Skip to content

Adding PagedAttention support for CausalLM models #1209

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

Open
vaibverm wants to merge 15 commits into
base: main
Choose a base branch
from
Open

Adding PagedAttention support for CausalLM models #1209

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
e6fdf91
Rebased PagedAttention support with latest Qeff for PR
vaibverm May 8, 2026
f5d449c
Added block_table and slot_id inputs + minor modelling_auto.py changes
vaibverm May 9, 2026
7bb398f
Working version with PagedAttention
vaibverm May 9, 2026
65b22ae
Minor fixes to specialization builder
vaibverm May 9, 2026
b73d5f2
Added support for Qwen2.5_VL PagedAttention
vaibverm May 12, 2026
760a41d
slot_id fix for Qwen2.5_VL PagedAttention decode
vaibverm May 13, 2026
9b59e9c
Added support for Qwen3_VL PagedAttention
vaibverm May 13, 2026
18352a8
Removed commented code corrected in rebase
vaibverm May 13, 2026
d2ca231
Minor fix for enum bug
vaibverm May 13, 2026
f9d7f15
Adding PagedAttetion support for Qwen3_VL_MOE
vaibverm May 13, 2026
3573adc
Adding PagedAttention specific test, unit tests and examples + lint/f…
vaibverm Jun 4, 2026
762de49
Lint/format fixes
vaibverm Jun 7, 2026
fb508e1
Rebasing with latest Qeff mainline
vaibverm Jul 22, 2026
a9e9696
Fix for get_compiler_dims returning 4 outputs
vaibverm Jul 24, 2026
6420a91
Fix the faulty blocking modes resolution in blocking_configurator.py
vaibverm Jul 27, 2026
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
54 changes: 52 additions & 2 deletions QEfficient/blocking/attention_blocking.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,26 +16,35 @@

from QEfficient.blocking.blocked_attention_forwards import (
blocked_bhqkv_attention_forward,
blocked_bhqkv_paged_attention_forward,
blocked_h_attention_forward,
blocked_h_mla_attention_forward,
blocked_hqkv_attention_forward,
blocked_hqkv_paged_attention_forward,
blocked_kv_attention_forward,
blocked_kv_mla_attention_forward,
blocked_kv_paged_attention_forward,
blocked_q_attention_forward,
blocked_qkv_attention_forward,
blocked_qkv_paged_attention_forward,
)


class BlockingMode(str, Enum):
NONE = ""
KV = "kv"
KV_PAGED = "kv_paged"

@kdulla kdulla Jul 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would be better to not changed BlockingMode and instead add a paged_attention_flag to AttentionBlockingConfig

Q = "q"
H = "h"
QKV = "qkv"
QKV_PAGED = "qkv_paged"
HQ = "hq"
HKV = "hkv"
HKV_PAGED = "hkv_paged"
HQKV = "hqkv"
HQKV_PAGED = "hqkv_paged"
BHQKV = "bhqkv"
BHQKV_PAGED = "bhqkv_paged"


@dataclass
Expand All @@ -52,15 +61,24 @@ def supports_blocked_kv(past_key_value: Optional[Cache]) -> bool:
return past_key_value is not None and hasattr(past_key_value, "read_only_blockedKV")


def supports_paged_attention_blocked_kv(past_key_value: Optional[Cache]) -> bool:

@anujgupt-github anujgupt-github Jul 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should paged behavior be selected by hasattr on the cache object rather than an explicit config signal?

return past_key_value is not None and hasattr(past_key_value, "read_only_pagedAttention")


_STRATEGIES: Dict[BlockingMode, Callable] = {

@kdulla kdulla Jul 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can create a second dictionary _STRATEGIES_PAGED that is checked if AttentionBlockingConfig has paged_attention flag True

