dnrops
/
transformers
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

updates

This commit is contained in:
Arthur Zucker 2024-05-15 16:49:18 +02:00
parent eaaf34f303
commit 45f20f5641
3 changed files with 205 additions and 161 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/transformers/models/gemma/diff_gemma.py
View File

@ -27,8 +27,8 @@ GemmaConverter.register("GemmaRotaryEmbedding", LlamaRotaryEmbedding)
GemmaConverter.register("GemmaMLP", LlamaMLP)
GemmaAttention = GemmaConverter.register("GemmaAttention", LlamaAttention)
GemmaSdpaAttention = GemmaConverter.register("GemmaSdpaAttention", LlamaAttention)
GemmaFlashAttention2 = GemmaConverter.register("GemmaFlashAttention2", LlamaAttention)
GemmaSdpaAttention = GemmaConverter.register("GemmaSdpaAttention", LlamaSdpaAttention)
GemmaFlashAttention2 = GemmaConverter.register("GemmaFlashAttention2", LlamaFlashAttention2)
COHERE_ATTENTION_CLASSES = {"eager": GemmaAttention, "flash_attention_2": GemmaFlashAttention2, "sdpa": GemmaSdpaAttention}

359
src/transformers/models/gemma/modeling_gemma.py
View File

@ -226,69 +226,14 @@ class GemmaAttention(nn.Module):
return attn_output, attn_weights, past_key_value
class GemmaSdpaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
if layer_idx is None:
logger.warning_once(
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
"when creating this class."
)
self.attention_dropout = config.attention_dropout
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_key_value_heads = config.num_key_value_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.max_position_embeddings = config.max_position_embeddings
self.rope_theta = config.rope_theta
self.is_causal = True
if (self.head_dim * self.num_heads) != self.hidden_size:
raise ValueError(
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
f" and `num_heads`: {self.num_heads})."
)
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
self._init_rope()
def _init_rope(self):
if self.config.rope_scaling is None:
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
else:
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = GemmaLinearScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
elif scaling_type == "dynamic":
self.rotary_emb = GemmaDynamicNTKScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
else:
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
class GemmaSdpaAttention(GemmaAttention):
"""
Gemma attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
`GemmaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
SDPA API.
"""
# Adapted from GemmaAttention.forward
def forward(
self,
hidden_states: torch.Tensor,
@ -299,6 +244,22 @@ class GemmaSdpaAttention(nn.Module):
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if output_attentions:
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
logger.warning_once(
"GemmaModel is using GemmaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
)
return super().forward(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
)
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
@ -311,6 +272,7 @@ class GemmaSdpaAttention(nn.Module):
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
if past_key_value is not None:
# sin and cos are specific to RoPE models; cache_position needed for the static cache
cache_kwargs = {"cache_position": cache_position}
@ -319,146 +281,130 @@ class GemmaSdpaAttention(nn.Module):
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
causal_mask = attention_mask
if attention_mask is not None:
causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
if attention_mask is not None: # no matter the length, we just slice it
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
attn_weights = attn_weights + causal_mask
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
# Reference: https://github.com/pytorch/pytorch/issues/112577.
if query_states.device.type == "cuda" and causal_mask is not None:
query_states = query_states.contiguous()
key_states = key_states.contiguous()
value_states = value_states.contiguous()
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_output = torch.matmul(attn_weights, value_states)
# We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
# inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
is_causal = True if causal_mask is None and q_len > 1 else False
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=causal_mask,
dropout_p=self.attention_dropout if self.training else 0.0,
is_causal=is_causal,
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.view(bsz, q_len, self.hidden_size)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, None, past_key_value
return attn_output, attn_weights, past_key_value
class GemmaFlashAttention2(GemmaAttention):
"""
Gemma flash attention module. This module inherits from `GemmaAttention` as the weights of the module stays
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
flash attention and deal with padding tokens in case the input contains any of them.
"""
class GemmaFlashAttention2(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
if layer_idx is None:
logger.warning_once(
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
"when creating this class."
)
self.attention_dropout = config.attention_dropout
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_key_value_heads = config.num_key_value_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.max_position_embeddings = config.max_position_embeddings
self.rope_theta = config.rope_theta
self.is_causal = True
if (self.head_dim * self.num_heads) != self.hidden_size:
raise ValueError(
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
f" and `num_heads`: {self.num_heads})."
)
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
self._init_rope()
def _init_rope(self):
if self.config.rope_scaling is None:
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
else:
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = GemmaLinearScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
elif scaling_type == "dynamic":
self.rotary_emb = GemmaDynamicNTKScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
else:
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if isinstance(past_key_value, StaticCache):
raise ValueError(
"`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
"make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
)
