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

Jump to bottom

Remove unnecessary unsqueeze - squeeze in rotary positional embedding #26162

Merged
merged 7 commits into from
Oct 6, 2023
Merged

Remove unnecessary unsqueeze - squeeze in rotary positional embedding #26162

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
21675a8
remove unnecessary unsqueeze-squeeze in llama
fxmarty Sep 14, 2023
768b8b0
correct other models
fxmarty Sep 16, 2023
4183f9f
fix
fxmarty Sep 16, 2023
e0bab59
revert gpt_neox_japanese
fxmarty Sep 16, 2023
5443100
Merge branch 'main' into remove-back-and-forth-unsqueeze-squeeze
fxmarty Oct 5, 2023
4f89883
fix copie
fxmarty Oct 5, 2023
69a4840
fix test
fxmarty Oct 5, 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
16 changes: 8 additions & 8 deletions src/transformers/models/deprecated/open_llama/modeling_open_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,17 +121,17 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)

def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)

return (
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.cos_cached[:seq_len].to(dtype=x.dtype),
self.sin_cached[:seq_len].to(dtype=x.dtype),
)


Expand All @@ -151,8 +151,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)


# Copied from transformers.models.llama.modeling_llama.LlamaDynamicNTKScalingRotaryEmbedding with Llama->OpenLlama
Expand All @@ -178,8 +178,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)


def rotate_half(x):
Expand Down
16 changes: 8 additions & 8 deletions src/transformers/models/falcon/modeling_falcon.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,8 +94,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
if dtype in [torch.float16, torch.bfloat16]:
emb = emb.float()

self.cos_cached = emb.cos()[None, :, :]
self.sin_cached = emb.sin()[None, :, :]
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()

self.cos_cached = self.cos_cached.type(dtype)
self.sin_cached = self.sin_cached.type(dtype)
Expand All @@ -107,8 +107,8 @@ def cos_sin(
if total_length > self.seq_len_cached:
self._set_cos_sin_cache(total_length, device, dtype)
# Gather cos, sin at the designated position ids
cos = self.cos_cached.squeeze(0)[position_ids] # [bs, seq_len, dim]
sin = self.sin_cached.squeeze(0)[position_ids] # [bs, seq_len, dim]
cos = self.cos_cached[position_ids] # [bs, seq_len, dim]
sin = self.sin_cached[position_ids] # [bs, seq_len, dim]
return cos, sin

def forward(self, query, key, past_key_values_length, position_ids):
Expand Down Expand Up @@ -155,8 +155,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
if dtype in [torch.float16, torch.bfloat16]:
emb = emb.float()

self.cos_cached = emb.cos()[None, :, :]
self.sin_cached = emb.sin()[None, :, :]
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()

self.cos_cached = self.cos_cached.type(dtype)
self.sin_cached = self.sin_cached.type(dtype)
Expand Down Expand Up @@ -189,8 +189,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
if dtype in [torch.float16, torch.bfloat16]:
emb = emb.float()

self.cos_cached = emb.cos()[None, :, :]
self.sin_cached = emb.sin()[None, :, :]
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()

self.cos_cached = self.cos_cached.type(dtype)
self.sin_cached = self.sin_cached.type(dtype)
Expand Down
20 changes: 9 additions & 11 deletions src/transformers/models/gpt_neox/modeling_gpt_neox.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -306,14 +306,14 @@ def _set_cos_sin_cache(self, seq_len, device):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.cos_cached = emb.cos()[None, None, :, :]
self.sin_cached = emb.sin()[None, None, :, :]
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()

def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len=seq_len, device=x.device)
return self.cos_cached[:seq_len, ...].to(x.device), self.sin_cached[:seq_len, ...].to(x.device)
return self.cos_cached[:seq_len].to(x.device), self.sin_cached[:seq_len].to(x.device)


class GPTNeoXLinearScalingRotaryEmbedding(GPTNeoXRotaryEmbedding):
Expand All @@ -331,8 +331,8 @@ def _set_cos_sin_cache(self, seq_len, device):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.cos_cached = emb.cos()[None, None, :, :]
self.sin_cached = emb.sin()[None, None, :, :]
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()


