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

fix(layers): fix SuRotaryEmbedding #2060

Merged
merged 3 commits into from
Jun 12, 2024
Merged

fix(layers): fix SuRotaryEmbedding #2060

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9cc1672
fix(layers): fix SuRotaryEmbedding
OlivierDehaene Jun 12, 2024
9775fac
change arange
OlivierDehaene Jun 12, 2024
8230226
remove logs
OlivierDehaene Jun 12, 2024
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
26 changes: 14 additions & 12 deletions server/text_generation_server/layers/rotary.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -267,19 +267,21 @@ def _update_cos_sin_cache(self, dtype, device, seqlen):
or self._cos_cached.dtype != dtype
):
self._seq_len_cached = seqlen
if seqlen > self.original_max_position_embeddings:
inv_freq = self.long_inv_freq
else:
inv_freq = self.short_inv_freq
t = torch.arange(seqlen, device=device, dtype=inv_freq.dtype)
if self.scaling_factor is not None:
t /= self.scaling_factor
# Don't do einsum, it converts fp32 to fp16
# freqs = torch.einsum("i,j->ij", t, self.inv_freq)

freqs = torch.outer(t, inv_freq.to(device=t.device))
self._cos_cached = torch.cos(freqs).to(dtype)
self._sin_cached = torch.sin(freqs).to(dtype)
t = torch.arange(seqlen, device=device, dtype=self.short_inv_freq.dtype)
short_freqs = torch.outer(
t[: self.original_max_position_embeddings],
self.short_inv_freq.to(device=t.device),
)
long_freqs = torch.outer(
t[self.original_max_position_embeddings :],
self.long_inv_freq.to(device=t.device),
)

freqs = torch.cat([short_freqs, long_freqs])

self._cos_cached = (torch.cos(freqs) * self.scaling_factor).to(dtype)
self._sin_cached = (torch.sin(freqs) * self.scaling_factor).to(dtype)


class DynamicPositionRotaryEmbedding(PositionRotaryEmbedding):
Expand Down
3 changes: 1 addition & 2 deletions server/text_generation_server/models/flash_phi.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,6 @@
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_phi_modeling import (
FlashPhiForCausalLM,
PhiConfig,
)
from text_generation_server.utils import (
initialize_torch_distributed,
Expand Down Expand Up @@ -44,7 +43,7 @@ def __init__(
trust_remote_code=trust_remote_code,
)

config = PhiConfig.from_pretrained(
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
Expand Down
Loading