Skip to content

[ DON'T MERGE ] Experimenting with make_dynamic_splash_mha #3186

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

Closed
parambole wants to merge 1 commit into from
Closed

[ DON'T MERGE ] Experimenting with make_dynamic_splash_mha #3186

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
35 changes: 16 additions & 19 deletions src/MaxText/layers/attention_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1303,27 +1303,24 @@ def wrap_flash_attention(
decoder_segment_ids_tuple = splash_attention_kernel.SegmentIds(decoder_segment_ids_q, decoder_segment_ids_kv)
else:
decoder_segment_ids_tuple = None

if self.config.use_tokamax_splash:
if self.config.use_sparse_indexer and index_mask is not None:
# Construct the splash kernel call with dynamic mask
def dynamic_mask_splash_kernel(q, k, v, segment, sinks, index_mask):
splash_kernel = tokamax_splash_kernel.make_dynamic_splash_mha(
mask=index_mask,
config=sa_config,
)
kernel = partial(splash_kernel, max_logit_value=max_logit_value)
return kernel(q, k, v, segment, sinks=sinks)

# Iterate over batch dimension for (query, key, value, segment, sinks, mask)
attn_fn = jax.vmap(dynamic_mask_splash_kernel, (0, 0, 0, 0, None, 0))
index_mask = jnp.isclose(index_mask, 0.0)
attention_output = attn_fn(query, key, value, decoder_segment_ids_tuple, sinks, index_mask)
else:
kernel = partial(splash_kernel, max_logit_value=max_logit_value)
attention_output = jax.vmap(lambda q, k, v, d, s: kernel(q, k, v, d, sinks=s), in_axes=(0, 0, 0, 0, None))(
query, key, value, decoder_segment_ids_tuple, sinks
)
q_len = query.shape[2]
kv_len = key.shape[2]

row_ids = jax.lax.broadcasted_iota(jnp.int32, (q_len, kv_len), 0)
col_ids = jax.lax.broadcasted_iota(jnp.int32, (q_len, kv_len), 1)
causal_bool_mask = row_ids >= col_ids

splash_kernel = tokamax_splash_kernel.make_dynamic_splash_mha(
mask=causal_bool_mask,
config=sa_config,
)
kernel = partial(splash_kernel, max_logit_value=max_logit_value)

attention_output = jax.vmap(lambda q, k, v, d, s: kernel(q, k, v, d, sinks=s), in_axes=(0, 0, 0, 0, None))(
query, key, value, decoder_segment_ids_tuple, sinks
)
elif self.config.use_jax_splash:
materialized_mask = jnp.asarray(mask[:, :])
attention_output = jax_flash_attention.flash_attention_block_masked(
Expand Down
Loading