k_proj", "k"),
|
||||
("qkv_proj", "v_proj", "v"),
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
]
|
||||
state_dict = self.state_dict()
|
||||
|
||||
params_dict = dict(self.named_parameters())
|
||||
for name, loaded_weight in hf_model_weights_iterator(
|
||||
model_name_or_path, cache_dir, load_format, revision):
|
||||
if "rotary_emb.inv_freq" in name:
|
||||
continue
|
||||
|
||||
packed_dim = None
|
||||
is_transposed = False
|
||||
if self.quant_config is not None:
|
||||
packed_dim = self.quant_config.get_packed_dim(name)
|
||||
is_transposed = self.quant_config.is_transposed(name)
|
||||
if is_transposed:
|
||||
loaded_weight = convert_pyslice_to_tensor(loaded_weight)
|
||||
loaded_weight = loaded_weight.T
|
||||
|
||||
is_attention_weight = False
|
||||
for weight_name, shard_size, offset in attention_weight_specs:
|
||||
for (param_name, weight_name, shard_id) in stacked_params_mapping:
|
||||
if weight_name not in name:
|
||||
continue
|
||||
param = state_dict[name.replace(weight_name, "qkv_proj")]
|
||||
if is_transposed:
|
||||
param = param.T
|
||||
|
||||
if packed_dim is not None:
|
||||
shard_dim = 0 if not is_transposed else 1
|
||||
if packed_dim == shard_dim:
|
||||
shard_size //= self.quant_config.pack_factor
|
||||
offset //= self.quant_config.pack_factor
|
||||
|
||||
if weight_name in ["k_proj", "v_proj"]:
|
||||
shard_id = tp_rank // num_kv_heads_replicas
|
||||
else:
|
||||
shard_id = tp_rank
|
||||
loaded_weight = loaded_weight[shard_size *
|
||||
shard_id:shard_size *
|
||||
(shard_id + 1)]
|
||||
param_slice = param.data[offset:offset + shard_size]
|
||||
assert param_slice.shape == loaded_weight.shape
|
||||
|
||||
param_slice.copy_(loaded_weight)
|
||||
is_attention_weight = True
|
||||
param = params_dict[name.replace(weight_name, param_name)]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
break
|
||||
if is_attention_weight:
|
||||
continue
|
||||
|
||||
is_gate_up_weight = False
|
||||
for stride_id, weight_name in enumerate(["gate_proj", "up_proj"]):
|
||||
if weight_name not in name:
|
||||
continue
|
||||
param = state_dict[name.replace(weight_name, "gate_up_proj")]
|
||||
if is_transposed:
|
||||
param = param.T
|
||||
|
||||
shard_size = param.shape[0] // 2
|
||||
loaded_weight = loaded_weight[shard_size * tp_rank:shard_size *
|
||||
(tp_rank + 1)]
|
||||
param_slice = param.data[shard_size * stride_id:shard_size *
|
||||
(stride_id + 1)]
|
||||
assert param_slice.shape == loaded_weight.shape
|
||||
param_slice.copy_(loaded_weight)
|
||||
is_gate_up_weight = True
|
||||
break
|
||||
if is_gate_up_weight:
|
||||
continue
|
||||
|
||||
param = state_dict[name]
|
||||
if is_transposed:
|
||||
param = param.T
|
||||
|
||||
if "embed_tokens" in name or "lm_head" in name:
|
||||
load_padded_tensor_parallel_vocab(param, loaded_weight,
|
||||
tp_rank)
|
||||
continue
|
||||
|
||||
load_tensor_parallel_weights(param, loaded_weight, name,
|
||||
column_parallel_weights,
|
||||
row_parallel_weights, tp_rank)
|
||||
else:
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param, loaded_weight)
|
||||
|
||||
Reference in New Issue
Block a user