BlockingMode.KV: blocked_kv_attention_forward,
BlockingMode.KV_PAGED: blocked_kv_paged_attention_forward,
BlockingMode.Q: blocked_q_attention_forward,
BlockingMode.H: blocked_h_attention_forward,
BlockingMode.QKV: blocked_qkv_attention_forward,
BlockingMode.QKV_PAGED: blocked_qkv_paged_attention_forward,
BlockingMode.HQ: blocked_hqkv_attention_forward,
BlockingMode.HKV: blocked_hqkv_attention_forward,
BlockingMode.HKV_PAGED: blocked_hqkv_paged_attention_forward,
BlockingMode.HQKV: blocked_hqkv_attention_forward,
BlockingMode.HQKV_PAGED: blocked_hqkv_paged_attention_forward,
BlockingMode.BHQKV: blocked_bhqkv_attention_forward,
BlockingMode.BHQKV_PAGED: blocked_bhqkv_paged_attention_forward,
}

_STRATEGIES_MLA: Dict[BlockingMode, Callable] = {
Expand All @@ -79,10 +97,17 @@ def past_key_value_update(
comp_ctx_lengths: Optional[torch.LongTensor] = None,
batch_index: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
block_table: Optional[torch.LongTensor] = None,

@kdulla kdulla Jul 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We only ever call this function when use_paged_kv_blocked is disabled, is this needed?

slot_id: Optional[torch.LongTensor] = None,
sliding_window: Optional[int] = None,
):
if past_key_value is not None:
cache_kwargs = {"batch_index": batch_index, "position_ids": position_ids}
cache_kwargs = {
"batch_index": batch_index,
"position_ids": position_ids,
"block_table": block_table,
"slot_id": slot_id,
}
if sliding_window is not None:
cache_kwargs.update(
{
Expand Down Expand Up @@ -110,6 +135,8 @@ def generic_blocked_attention_interface(
comp_ctx_lengths: Optional[torch.LongTensor] = None,
batch_index: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
block_table: Optional[torch.LongTensor] = None,
slot_id: Optional[torch.LongTensor] = None,
past_seen_tokens: Optional[int] = None,
non_blocked_forward: Callable = None,
score_mod: Optional[Callable] = None,
Expand All @@ -122,8 +149,29 @@ def generic_blocked_attention_interface(
blocking_config is not None and "kv" in blocking_config.mode and supports_blocked_kv(past_key_value)
)

use_paged_kv_blocked = (
blocking_config is not None
and "paged" in blocking_config.mode
and supports_paged_attention_blocked_kv(past_key_value)
)

if past_key_value is not None:
if use_kv_blocked and sliding_window is None:
if use_paged_kv_blocked and sliding_window is None:
cache_kwargs = {
"batch_index": batch_index,
"position_ids": position_ids,
"block_table": block_table,
"slot_id": slot_id,
}
if sliding_window is not None:

@anujgupt-github anujgupt-github Jul 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why separate handling? are you assuming that block size will always be > SWA size?

@vaibverm vaibverm Jul 28, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In the current implementation, PagedAttention hasn't been evaluated with sliding window attention yet. The reason is that in sliding window attention, different layers will require different max num_kv_blocks and how VLLM handles this heterogeneity hasn't been studied yet. This will be done in a future implementation.

@anujgupt-github anujgupt-github Jul 29, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Better to assert the path which is not supported instead of letting it fall through

@quic-hemagnih quic-hemagnih Jul 27, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This peice of code will never execute, as at line#159, sliding _window should be none to enter in the code leg.
Now at this line you are checking that sliding window is not NONE.

cache_kwargs.update(
{
"is_sliding": sliding_window is not None,
"sliding_window": past_key_value.sliding_window_len,
}
)
past_key_value.write_only_pagedAttention(key, value, module.layer_idx, cache_kwargs)
elif use_kv_blocked and sliding_window is None:
cache_kwargs = {
"batch_index": batch_index,
"position_ids": position_ids,
Expand All @@ -147,6 +195,8 @@ def generic_blocked_attention_interface(
comp_ctx_lengths=comp_ctx_lengths,
batch_index=batch_index,
position_ids=position_ids,
block_table=block_table,

@kdulla kdulla Jul 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this needed if we are not doing paged attention in this else condition?

slot_id=slot_id,
sliding_window=sliding_window,
)

Expand Down
Loading