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

[Quality] Add code formatter and linter (#326)

Browse Source
This commit is contained in:
Zhuohan Li
2023-07-03 11:31:55 -07:00
committed by GitHub
parent 0ffded812a
commit d6fa1be3a8
47 changed files with 1547 additions and 617 deletions
Download Patch File Download Diff File Expand all files Collapse all files

104
vllm/model_executor/models/gpt_bigcode.py
View File

@@ -1,5 +1,6 @@
# coding=utf-8
# Adapted from https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt2/modeling_gpt2.py
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt2/modeling_gpt2.py
# Copyright 2023 The vLLM team.
# Copyright 2023 CTranslate2, and Michael Feil
# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
@@ -49,19 +50,25 @@ class GPTBigCodeAttention(nn.Module):
super().__init__()
self.hidden_size = config.hidden_size
total_num_heads = config.num_attention_heads
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
tensor_model_parallel_world_size = (
get_tensor_model_parallel_world_size())
assert total_num_heads % tensor_model_parallel_world_size == 0
self.num_heads = total_num_heads // tensor_model_parallel_world_size
self.head_dim = self.hidden_size // total_num_heads
self.scale = self.head_dim ** -0.5
self.scale = self.head_dim**-0.5
self.c_attn = ColumnParallelLinear(self.hidden_size, 3 * self.hidden_size,
bias=True, gather_output=False,
self.c_attn = ColumnParallelLinear(self.hidden_size,
3 * self.hidden_size,
bias=True,
gather_output=False,
perform_initialization=False)
self.c_proj = RowParallelLinear(self.hidden_size, self.hidden_size,
bias=True, input_is_parallel=True,
self.c_proj = RowParallelLinear(self.hidden_size,
self.hidden_size,
bias=True,
input_is_parallel=True,
perform_initialization=False)
self.attn = PagedAttention(self.num_heads, self.head_dim,
self.attn = PagedAttention(self.num_heads,
self.head_dim,
scale=self.scale)
def forward(
@@ -74,8 +81,8 @@ class GPTBigCodeAttention(nn.Module):
qkv, _ = self.c_attn(hidden_states)
q, k, v = qkv.chunk(chunks=3, dim=-1)
key_cache, value_cache = kv_cache
attn_output = self.attn(
q, k, v, key_cache, value_cache, input_metadata, cache_event)
attn_output = self.attn(q, k, v, key_cache, value_cache,
input_metadata, cache_event)
attn_output, _ = self.c_proj(attn_output)
return attn_output
@@ -89,11 +96,15 @@ class GPTBigMLP(nn.Module):
):
super().__init__()
hidden_size = config.hidden_size
self.c_fc = ColumnParallelLinear(hidden_size, intermediate_size,
bias=True, gather_output=False,
self.c_fc = ColumnParallelLinear(hidden_size,
intermediate_size,
bias=True,
gather_output=False,
perform_initialization=False)
self.c_proj = RowParallelLinear(intermediate_size, hidden_size,
bias=True, input_is_parallel=True,
self.c_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=True,
input_is_parallel=True,
perform_initialization=False)
self.act = get_act_fn(config.activation_function)
@@ -109,7 +120,8 @@ class GPTBigCodeBlock(nn.Module):
def __init__(self, config: GPTBigCodeConfig):
super().__init__()
hidden_size = config.hidden_size
inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
inner_dim = (config.n_inner if config.n_inner is not None else 4 *
hidden_size)
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.attn = GPTBigCodeAttention(config)
@@ -147,7 +159,7 @@ class GPTBigCodeModel(nn.Module):
def __init__(self, config: GPTBigCodeConfig):
super().__init__()
self.config = config
assert config.add_cross_attention == False
assert not config.add_cross_attention
self.embed_dim = config.hidden_size
@@ -181,8 +193,8 @@ class GPTBigCodeModel(nn.Module):
else:
cache_event = cache_events[i]
layer = self.h[i]
hidden_states = layer(
hidden_states, kv_caches[i], input_metadata, cache_event)
hidden_states = layer(hidden_states, kv_caches[i], input_metadata,
cache_event)
hidden_states = self.ln_f(hidden_states)
return hidden_states
@@ -207,24 +219,26 @@ class GPTBigCodeForCausalLM(nn.Module):
input_metadata: InputMetadata,
cache_events: Optional[List[torch.cuda.Event]],
) -> Dict[int, SequenceOutputs]:
hidden_states = self.transformer(
input_ids, positions, kv_caches, input_metadata, cache_events)
next_tokens = self.sampler(
self.lm_head_weight, hidden_states, input_metadata)
