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

Only write to HBM at the last iteration. #8393

Merged
merged 2 commits into from
Nov 20, 2024
Merged

Only write to HBM at the last iteration. #8393

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d52c6f2
Only write to HBM at the last iteration.
vanbasten23 Nov 19, 2024
1954394
fix the test
vanbasten23 Nov 20, 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
7 changes: 4 additions & 3 deletions test/test_tpu_paged_attention_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,7 +162,8 @@ def test_paged_attention_without_query_padding(
num_kv_pages_per_compute_block=num_kv_pages_per_compute_block,
num_queries_per_compute_block=num_queries_per_compute_block,
)
# actual_output = jax.block_until_ready(actual_output)
# Note kernel execution is async. Without blocking, if an error happens in the kernel, the error may point to some irrelevant and confusing places. See https://github.com/pytorch/xla/pull/8356#issuecomment-2486861631
actual_output = jax.block_until_ready(actual_output)

# Run the ref impl.
expected_output = _ref_jax_extended_paged_attention(
Expand Down Expand Up @@ -219,7 +220,7 @@ def test_paged_attention_with_query_padding(
block_kv_size,
):

max_kv_len = 2048
max_kv_len = 512
# Set query_len>kv_seq_lens
query_len = max_kv_len
batch_size = 3
Expand Down Expand Up @@ -259,7 +260,7 @@ def test_paged_attention_with_query_padding(
num_kv_pages_per_compute_block=num_kv_pages_per_compute_block,
num_queries_per_compute_block=num_queries_per_compute_block,
)
# actual_output = jax.block_until_ready(actual_output)
actual_output = jax.block_until_ready(actual_output)

# Run the ref impl.
expected_output = _ref_jax_extended_paged_attention(
Expand Down
15 changes: 8 additions & 7 deletions torch_xla/experimental/pallas_kernels/multi_queries_paged_attention_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,13 +198,14 @@ def start_new_sequence():
o_curr = jax.lax.dot(p.astype(v.dtype), v, preferred_element_type=jnp.float32)
acc_scratch_ref[q_head_idx_per_kv] += o_curr * l_broadcast(l_next_inv_safe)

# TODO: To potentially improve the perf, consider not to update o_ref, l_ref, and m_ref at every kv_blk_idx. Instead, use a proper @pl.when(kv_blk_idx == ...) at the last kv_block.
o_ref[0, q_head_idx_per_kv] = acc_scratch_ref[q_head_idx_per_kv].astype(
o_ref.dtype)
l_ref[0, q_head_idx_per_kv] = l_scratch_ref[q_head_idx_per_kv].astype(
l_ref.dtype)
m_ref[0, q_head_idx_per_kv] = m_scratch_ref[q_head_idx_per_kv].astype(
m_ref.dtype)
@pl.when(kv_blk_idx == kv_len // kv_seq_len_per_kv_compute_blk)
def store_to_output():
o_ref[0, q_head_idx_per_kv] = acc_scratch_ref[q_head_idx_per_kv].astype(
o_ref.dtype)
l_ref[0, q_head_idx_per_kv] = l_scratch_ref[q_head_idx_per_kv].astype(
l_ref.dtype)
m_ref[0, q_head_idx_per_kv] = m_scratch_ref[q_head_idx_per_kv].astype(
m_ref.dtype)


def paged_flash_attention_kernel(
Expand Down
Loading