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

[Pallas] Improve FlashAttention segment_ids test case #7034

Merged
merged 6 commits into from
May 8, 2024
Merged

[Pallas] Improve FlashAttention segment_ids test case #7034

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
407e83f
tmp
alanwaketan May 7, 2024
c28f379
some minor changes
alanwaketan May 7, 2024
d81d869
Fix everything
alanwaketan May 7, 2024
3470ac3
Enable backward as well
alanwaketan May 7, 2024
9d6955e
Delete customer code
alanwaketan May 7, 2024
faf570e
Improve test case
alanwaketan May 7, 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
33 changes: 16 additions & 17 deletions test/test_pallas.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -643,24 +643,23 @@ def test_flash_attention_wrapper_segment_ids_1(self):
q = torch.randn(3, 2, 128, 4)
k = torch.randn(3, 2, 128, 4)
v = torch.randn(3, 2, 128, 4)
q_segment_ids = torch.zeros(3, 128)
kv_segment_ids = torch.zeros(3, 128)
zeros = torch.zeros(3, 32)
segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
o = flash_attention(
q.to("xla"), k.to("xla"), v.to("xla"), False, q_segment_ids.to("xla"),
kv_segment_ids.to("xla"))
q.to("xla"), k.to("xla"), v.to("xla"), False, segment_ids.to("xla"),
segment_ids.to("xla"))

jax_q = jnp.array(q.numpy(), dtype=jnp.float32)
jax_k = jnp.array(k.numpy(), dtype=jnp.float32)
jax_v = jnp.array(v.numpy(), dtype=jnp.float32)
jax_q_segment_ids = jnp.array(q_segment_ids.numpy(), dtype=jnp.float32)
jax_kv_segment_ids = jnp.array(kv_segment_ids.numpy(), dtype=jnp.float32)
jax_segment_ids = jnp.array(segment_ids.numpy(), dtype=jnp.float32)
expected_o = torch.from_numpy(
np.array(
jax_flash_attention(
jax_q,
jax_k,
jax_v,
segment_ids=SegmentIds(jax_q_segment_ids, jax_kv_segment_ids),
segment_ids=SegmentIds(jax_segment_ids, jax_segment_ids),
)))

self.assertTrue(torch.allclose(o.cpu(), expected_o.cpu(), atol=1e-05))
Expand All @@ -674,16 +673,16 @@ def test_flash_attention_wrapper_segment_ids_2(self):
q = torch.randn(3, 2, 128, 4).to("xla")
k = torch.randn(3, 2, 128, 4).to("xla")
v = torch.randn(3, 2, 128, 4).to("xla")
q_segment_ids = torch.zeros(3, 128).to("xla")
kv_segment_ids = torch.zeros(3, 128).to("xla")
o = flash_attention(q, k, v, False, q_segment_ids, kv_segment_ids)
zeros = torch.zeros(3, 32).to("xla")
segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
o = flash_attention(q, k, v, False, segment_ids, segment_ids)

expected_o = self._attention(
q,
k,
v,
attn_mask=self._make_attention_mask_from_segment_ids(
q_segment_ids, kv_segment_ids))
segment_ids, segment_ids))
self.assertTrue(torch.allclose(o.cpu(), expected_o.cpu(), atol=1e-05))
jax.config.update('jax_default_matmul_precision', jax.lax.Precision.DEFAULT)

Expand All @@ -697,13 +696,13 @@ def test_flash_attention_backward_segment_ids(self):
q = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
k = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
v = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
q_segment_ids = torch.zeros(4, 128).to("xla")
kv_segment_ids = torch.zeros(4, 128).to("xla")
zeros = torch.zeros(4, 32).to("xla")
segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
q.retain_grad()
k.retain_grad()
v.retain_grad()

o = flash_attention(q, k, v, False, q_segment_ids, kv_segment_ids)
o = flash_attention(q, k, v, False, segment_ids, segment_ids)
loss = o.sum()
loss.backward()
xm.mark_step()
Expand All @@ -716,8 +715,8 @@ def test_flash_attention_backward_segment_ids(self):
q = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
k = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
v = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
q_segment_ids = torch.zeros(4, 128).to("xla")
kv_segment_ids = torch.zeros(4, 128).to("xla")
zeros = torch.zeros(4, 32).to("xla")
segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
q.retain_grad()
k.retain_grad()
v.retain_grad()
Expand All @@ -727,7 +726,7 @@ def test_flash_attention_backward_segment_ids(self):
k,
v,
attn_mask=self._make_attention_mask_from_segment_ids(
q_segment_ids, kv_segment_ids))
segment_ids, segment_ids))
loss = o.sum()
loss.backward()
xm.mark_step()
Expand Down
Loading