SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[RoPE] abstract dynamic RoPE update under a decorator ✨ (#37249)

Browse Source 
* dynamic rope decorator

* longrope; shorter fwd pass

* propper docstring

* make fixup
This commit is contained in:
Joao Gante
2025-04-04 14:27:28 +01:00
committed by GitHub
parent 3d40bda30e
commit ad3d157188
52 changed files with 527 additions and 1833 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
src/transformers/models/aria/modeling_aria.py
View File

@@ -28,7 +28,7 @@ from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import AttentionMaskConverter
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, ModelOutput
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
@@ -752,47 +752,18 @@ class AriaTextRotaryEmbedding(nn.Module):
self.register_buffer("inv_freq", inv_freq, persistent=False)
self.original_inv_freq = self.inv_freq
def _dynamic_frequency_update(self, position_ids, device):
"""
dynamic RoPE layers should recompute `inv_freq` in the following situations:
1 - growing beyond the cached sequence length (allow scaling)
2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
"""
seq_len = torch.max(position_ids) + 1
if seq_len > self.max_seq_len_cached: # growth
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation
self.max_seq_len_cached = seq_len
if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset
# This .to() is needed if the model has been moved to a device after being initialized (because
# the buffer is automatically moved, but not the original copy)
self.original_inv_freq = self.original_inv_freq.to(device)
self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
self.max_seq_len_cached = self.original_max_seq_len
@torch.no_grad()
@dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope)
def forward(self, x, position_ids):
if "dynamic" in self.rope_type:
self._dynamic_frequency_update(position_ids, device=x.device)
# Core RoPE block
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
position_ids_expanded = position_ids[:, None, :].float()
# Force float32 (see https://github.com/huggingface/transformers/pull/29285)
device_type = x.device.type
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
with torch.autocast(device_type=device_type, enabled=False):
freqs = (
inv_freq_expanded.to(device=x.device, dtype=torch.float) @ position_ids_expanded.float()
).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
cos = emb.cos()
sin = emb.sin()
# Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
cos = cos * self.attention_scaling
sin = sin * self.attention_scaling
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
with torch.autocast(device_type=device_type, enabled=False): # Force float32
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
emb = torch.cat((freqs, freqs), dim=-1)
cos = emb.cos() * self.attention_scaling
sin = emb.sin() * self.attention_scaling
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)