output_attentions = False
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
# Flash attention requires the input to have the shape
# batch_size x seq_length x head_dim x hidden_dim
# therefore we just need to keep the original shape
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
query_states, key_states = self.rotary_emb(query_states, key_states, position_ids)
if past_key_value is not None:
# sin and cos are specific to RoPE models; cache_position needed for the static cache
cache_kwargs = {"cache_position": cache_position}
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
# TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
# to be able to avoid many of these transpose/reshape/view.
query_states = query_states.transpose(1, 2)
key_states = key_states.transpose(1, 2)
value_states = value_states.transpose(1, 2)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
dropout_rate = self.attention_dropout if self.training else 0.0
if attention_mask is not None: # no matter the length, we just slice it
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
attn_weights = attn_weights + causal_mask
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
# therefore the input hidden states gets silently casted in float32. Hence, we need
# cast them back in the correct dtype just to be sure everything works as expected.
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
# in fp32. (GemmaRMSNorm handles it correctly)
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_output = torch.matmul(attn_weights, value_states)
input_dtype = query_states.dtype
if input_dtype == torch.float32:
if torch.is_autocast_enabled():
target_dtype = torch.get_autocast_gpu_dtype()
# Handle the case where the model is quantized
elif hasattr(self.config, "_pre_quantization_dtype"):
target_dtype = self.config._pre_quantization_dtype
else:
target_dtype = self.q_proj.weight.dtype
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
logger.warning_once(
f"The input hidden states seems to be silently casted in float32, this might be related to"
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
f" {target_dtype}."
)
attn_output = attn_output.transpose(1, 2).contiguous()
query_states = query_states.to(target_dtype)
key_states = key_states.to(target_dtype)
value_states = value_states.to(target_dtype)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self._flash_attention_forward(
query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
attn_output = self.o_proj(attn_output)
if not output_attentions:
@ -466,6 +412,103 @@ class GemmaFlashAttention2(nn.Module):
return attn_output, attn_weights, past_key_value
def _flash_attention_forward(
self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
):
"""
Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
first unpad the input, then computes the attention scores and pad the final attention scores.
Args:
query_states (`torch.Tensor`):
Input query states to be passed to Flash Attention API
key_states (`torch.Tensor`):
Input key states to be passed to Flash Attention API
value_states (`torch.Tensor`):
Input value states to be passed to Flash Attention API
attention_mask (`torch.Tensor`):
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
position of padding tokens and 1 for the position of non-padding tokens.
dropout (`float`):
Attention dropout
softmax_scale (`float`, *optional*):
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
"""
if not self._flash_attn_uses_top_left_mask:
causal = self.is_causal
else:
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in GemmaFlashAttention2 __init__.
causal = self.is_causal and query_length != 1
# Contains at least one padding token in the sequence
if attention_mask is not None:
batch_size = query_states.shape[0]
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
query_states, key_states, value_states, attention_mask, query_length
)
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
attn_output_unpad = flash_attn_varlen_func(
query_states,
key_states,
value_states,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
max_seqlen_q=max_seqlen_in_batch_q,
max_seqlen_k=max_seqlen_in_batch_k,
dropout_p=dropout,
softmax_scale=softmax_scale,
causal=causal,
)
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
else:
attn_output = flash_attn_func(
query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
)
return attn_output
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
key_layer = index_first_axis(
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
)
value_layer = index_first_axis(
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
)
if query_length == kv_seq_len:
query_layer = index_first_axis(
query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
)
cu_seqlens_q = cu_seqlens_k
max_seqlen_in_batch_q = max_seqlen_in_batch_k
indices_q = indices_k
elif query_length == 1:
max_seqlen_in_batch_q = 1
cu_seqlens_q = torch.arange(
batch_size + 1, dtype=torch.int32, device=query_layer.device
) # There is a memcpy here, that is very bad.
indices_q = cu_seqlens_q[:-1]
query_layer = query_layer.squeeze(1)
else:
# The -q_len: slice assumes left padding.
attention_mask = attention_mask[:, -query_length:]
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
return (
query_layer,
key_layer,
value_layer,
indices_q,
(cu_seqlens_q, cu_seqlens_k),
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
)
COHERE_ATTENTION_CLASSES = {"eager": GemmaAttention, "flash_attention_2": GemmaFlashAttention2, "sdpa": GemmaSdpaAttention}

3
utils/convert_diff_model.py
View File

@ -89,6 +89,7 @@ def create_single_model_file(converter):
parent_identifier_camel = re.findall(r'[A-Z][a-z0-9]*', parent_class)[0]
child_identifier_camel = re.findall(r'[A-Z][a-z0-9]*', class_name)[0]
print(f"`{class_name}` -> `{parent_class}`")
child_function_set = parent_function_set.copy()
class_def = inspect.getsource(eval(class_name))
@ -98,8 +99,8 @@ def create_single_model_file(converter):
full_function = match.group()
if "def" in full_function:
function_name = full_function.split("def")[1].split("(")[0]
if (f"super()." in full_function or f"return super()." in full_function) and parent_identifier_camel != child_identifier_camel:
print(f"`{parent_identifier_camel}` `{child_identifier_camel}`")
replaced_function = replace_super_calls_in_method(full_function,
parent_function_set.get(function_name,
""),