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

updates

This commit is contained in:
Arthur Zucker 2024-05-10 16:41:43 +02:00
parent e467d2fede
commit 1632e0f4bd
7 changed files with 533 additions and 2646 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

740
result.py
View File

@ -1,28 +1,26 @@
from transformers.models.llama.modeling_llama import *
import torch.nn as nn
from transformers import CohereConfig
from transformers.utils import ModelConverter
CohereConverter = ModelConverter(__file__)
# now should the cohere converted be added to all model converters?
GemmaConverter = ModelConverter(__file__)
class CohereLayerNorm(nn.Module):
def __init__(self, hidden_size=None, eps=1e-5, bias=False):
"""The hidden size can be a tuple or an int. The tuple is used for QKNorm to normalize across head_dim"""
class GemmaRMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
self.eps = eps
self.weight = nn.Parameter(torch.zeros(dim))
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
hidden_states = hidden_states.to(torch.float32)
mean = hidden_states.mean(-1, keepdim=True)
variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
hidden_states = (hidden_states - mean) * torch.rsqrt(variance + self.variance_epsilon)
hidden_states = self.weight.to(torch.float32) * hidden_states
return hidden_states.to(input_dtype)
def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
class CohereRotaryEmbedding(nn.Module):
def forward(self, x):
output = self._norm(x.float())
# Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
# See https://github.com/huggingface/transformers/pull/29402
output = output * (1.0 + self.weight.float())
return output.type_as(x)
class GemmaRotaryEmbedding(nn.Module):
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
super().__init__()
self.scaling_factor = scaling_factor
@ -33,7 +31,7 @@ class CohereRotaryEmbedding(nn.Module):
self.register_buffer("inv_freq", inv_freq, persistent=False)
@torch.no_grad()
def comput_cos_sin(self, x, position_ids):
def compute_cos_sin(self, x, position_ids):
# x: [bs, num_attention_heads, seq_len, head_size]
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
position_ids_expanded = position_ids[:, None, :].float()
@ -44,64 +42,59 @@ class CohereRotaryEmbedding(nn.Module):
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(). emb.sin()
return emb.cos(), emb.sin()
def rotate_half(self, x):
# Split and rotate
x1 = x[..., ::2]
x2 = x[..., 1::2]
rot_x = torch.stack([-x2, x1], dim=-1).flatten(-2)
return rot_x
"""Rotates half the hidden dims of the input."""
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
@torch.no_grad()
@add_start_docstrings("")
def forward(self, q, k, position_ids=None, unsqueeze_dim=1):
dtype = q.dtype
q,k = q.float(), k.float()
cos, sin = self.comput_cos_sin(q, position_ids)
cos = cos.unsqueeze(unsqueeze_dim)
sin = sin.unsqueeze(unsqueeze_dim)
"""Applies Rotary Position Embedding to the query and key tensors.
Args:
q (`torch.Tensor`): The query tensor.
k (`torch.Tensor`): The key tensor.
position_ids (`torch.Tensor`, *optional*):
Deprecated and unused.
unsqueeze_dim (`int`, *optional*, defaults to 1):
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
Returns:
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
"""
cos, sin = self.compute_cos_sin(q, position_ids)
cos = cos.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
sin = sin.unsqueeze(unsqueeze_dim).to(dtype=q.dtype)
q_embed = (q * cos) + (self.rotate_half(q) * sin)
k_embed = (k * cos) + (self.rotate_half(k) * sin)
return q_embed.to(dtype=dtype), k_embed.to(dtype=dtype)
class CohereMLP(nn.Module):
return q_embed, k_embed
class GemmaMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.hidden_size = config.hidden_size
self.intermediate_size = config.intermediate_size
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
self.act_fn = ACT2FN[config.hidden_act]
def forward(self, x):
if self.config.pretraining_tp > 1:
slice = self.intermediate_size // self.config.pretraining_tp
gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
up_proj_slices = self.up_proj.weight.split(slice, dim=0)
down_proj_slices = self.down_proj.weight.split(slice, dim=1)
gate_proj = torch.cat(
[F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
)
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
down_proj = [
F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
]
down_proj = sum(down_proj)
else:
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
return down_proj
class CohereAttention(nn.Module):
class GemmaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config: CohereConfig, layer_idx: Optional[int] = None):
def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
@ -136,7 +129,7 @@ class CohereAttention(nn.Module):
def _init_rope(self):
if self.config.rope_scaling is None:
self.rotary_emb = CohereRotaryEmbedding(
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
@ -145,14 +138,14 @@ class CohereAttention(nn.Module):
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = CohereLinearScalingRotaryEmbedding(
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 = CohereDynamicNTKScalingRotaryEmbedding(
self.rotary_emb = GemmaDynamicNTKScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
@ -170,43 +163,22 @@ class CohereAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
past_key_value = getattr(self, "past_key_value", past_key_value)
cos, sin = self.rotary_emb(value_states, position_ids)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
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 = {"sin": sin, "cos": cos, "cache_position": cache_position}
cache_kwargs = {"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)
@ -232,22 +204,16 @@ class CohereAttention(nn.Module):
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
if self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
class CohereSdpaAttention(nn.Module):
class GemmaSdpaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config: CohereConfig, layer_idx: Optional[int] = None):
def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
@ -282,7 +248,7 @@ class CohereSdpaAttention(nn.Module):
def _init_rope(self):
if self.config.rope_scaling is None:
self.rotary_emb = CohereRotaryEmbedding(
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
@ -291,14 +257,14 @@ class CohereSdpaAttention(nn.Module):
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = CohereLinearScalingRotaryEmbedding(
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 = CohereDynamicNTKScalingRotaryEmbedding(
self.rotary_emb = GemmaDynamicNTKScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
@ -316,43 +282,22 @@ class CohereSdpaAttention(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
past_key_value = getattr(self, "past_key_value", past_key_value)
cos, sin = self.rotary_emb(value_states, position_ids)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
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 = {"sin": sin, "cos": cos, "cache_position": cache_position}
cache_kwargs = {"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)
@ -378,22 +323,16 @@ class CohereSdpaAttention(nn.Module):
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
if self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
class CohereFlashAttention2(nn.Module):
class GemmaFlashAttention2(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config: CohereConfig, layer_idx: Optional[int] = None):
def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
@ -428,7 +367,7 @@ class CohereFlashAttention2(nn.Module):
def _init_rope(self):
if self.config.rope_scaling is None:
self.rotary_emb = CohereRotaryEmbedding(
self.rotary_emb = GemmaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
@ -437,14 +376,14 @@ class CohereFlashAttention2(nn.Module):
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = CohereLinearScalingRotaryEmbedding(
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 = CohereDynamicNTKScalingRotaryEmbedding(
self.rotary_emb = GemmaDynamicNTKScalingRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
@ -462,43 +401,22 @@ class CohereFlashAttention2(nn.Module):
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
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)
past_key_value = getattr(self, "past_key_value", past_key_value)
cos, sin = self.rotary_emb(value_states, position_ids)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
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 = {"sin": sin, "cos": cos, "cache_position": cache_position}
cache_kwargs = {"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)
@ -524,46 +442,57 @@ class CohereFlashAttention2(nn.Module):
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
if self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
COHERE_ATTENTION_CLASSES = {"eager": CohereAttention, "flash_attention_2": CohereFlashAttention2, "sdpa": CohereSdpaAttention}
COHERE_ATTENTION_CLASSES = {"eager": GemmaAttention, "flash_attention_2": GemmaFlashAttention2, "sdpa": GemmaSdpaAttention}
class CohereDecoderLayer(nn.Module):
def __init__(self, config: CohereConfig, layer_idx: int):
class GemmaDecoderLayer(nn.Module):
def __init__(self, config: GemmaConfig, layer_idx: int):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = COHERE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
self.self_attn = LLAMA_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
self.mlp = CohereMLP(config)
self.input_layernorm = CohereLayerNorm(hidden_size=(config.hidden_size), eps=config.layer_norm_eps)
self.mlp = GemmaMLP(config)
self.input_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
past_key_value: Optional[Cache] = None,
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = False,
cache_position: Optional[torch.LongTensor] = None,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`, *optional*):
attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
query_sequence_length, key_sequence_length)` if default attention is used.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
# Self Attention
hidden_states_attention, self_attn_weights, present_key_value = self.self_attn(
hidden_states, self_attn_weights, present_key_value = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
@ -572,12 +501,13 @@ class CohereDecoderLayer(nn.Module):
use_cache=use_cache,
cache_position=cache_position,
)
hidden_states = residual + hidden_states
# Fully Connected
hidden_states_mlp = self.mlp(hidden_states)
# Add everything together (main diff with llama )
hidden_states = residual + hidden_states_attention + hidden_states_mlp
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
@ -588,16 +518,15 @@ class CohereDecoderLayer(nn.Module):
outputs += (present_key_value,)
return outputs
@add_start_docstrings(
"The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
LLAMA_START_DOCSTRING,
)
class CoherePreTrainedModel(PreTrainedModel):
config_class = CohereConfig
class GemmaPreTrainedModel(PreTrainedModel):
config_class = GemmaConfig
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["CohereDecoderLayer"]
_no_split_modules = ["GemmaDecoderLayer"]
_skip_keys_device_placement = ["past_key_values"]
_supports_flash_attn_2 = True
_supports_sdpa = True
@ -614,443 +543,4 @@ class CoherePreTrainedModel(PreTrainedModel):
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
def _setup_cache(self, cache_cls, max_batch_size, max_cache_len: Optional[int] = None):
if self.config._attn_implementation == "flash_attention_2" and cache_cls == 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"
)
for layer in self.model.layers:
device = layer.input_layernorm.weight.device
if hasattr(self.config, "_pre_quantization_dtype"):
dtype = self.config._pre_quantization_dtype
else:
dtype = layer.self_attn.o_proj.weight.dtype
layer.self_attn.past_key_value = cache_cls(
self.config, max_batch_size, max_cache_len, device=device, dtype=dtype
)
def _reset_cache(self):
for layer in self.model.layers:
layer.self_attn.past_key_value = None
@add_start_docstrings(
"The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
LLAMA_START_DOCSTRING,
)
class CohereModel(CoherePreTrainedModel):
"""
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`CohereDecoderLayer`]
Args:
config: CohereConfig
"""
def __init__(self, config: CohereConfig):
super().__init__(config)
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
self.layers = nn.ModuleList(
[CohereDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
)
self.norm = CohereRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if (input_ids is None) ^ (inputs_embeds is not None):
raise ValueError(
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
)
if self.gradient_checkpointing and self.training and use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
)
use_cache = False
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
past_seen_tokens = 0
if use_cache: # kept for BC (cache positions)
if not isinstance(past_key_values, StaticCache):
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
past_seen_tokens = past_key_values.get_seq_length()
if cache_position is None:
if isinstance(past_key_values, StaticCache):
raise ValueError("cache_position is a required argument when using StaticCache.")
cache_position = torch.arange(
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
)
if position_ids is None:
position_ids = cache_position.unsqueeze(0)
causal_mask = self._update_causal_mask(
attention_mask, inputs_embeds, cache_position, past_seen_tokens + inputs_embeds.shape[1]
)
# embed positions
hidden_states = inputs_embeds
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
next_decoder_cache = None
for decoder_layer in self.layers:
if output_hidden_states:
all_hidden_states += (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
causal_mask,
position_ids,
past_key_values,
output_attentions,
use_cache,
cache_position,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=causal_mask,
position_ids=position_ids,
past_key_value=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
if output_attentions:
all_self_attns += (layer_outputs[1],)
hidden_states = self.norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = None
if use_cache:
next_cache = (
next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
)
if not return_dict:
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
return BaseModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=next_cache,
hidden_states=all_hidden_states,
attentions=all_self_attns,
)
# TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
# KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
# (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
# `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
def _update_causal_mask(self, attention_mask, input_tensor, cache_position, current_length):
if self.config._attn_implementation == "flash_attention_2":
if attention_mask is not None and 0.0 in attention_mask:
return attention_mask
return None
dtype, device = input_tensor.dtype, input_tensor.device
min_dtype = torch.finfo(dtype).min
sequence_length = input_tensor.shape[1]
if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"): # static cache
target_length = self.config.max_position_embeddings
else: # dynamic cache
target_length = (
attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
)
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
if sequence_length != 1:
causal_mask = torch.triu(causal_mask, diagonal=1)
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
if attention_mask is not None:
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
if attention_mask.dim() == 2:
mask_length = attention_mask.shape[-1]
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
elif attention_mask.dim() == 4:
# backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
# cache. In that case, the 4D attention mask attends to the newest tokens only.
if attention_mask.shape[-2] < cache_position[0] + sequence_length:
offset = cache_position[0]
else:
offset = 0
mask_shape = attention_mask.shape
mask_slice = (attention_mask.eq(0.0)).to(dtype=dtype) * min_dtype
causal_mask[
: mask_shape[0], : mask_shape[1], offset : mask_shape[2] + offset, : mask_shape[3]
] = mask_slice
if (
self.config._attn_implementation == "sdpa"
and attention_mask is not None
and attention_mask.device.type == "cuda"
):
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
return causal_mask
class CohereForCausalLM(CoherePreTrainedModel):
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config):
super().__init__(config)
self.model = CohereModel(config)
self.vocab_size = config.vocab_size
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.model.embed_tokens
def set_input_embeddings(self, value):
self.model.embed_tokens = value
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def set_decoder(self, decoder):
self.model = decoder
def get_decoder(self):
return self.model
@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
) -> Union[Tuple, CausalLMOutputWithPast]:
r"""
Args:
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Returns:
Example:
```python
>>> from transformers import AutoTokenizer, CohereForCausalLM
>>> model = CohereForCausalLM.from_pretrained("meta-llama/Cohere-2-7b-hf")
>>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Cohere-2-7b-hf")
>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
```"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
hidden_states = outputs[0]
if self.config.pretraining_tp > 1:
lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
logits = torch.cat(logits, dim=-1)
else:
logits = self.lm_head(hidden_states)
logits = logits.float()
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
shift_logits = shift_logits.view(-1, self.config.vocab_size)
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
loss = loss_fct(shift_logits, shift_labels)
if not return_dict:
output = (logits,) + outputs[1:]
return (loss,) + output if loss is not None else output
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def prepare_inputs_for_generation(
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, **kwargs
):
# With static cache, the `past_key_values` is None
# TODO joao: standardize interface for the different Cache classes and remove of this if
has_static_cache = False
if past_key_values is None:
past_key_values = getattr(getattr(self.model.layers[0], "self_attn", {}), "past_key_value", None)
has_static_cache = past_key_values is not None
past_length = 0
if past_key_values is not None:
if isinstance(past_key_values, Cache):
past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length()
max_cache_length = (
torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
if past_key_values.get_max_length() is not None
else None
)
cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)
# TODO joao: remove this `else` after `generate` prioritizes `Cache` objects
else:
cache_length = past_length = past_key_values[0][0].shape[2]
max_cache_length = None
# Keep only the unprocessed tokens:
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
# input)
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
# input_ids based on the past_length.
elif past_length < input_ids.shape[1]:
input_ids = input_ids[:, past_length:]
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
if (
max_cache_length is not None
and attention_mask is not None
and cache_length + input_ids.shape[1] > max_cache_length
):
attention_mask = attention_mask[:, -max_cache_length:]
position_ids = kwargs.get("position_ids", None)
if attention_mask is not None and position_ids is None:
# create position_ids on the fly for batch generation
position_ids = attention_mask.long().cumsum(-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
if past_key_values:
position_ids = position_ids[:, -input_ids.shape[1] :]
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
if inputs_embeds is not None and past_key_values is None:
model_inputs = {"inputs_embeds": inputs_embeds}
else:
# The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
# recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
# TODO: use `next_tokens` directly instead.
model_inputs = {"input_ids": input_ids.contiguous()}
input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1]
if cache_position is None:
cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device)
else:
cache_position = cache_position[-input_length:]
if has_static_cache:
past_key_values = None
model_inputs.update(
{
"position_ids": position_ids,
"cache_position": cache_position,
"past_key_values": past_key_values,
"use_cache": kwargs.get("use_cache"),
"attention_mask": attention_mask,
}
)
return model_inputs
@staticmethod
def _reorder_cache(past_key_values, beam_idx):
reordered_past = ()
for layer_past in past_key_values:
reordered_past += (
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
)
return reordered_past

850
src/transformers/models/cohere/modeling_cohere.py
View File

File diff suppressed because it is too large Load Diff

1
src/transformers/models/starcoder2/diff_startcoder2.py → src/transformers/models/starcoder2/diff_starcoder2.py
View File

@ -1,6 +1,5 @@
from typing import List, Tuple
from torch import FloatTensor, LongTensor, Tensor
from torch._C import FloatTensor, LongTensor
from transformers.modeling_outputs import BaseModelOutputWithPast
from transformers.models.llama.configuration_llama import LlamaConfig
from transformers.models.llama.modeling_llama import *

1419
src/transformers/models/starcoder2/modeling_starcoder2.py
View File

File diff suppressed because it is too large Load Diff

2
src/transformers/utils/__init__.py
View File

@ -215,7 +215,7 @@ from .peft_utils import (
check_peft_version,
find_adapter_config_file,
)
from .convert_diff_to_single_file import ModelConverter
from .model_converter import ModelConverter
WEIGHTS_NAME = "pytorch_model.bin"
WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"

70
src/transformers/utils/convert_diff_to_single_file.py
View File

@ -1,70 +0,0 @@
"""
running this script on `src/transformers/models/**_diff.py` should produce the equivalent single model single files
1. Iterate though `**_diff.py` files
2. How to handle the imports?
a. `model_type` should always be present?
b. `ConfigClass` should be defined as well?
3. Copy each class and function one by one.
a. if there is a class registered for this file like `@__file__.register(MyNewClass, OldClass)`
then copy the content of `OldClass`, replacing all names of `Old` with `MyNew`.
Also copy the decorators that are on top of this class.
b. if there is inheritance, copy non-overloaded functions from base, and overloaded from non base.
4. Register things?
new = type("new_class", (torch.nn.Linear,),{})
new__main__.new_class
new(10,10)
new_class(in_features=10, out_features=10, bias=True)
CohereConverter = ModelConverter(__file__)
CohereMLP = CohereConverter.register("CohereMLP", LlamaMLP)
CohereMLP
<class 'transformers.models.cohere.modeling_cohere.CohereMLP'>
CohereMLP(LlamaConfig())
CohereMLP(
(gate_proj): Linear(in_features=4096, out_features=11008, bias=False)
(up_proj): Linear(in_features=4096, out_features=11008, bias=False)
(down_proj): Linear(in_features=11008, out_features=4096, bias=False)
(act_fn): SiLU()
)
>>> CohereMLP(LlamaConfig())(torch.ones(1,1,4096))
How to deal with submodules?
CohereSdpaAttention(
(q_proj): Linear(in_features=4096, out_features=4096, bias=False)
(k_proj): Linear(in_features=4096, out_features=4096, bias=False)
(v_proj): Linear(in_features=4096, out_features=4096, bias=False)
(o_proj): Linear(in_features=4096, out_features=4096, bias=False)
(rotary_emb): LlamaRotaryEmbedding()
)
"""
import regex as re
class ModelConverter:
def __init__(self, file):
self.diff_file = file
self.model_name = re.search(r'models/(.*?)/diff', self.diff_file).group(1)
self.modeling_file = file.replace("diff", "modeling")
self.registered_classes = {}
self.modules_to_import = []
def register(self, new_class, old_class):
# registering. Returns the old class to be usable with a new name
self.registered_classes[new_class] = old_class
self.modules_to_import.append([old_class, old_class.__module__])
new_class = type(new_class, (old_class,), {})
base_model_name = re.search(r'models\.(.*?)\.modeling', old_class.__module__).group(1)
new_class.__module__ = re.sub(base_model_name, self.model_name, old_class.__module__)
return new_class
def __repr__(self) -> str:
return f"ModelConverter({self.diff_file}, {self.model_name}, {self.registered_classes})"

91
utils/convert_diff_model.py
View File

@ -10,47 +10,78 @@ import inspect
from transformers.models.llama.modeling_llama import *
from transformers.models.cohere.diff_cohere import *
from transformers.models.starcoder2.modeling_starcoder2 import *
# 1. all the imports from the original file should be copied until end of header? __HEADER__
# with open(CohereConverter.original_file, 'r') as file, open("result.py", "w+") as modeling:
# pass
# TODO also copy and import from all modules in CohereConverter.modules_to_import to be able to use inspect
# 2. Write all the classes. Use the `CohereConverter` class for this.
with open(CohereConverter.diff_file, 'r') as file, open("result.py", "w+") as modeling:
function_set = {}
for line in file:
if "Converter.register" in line: # TODO use map() to map lines to this
# write the code of the original model
class_to_use, old_class = re.search(r'Converter\.register\(\"(.*?)\", (.*?)\)', line).groups()
model_identifier_camel = re.findall(r'[A-Z][a-z0-9]*', class_to_use)[0]
old_model_identifier_camel = re.findall(r'[A-Z][a-z0-9]*', old_class)[0]
source_code = inspect.getsource(CohereConverter.registered_classes[class_to_use]).replace(old_class, class_to_use)
source_code = source_code.replace(old_model_identifier_camel, model_identifier_camel)
modeling.write(source_code)
elif match:=re.match(r"class (\w+)\((\w+)\):", line):
class_name, parent_class = match.groups()
pattern = re.compile(r"(( [\s\S]*?)\n(\n)?(?= \S))|(( [\s\S]*?)(?=\Z))", re.MULTILINE)
def create_single_model_file(converter):
model_identifier = converter.diff_file.split("diff_")
# temporarily add the source to the path in order to load everything?
with open(converter.diff_file, 'r') as file, open(f"{model_identifier[0]}modeling_{model_identifier[1]}", "w+") as modeling:
function_set = {}
for line in file:
if "Converter.register" in line: # TODO use map() to map lines to this
# write the code of the original model
class_to_use, old_class = re.search(r'Converter\.register\(\"(.*?)\", (.*?)\)', line).groups()
model_identifier_camel = re.findall(r'[A-Z][a-z0-9]*', class_to_use)[0]
old_model_identifier_camel = re.findall(r'[A-Z][a-z0-9]*', old_class)[0]
source_code = inspect.getsource(converter.registered_classes[class_to_use]).replace(old_class, class_to_use)
source_code = source_code.replace(old_model_identifier_camel, model_identifier_camel)
modeling.write(source_code)
elif match:=re.match(r"class (\w+)\((\w+)\):", line):
class_name, parent_class = match.groups()
pattern = re.compile(r"(( [\s\S]*?)\n(\n)?(?= \S))|(( [\s\S]*?)(?=\Z))", re.MULTILINE)
parent_class_def = inspect.getsource(eval(parent_class))
modeling.write(parent_class_def.split('\n')[0].replace(parent_class,class_name)+"\n")
parent_class_def = inspect.getsource(eval(parent_class))
modeling.write(parent_class_def.split('\n')[0].replace(parent_class,class_name)+"\n")
matches = pattern.finditer(parent_class_def)
function_set = {}
for match in matches:
full_function = match.group()
function_set[full_function.split("(")[0]] = full_function
class_def = inspect.getsource(eval(class_name))
matches = pattern.finditer(class_def)
for match in matches:
# TODO handle call to super!
matches = pattern.finditer(parent_class_def)
function_set = {}
for match in matches:
full_function = match.group()
function_set[full_function.split("(")[0]] = full_function
modeling.write("".join(function_set.values())) # TODO we wrote the code, next lines shall be ignored
elif line not in "".join(function_set.values()) or line=="\n":
modeling.write(line)
class_def = inspect.getsource(eval(class_name))
matches = pattern.finditer(class_def)
for match in matches:
# TODO handle call to super!
full_function = match.group()
function_set[full_function.split("(")[0]] = full_function
modeling.write("".join(function_set.values())) # TODO we wrote the code, next lines shall be ignored
elif line not in "".join(function_set.values()) or line=="\n":
modeling.write(line)
def dynamically_import_object(module_path, object_name):
try:
module = importlib.import_module(module_path)
obj = getattr(module, object_name)
return obj
except (ImportError, AttributeError) as e:
print(f"Failed to import object '{object_name}' from module '{module_path}'")
print(e)
exit(0)
# 3. Apply ruff fix to remove unused imports
# 4. Run a tiny test to import from this new file.
import importlib
import argparse
import glob
from transformers import MODEL_NAMES_MAPPING
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--files_to_parse", default="all", help="A list of `diff_xxxx` files that should be converted to single model file")
args = parser.parse_args()
if args.files_to_parse == "all":
args.files_to_parse = glob.glob("src/transformers/models/**/diff_*.py", recursive=True)
for file_name in args.files_to_parse:
print(f"Converting {file_name} to a single model single file format")
module_path = file_name.replace("/",".").replace(".py","").replace("src.","")
model_name = MODEL_NAMES_MAPPING[module_path.split('_')[-1]]
converter = dynamically_import_object(module_path, f"{model_name}Converter")
create_single_model_file(converter)