hidden_states = self.transformer(input_ids, positions, kv_caches,
input_metadata, cache_events)
next_tokens = self.sampler(self.lm_head_weight, hidden_states,
input_metadata)
return next_tokens
_column_parallel_weights = ["wte.weight", "c_fc.weight", "c_fc.bias"]
_row_parallel_weights = ["c_proj.weight"]
def load_weights(self, model_name_or_path: str,
def load_weights(self,
model_name_or_path: str,
cache_dir: Optional[str] = None,
use_np_cache: bool = False):
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
tensor_model_parallel_world_size = (
get_tensor_model_parallel_world_size())
tensor_model_parallel_rank = get_tensor_model_parallel_rank()
state_dict = self.state_dict()
for name, loaded_weight in hf_model_weights_iterator(
model_name_or_path, cache_dir, use_np_cache):
model_name_or_path, cache_dir, use_np_cache):
if "lm_head.weight" in name:
# GPT-2 ties the weights of the embedding layer and the final
# linear layer.
@@ -241,9 +255,11 @@ class GPTBigCodeForCausalLM(nn.Module):
if name == "transformer.wte.weight":
# Consider padding in the vocab size.
padded_vocab_size = param.shape[0] * tensor_model_parallel_world_size
padded_vocab_size = param.shape[
0] * tensor_model_parallel_world_size
num_extra_rows = padded_vocab_size - self.config.vocab_size
extra_rows = torch.empty(num_extra_rows, loaded_weight.shape[1])
extra_rows = torch.empty(num_extra_rows,
loaded_weight.shape[1])
extra_rows = extra_rows.to(loaded_weight)
loaded_weight = torch.cat([loaded_weight, extra_rows], dim=0)
@@ -258,25 +274,31 @@ class GPTBigCodeForCausalLM(nn.Module):
qkv_array = qkv_array.numpy()
dims_q = n_head * head_dim
q, k, v = np.split(qkv_array, (dims_q, dims_q + head_dim), axis=0)
# q is fine, but k & v have not replicated shape along the first axis
# as long as MQA is not nativly supported, increase memory and replicated
# (head_dim, hidden_dim) to (n_heads * head_dim, hidden_dim)
# pylint: disable=unbalanced-tuple-unpacking
q, k, v = np.split(qkv_array, (dims_q, dims_q + head_dim),
axis=0)
# q is fine, but k & v have not replicated shape along the first
# axis as long as MQA is not nativly supported, increase memory
# and replicated (head_dim, hidden_dim) to
# (n_heads * head_dim, hidden_dim)
if k.ndim == 2 and v.ndim == 2:
replication = (n_head, 1) # weights
else:
replication = n_head # biases
# replicate n_head times for q, v
k, v = np.tile(k, replication), np.tile(v, replication)
# concat q, k, v along the first axis (n_heads * head_dim, hidden_dim)
# concat q, k, v along the first axis
# (n_heads * head_dim, hidden_dim)
# to (3 * n_heads * head_dim, hidden_dim)
qkv_array = np.concatenate((q, k, v), axis=0)
return torch.from_numpy(qkv_array)
# For the fused QKV linear layer, manually shard the weights.
if "c_attn" in name:
# GPT-2's fused QKV has the shape of [3 * num_heads * head_size, hidden_size].
# When tensor parallelism is used, we shard the weights along the head dimension.
# GPT-2's fused QKV has the shape of
# [3 * num_heads * head_size, hidden_size].
# When tensor parallelism is used, we shard the weights along
# the head dimension.
total_num_heads = self.config.num_attention_heads
hidden_size = self.config.hidden_size
head_size = hidden_size // total_num_heads
@@ -285,13 +307,19 @@ class GPTBigCodeForCausalLM(nn.Module):
head_end = (tensor_model_parallel_rank + 1) * num_heads
if name.endswith(".weight"):
loaded_weight = _expand_mqa_mha(loaded_weight, n_head=total_num_heads, head_dim=head_size)
loaded_weight = loaded_weight.view(3, total_num_heads, head_size, hidden_size)
loaded_weight = _expand_mqa_mha(loaded_weight,
n_head=total_num_heads,
head_dim=head_size)
loaded_weight = loaded_weight.view(3, total_num_heads,
head_size, hidden_size)
loaded_weight = loaded_weight[:, head_start:head_end, :, :]
loaded_weight = loaded_weight.reshape(-1, hidden_size)
elif name.endswith(".bias"):
loaded_weight = _expand_mqa_mha(loaded_weight, n_head=total_num_heads, head_dim=head_size)
loaded_weight = loaded_weight.view(3, total_num_heads, head_size)
loaded_weight = _expand_mqa_mha(loaded_weight,
n_head=total_num_heads,
head_dim=head_size)
loaded_weight = loaded_weight.view(3, total_num_heads,
head_size)
loaded_weight = loaded_weight[:, head_start:head_end, :]
loaded_weight = loaded_weight.reshape(-1)
else: