Skip to content

Fix CI after local sharding local_map stack lands #188

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 1 commit into from
Oct 13, 2025
Merged

Fix CI after local sharding local_map stack lands #188

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
16 changes: 5 additions & 11 deletions examples/example_local_map.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,20 +61,14 @@ def replicate_linear(w, x):


@local_map(
out_placements=(
(Shard(0), Shard(0), Replicate()),
None,
),
in_placements=(
(Shard(0), Shard(0), Replicate()),
None,
),
out_placements=((Shard(0), Shard(0), Replicate()),),
in_placements=((Shard(0), Shard(0), Replicate()),),
redistribute_inputs=True,
in_grad_placements=None,
device_mesh=mesh,
)
def sharded_pointwise(x, scalar):
return x + scalar, scalar
def sharded_pointwise(x):
return x + 10


@local_map(
Expand Down Expand Up @@ -114,7 +108,7 @@ def init_weights(self):
torch.nn.init.normal_(lin.bias)

def _compute_attention(self, x):
boosted_weight, scalar = sharded_pointwise(self.wq.weight, 10)
boosted_weight = sharded_pointwise(self.wq.weight)
q = replicate_linear(boosted_weight, x)
k = self.wk(x)
v = self.wv(x)
Expand Down
7 changes: 4 additions & 3 deletions tests/test_optimize_placement.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -372,7 +372,7 @@ def input_fn():


class LocalMapTransformerBlock(nn.Module):
def __init__(self, nheads, dim1, dim2):
def __init__(self, nheads, dim1, dim2, mesh):
super().__init__()
self.nheads = nheads
bias = False
Expand All @@ -382,6 +382,7 @@ def __init__(self, nheads, dim1, dim2):
self.wo = nn.Linear(dim1, dim1, bias=bias)
self.w1 = nn.Linear(dim1, dim2, bias=bias)
self.w2 = nn.Linear(dim2, dim1, bias=bias)
self.mesh = mesh

def forward(self, x):
@local_map(
Expand All @@ -393,7 +394,7 @@ def forward(self, x):
),
redistribute_inputs=True,
in_grad_placements=None,
device_mesh=None,
device_mesh=self.mesh,
)
def _context_parallel_attention(query, key, value):
out = F.scaled_dot_product_attention(
Expand Down Expand Up @@ -435,7 +436,7 @@ def test_local_map_placement_respected(device_mesh_local_map, device="cuda"):
seq_len = 256

def model_fn():
return LocalMapTransformerBlock(nheads, dim1, dim2)
return LocalMapTransformerBlock(nheads, dim1, dim2, device_mesh_local_map)

def input_fn():
return torch.randn(bs, seq_len, dim1, device=device, requires_grad=True)
Expand Down
Loading