class GPTNeoXDynamicNTKScalingRotaryEmbedding(GPTNeoXRotaryEmbedding):
Expand All @@ -357,8 +357,8 @@ def _set_cos_sin_cache(self, seq_len, device):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.cos_cached = emb.cos()[None, None, :, :]
self.sin_cached = emb.sin()[None, None, :, :]
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()


def rotate_half(x):
Expand All @@ -369,10 +369,8 @@ def rotate_half(x):


def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
gather_indices = position_ids[:, None, :, None] # [bs, 1, seq_len, 1]
gather_indices = gather_indices.repeat(1, cos.shape[1], 1, cos.shape[3])
cos = torch.gather(cos.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
sin = torch.gather(sin.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
cos = cos[position_ids].unsqueeze(1) # [seq_len, dim] -> [batch_size, 1, seq_len, head_dim]
sin = sin[position_ids].unsqueeze(1)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
Expand Down
19 changes: 9 additions & 10 deletions src/transformers/models/idefics/modeling_idefics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -507,8 +507,8 @@ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
self.register_buffer("cos_cached", emb.cos(), persistent=False)
self.register_buffer("sin_cached", emb.sin(), persistent=False)

def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
Expand All @@ -519,11 +519,11 @@ def forward(self, x, seq_len=None):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
self.register_buffer("cos_cached", emb.cos(), persistent=False)
self.register_buffer("sin_cached", emb.sin(), persistent=False)
return (
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.cos_cached[:seq_len].to(dtype=x.dtype),
self.sin_cached[:seq_len].to(dtype=x.dtype),
)


Expand All @@ -534,11 +534,10 @@ def rotate_half(x):
return torch.cat((-x2, x1), dim=-1)


# Copied from transformers.models.gpt_neox.modeling_gpt_neox.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
gather_indices = position_ids[:, None, :, None] # [bs, 1, seq_len, 1]
gather_indices = gather_indices.repeat(1, cos.shape[1], 1, cos.shape[3])
cos = torch.gather(cos.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
sin = torch.gather(sin.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
cos = cos[position_ids].unsqueeze(1) # [seq_len, dim] -> [batch_size, 1, seq_len, head_dim]
sin = sin[position_ids].unsqueeze(1)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
Expand Down
24 changes: 11 additions & 13 deletions src/transformers/models/llama/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,17 +111,17 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)

def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)

return (
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.cos_cached[:seq_len].to(dtype=x.dtype),
self.sin_cached[:seq_len].to(dtype=x.dtype),
)


Expand All @@ -140,8 +140,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)


class LlamaDynamicNTKScalingRotaryEmbedding(LlamaRotaryEmbedding):
Expand All @@ -166,8 +166,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)


def rotate_half(x):
Expand All @@ -177,12 +177,10 @@ def rotate_half(x):
return torch.cat((-x2, x1), dim=-1)


# Copied from transformers.models.gpt_neox.modeling_gpt_neox.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [seq_len, dim] -> [batch_size, 1, seq_len, head_dim]
sin = sin[position_ids].unsqueeze(1)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
Expand Down
19 changes: 8 additions & 11 deletions src/transformers/models/persimmon/modeling_persimmon.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,17 +94,17 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)

def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)

return (
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.cos_cached[:seq_len].to(dtype=x.dtype),
self.sin_cached[:seq_len].to(dtype=x.dtype),
)


Expand All @@ -124,8 +124,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)


# Copied from transformers.models.llama.modeling_llama.LlamaDynamicNTKScalingRotaryEmbedding with Llama->Persimmon
Expand All @@ -151,8 +151,8 @@ def _set_cos_sin_cache(self, seq_len, device, dtype):
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)


# Copied from transformers.models.llama.modeling_llama.rotate_half
Expand All @@ -165,9 +165,6 @@ def rotate_half(x):

# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
q_embed = (q * cos) + (rotate_half(q) * sin)
Expand Down
Loading