Skip to content

allow customized head_dim #6872

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

Merged
merged 5 commits into from
Nov 25, 2024
Merged

allow customized head_dim #6872

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
c083d85
allow customized head_dim
helunwencser Nov 14, 2024
0120876
Update on "allow customized head_dim"
helunwencser Nov 18, 2024
1c87ce3
Update on "allow customized head_dim"
helunwencser Nov 18, 2024
2a1ace5
Update on "allow customized head_dim"
helunwencser Nov 18, 2024
811ad8b
Update on "allow customized head_dim"
helunwencser Nov 25, 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
12 changes: 8 additions & 4 deletions examples/models/llama/llama_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,6 +85,7 @@ class ModelArgs:
n_kv_heads: Optional[int] = None
vocab_size: int = -1 # defined later by tokenizer
hidden_dim: Optional[int] = None
head_dim: Optional[int] = None # Optional customized head_dim
multiple_of: int = 256 # make SwiGLU hidden layer size multiple of large power of 2
ffn_dim_multiplier: Optional[float] = None
norm_eps: float = 1e-5
Expand Down Expand Up @@ -142,6 +143,9 @@ def __post_init__(self):
hidden_dim = int(self.ffn_dim_multiplier * hidden_dim)
self.hidden_dim = find_multiple(hidden_dim, multiple_of)

if self.head_dim is None:
self.head_dim = self.dim // self.n_heads


class KVCache(nn.Module):
def __init__(
Expand Down Expand Up @@ -272,7 +276,7 @@ def __init__(self, args: ModelArgs, layer_id: int):
self.n_local_heads = self.n_heads // model_parallel_size
self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
self.n_rep = self.n_local_heads // self.n_local_kv_heads
self.head_dim = args.dim // self.n_heads
self.head_dim = args.head_dim
self.max_batch_size = args.max_batch_size
self.max_seq_len = args.max_seq_len
self.dim = args.dim
Expand Down Expand Up @@ -304,7 +308,7 @@ def __init__(self, args: ModelArgs, layer_id: int):
)
self.SDPA = SDPA(
kv_cache=self.kv_cache,
dim=self.dim,
dim=self.n_local_heads * self.head_dim,
head_dim=self.head_dim,
n_rep=self.n_rep,
max_seq_len=self.max_seq_len,
Expand Down Expand Up @@ -425,7 +429,7 @@ def __init__(self, layer_id: int, args: ModelArgs):
self.use_kv_cache = args.use_kv_cache
self.n_heads = args.n_heads
self.dim = args.dim
self.head_dim = args.dim // args.n_heads
self.head_dim = args.head_dim
self.attention = Attention(args, layer_id)
if args.moe:
self.block_sparse_moe = MOEFeedForward(args)
Expand Down Expand Up @@ -472,7 +476,7 @@ def __init__(self, params: ModelArgs):
precompute_freqs_cis, use_scaled=params.use_scaled_rope
)
freqs_cos, freqs_sin = self.precompute_freqs_cis(
params.dim // params.n_heads,
params.head_dim,
(
params.max_seq_len # Normal llama2.
if params.ffn_dim_multiplier is None
Expand Down
Loading