Skip to content

[TP] Infer local n_heads instead of ad-hoc model changes #498

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 3 commits into from
Aug 2, 2024
Merged

[TP] Infer local n_heads instead of ad-hoc model changes #498

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
bb08236
Update
kwen2501 Aug 2, 2024
a1ef2b4
Update
kwen2501 Aug 2, 2024
8fd1308
Update
kwen2501 Aug 2, 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
9 changes: 6 additions & 3 deletions torchtitan/models/llama/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,9 +190,12 @@ def forward(
bs, seqlen, _ = x.shape
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)

xq = xq.view(bs, seqlen, self.n_heads, self.head_dim)
xk = xk.view(bs, seqlen, self.n_kv_heads, self.head_dim)
xv = xv.view(bs, seqlen, self.n_kv_heads, self.head_dim)
# Use -1 instead of `n_heads` (or `n_kv_heads`) to infer the actual
# local heads from sizes of xq, xk, and xv as TP may have sharded them
# after the above linear ops.
xq = xq.view(bs, seqlen, -1, self.head_dim)
xk = xk.view(bs, seqlen, -1, self.head_dim)
xv = xv.view(bs, seqlen, -1, self.head_dim)

xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)

Expand Down
5 changes: 0 additions & 5 deletions torchtitan/parallelisms/parallelize_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -398,11 +398,6 @@ def apply_tp(
"feed_forward.w3": colwise_parallel_weight(),
}

# Adjust attention module to use the local number of heads
attn_layer = transformer_block.attention
attn_layer.n_heads = attn_layer.n_heads // tp_mesh.size()
attn_layer.n_kv_heads = attn_layer.n_kv_heads // tp_mesh.size()

parallelize_module(
module=transformer_block,
device_mesh=tp_mesh,
Expand Down