Add ragged paged attention #8659
Open
+1,359
−1
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
bythew3i
reviewed
Jan 31, 2025
torch_xla/experimental/pallas_kernels/ragged_paged_attention_kernel.py
Outdated
Show resolved
Hide resolved
torch_xla/experimental/pallas_kernels/ragged_paged_attention_kernel.py
Outdated
Show resolved
Hide resolved
torch_xla/experimental/pallas_kernels/ragged_paged_attention_kernel.py
Outdated
Show resolved
Hide resolved
torch_xla/experimental/pallas_kernels/ragged_paged_attention_kernel.py
Outdated
Show resolved
Hide resolved
torch_xla/experimental/pallas_kernels/ragged_paged_attention_kernel.py
Outdated
Show resolved
Hide resolved
torch_xla/experimental/pallas_kernels/ragged_paged_attention_kernel.py
Outdated
Show resolved
Hide resolved
| last_time_seeing_cur_physical_q_blk = jnp.logical_or(is_last_logical_q_blk, physical_q_blk_will_change) | ||
| should_store_to_hbm = jnp.logical_and(is_last_kv_blk_idx, last_time_seeing_cur_physical_q_blk) | ||
| @pl.when(should_store_to_hbm) | ||
| def store_to_hbm(): # pylint: disable=unused-variable |
There was a problem hiding this comment.
This function actually only stores to VMEM and all the ref used are actually VMEM ref. We rely on pipeline emitter to send vmem block back to HBM. And we can't store vregs directly to HBM in kernel. So I think original store_to_output makes more sense here.
Comment on lines
894
to
884
| pages_per_sequence=pages_per_sequence, | ||
| num_tokens=num_tokens, | ||
| num_seqs=num_seqs, # it they changes, need to recompile. | ||
| num_kv_pages_per_compute_block=num_kv_pages_per_compute_block, | ||
| mask_value=mask_value, |
There was a problem hiding this comment.
These are all calculated from either static shape or static_argnames.
The overhead could be very large if some are too dynamic (num_seqs, num_tokens) because we need to recompile many times. Maybe to add a note or TODO here.
How is 65536 calculated? |
I found a ticket and someone uses it. I remember the number is the vmem limit on a TPU generation. |
vanbasten23
force-pushed
the
xiowei/add_ragged_paged_attention
branch
from
ad2f87c to
9e4b227
Compare
bythew3i
approved these changes
Feb 1, 2025
…_extreme_one_tokens_per_sequence.
…added runtime check.
vanbasten23
force-pushed
the
xiowei/add_ragged_paged_attention
branch
from
9e4b227 to
7fe5071
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Test plan: