Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add fused linear for the LLAMA MLP block and multi-head attention block #6425

Merged
merged 4 commits into from
Jul 26, 2023
Merged

Add fused linear for the LLAMA MLP block and multi-head attention block #6425

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0c1a534
Add fused linear for the LLAMA MLP block and multi-head attention block
littsk Jul 17, 2023
178e63b
Refactor the config name 'fuse_attn_qkv' to 'fuse_attention_qkv' for …
littsk Jul 19, 2023
6412892
Added a switch for fuse_mlp_linear and improved the organization of t…
littsk Jul 24, 2023
394608f
add tensor parallel mappings for fused linears
littsk Jul 26, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions llm/llama/run_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,8 +133,8 @@ class ModelArguments:
metadata={"help": "llama, use_fused_rms_norm"},
)
fuse_attention_qkv: bool = field(
default=True,
metadata={"help": "gpt, fuse_attention_qkv"},
default=False,
metadata={"help": "whether to fuse attention qkv"},
)
recompute_granularity: str = field(
default="full",
Expand Down
2 changes: 2 additions & 0 deletions paddlenlp/transformers/llama/configuration.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -208,6 +208,7 @@ def __init__(
use_cache=True,
use_recompute=False,
recompute_granularity="full",
fuse_attention_qkv=False,
use_flash_attention=False,
use_fused_rms_norm=False,
tensor_parallel_output=True,
Expand All @@ -230,6 +231,7 @@ def __init__(
self.use_cache = use_cache
self.use_recompute = use_recompute
self.recompute_granularity = recompute_granularity
self.fuse_attention_qkv = fuse_attention_qkv
self.use_flash_attention = use_flash_attention
self.use_fused_rms_norm = use_fused_rms_norm
self.tensor_parallel_output = tensor_parallel_output
Expand Down
112 changes: 67 additions & 45 deletions paddlenlp/transformers/llama/modeling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -316,11 +316,13 @@ def __init__(self, config):
super().__init__()
self.hidden_size = config.hidden_size
self.intermediate_size = config.intermediate_size
self.tensor_parallel_degree = config.tensor_parallel_degree

if config.tensor_parallel_degree > 1:
self.gate_proj = mpu.ColumnParallelLinear(
# 为了减少张量并行的通信量，将两个linear合并成一个
self.gate_up_fused_proj = mpu.ColumnParallelLinear(
littsk marked this conversation as resolved.
Show resolved Hide resolved
sijunhe marked this conversation as resolved.
Show resolved Hide resolved
self.hidden_size,
self.intermediate_size,
self.intermediate_size * 2,
gather_output=False,
has_bias=False,
)
Expand All @@ -330,19 +332,18 @@ def __init__(self, config):
input_is_parallel=True,
has_bias=False,
)
self.up_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.intermediate_size,
gather_output=False,
has_bias=False,
)
else:
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias_attr=False)
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias_attr=False)
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias_attr=False)

def forward(self, x):
return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
if self.tensor_parallel_degree > 1:
gate_out, up_out = paddle.chunk(self.gate_up_fused_proj(x), chunks=2, axis=-1)
littsk marked this conversation as resolved.
Show resolved Hide resolved
out = self.down_proj(F.silu(gate_out) * up_out)
else:
out = self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
return out


class LlamaAttention(nn.Layer):
Expand All @@ -354,47 +355,63 @@ def __init__(self, config):
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.max_position_embeddings = config.max_position_embeddings
self.fuse_attention_qkv = config.fuse_attention_qkv
if config.tensor_parallel_degree > 1:
assert (
self.num_heads % config.tensor_parallel_degree == 0
), f"num_heads: {self.num_heads}, tensor_parallel_degree: {config.tensor_parallel_degree}"
self.num_heads = self.num_heads // config.tensor_parallel_degree

if config.tensor_parallel_degree > 1:
self.q_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.hidden_size,
has_bias=False,
gather_output=False,
)
self.k_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.hidden_size,
has_bias=False,
gather_output=False,
)
self.v_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.hidden_size,
has_bias=False,
gather_output=False,
)
if self.fuse_attention_qkv:
self.qkv_proj = mpu.ColumnParallelLinear(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

预训练参数加载转换，看看能不能根据fuse 搞成自动的。

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

请问大佬，具体是啥意思呀

self.hidden_size,
3 * self.hidden_size,
has_bias=False,
gather_output=False,
)
else:
self.q_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.hidden_size,
has_bias=False,
gather_output=False,
)
self.k_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.hidden_size,
has_bias=False,
gather_output=False,
)
self.v_proj = mpu.ColumnParallelLinear(
self.hidden_size,
self.hidden_size,
has_bias=False,
gather_output=False,
)
else:
self.q_proj = nn.Linear(
self.hidden_size,
self.hidden_size,
bias_attr=False,
)
self.k_proj = nn.Linear(
self.hidden_size,
self.hidden_size,
bias_attr=False,
)
self.v_proj = nn.Linear(
self.hidden_size,
self.hidden_size,
bias_attr=False,
)
if self.fuse_attention_qkv:
self.qkv_proj = nn.Linear(
self.hidden_size,
3 * self.hidden_size,
bias_attr=False,
)
else:
self.q_proj = nn.Linear(
self.hidden_size,
self.hidden_size,
bias_attr=False,
)
self.k_proj = nn.Linear(
self.hidden_size,
self.hidden_size,
bias_attr=False,
)
self.v_proj = nn.Linear(
self.hidden_size,
self.hidden_size,
bias_attr=False,
)
littsk marked this conversation as resolved.
Show resolved Hide resolved

if config.tensor_parallel_degree > 1:
self.o_proj = mpu.RowParallelLinear(
Expand Down Expand Up @@ -422,9 +439,14 @@ def forward(
) -> Tuple[paddle.Tensor, Optional[paddle.Tensor], Optional[Tuple[paddle.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
bsz, q_len, _ = hidden_states.shape
query_states = self.q_proj(hidden_states).reshape(shape=[bsz, q_len, self.num_heads, self.head_dim])
key_states = self.k_proj(hidden_states).reshape(shape=[bsz, q_len, self.num_heads, self.head_dim])
value_states = self.v_proj(hidden_states).reshape(shape=[bsz, q_len, self.num_heads, self.head_dim])
if self.fuse_attention_qkv:
mix_layer = self.qkv_proj(hidden_states)
mix_layer = paddle.reshape_(mix_layer, [0, 0, self.num_heads, 3 * self.head_dim])
query_states, key_states, value_states = paddle.split(mix_layer, num_or_sections=3, axis=-1)
else:
query_states = self.q_proj(hidden_states).reshape(shape=[bsz, q_len, self.num_heads, self.head_dim])
key_states = self.k_proj(hidden_states).reshape(shape=[bsz, q_len, self.num_heads, self.head_dim])
value_states = self.v_proj(hidden_states).reshape(shape=[bsz, q_len, self.num_heads, self.head_dim])

kv_seq_len = key_states.shape[-3]
offset = 0
Expand Down