Skip to content

Draft: Add FlashAttention online merge in Unified Attention #785

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

Draft
kzawora-intel wants to merge 10 commits into
base: main
Choose a base branch
from
+679 −117
Draft

Draft: Add FlashAttention online merge in Unified Attention #785

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
689431e
Handle spec decode optionals in unified batch
kzawora-intel Jan 7, 2026
0068dd0
[WIP] Chunked Shared Attention with Dense Biases
kzawora-intel Jan 7, 2026
b4751ce
cleanup
kzawora-intel Jan 7, 2026
1f0f17a
remove FA changes
kzawora-intel Jan 7, 2026
019a1cf
pad shared blocks to chunk size
kzawora-intel Jan 7, 2026
2d708e4
Add FA online merge to UA
kzawora-intel Jan 7, 2026
ef3c73f
Unified Attention - multi-step low-level profiling
kzawora-intel Jan 8, 2026
26cfc38
Merge remote-tracking branch 'origin' into origin/main
kzawora-intel Jan 14, 2026
527304e
cleanup flags + add graph splitting
kzawora-intel Jan 14, 2026
8a5eac0
reduce chunk size, fix mla
kzawora-intel Jan 14, 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
2 changes: 1 addition & 1 deletion vllm_gaudi/attention/backends/hpu_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1030,7 +1030,7 @@ def __init__(
self.qk_head_dim = qk_head_dim
self.v_head_dim = v_head_dim
self.kv_b_proj = kv_b_proj # Used to expand latent → full KV in causal path

self.use_online_merge = get_config().unified_attn_online_merge
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
self.latent_cache_k = VLLMKVCache() if not self.enable_fp8_attn \
Expand Down
20 changes: 17 additions & 3 deletions vllm_gaudi/extension/features.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,21 @@ def get_experimental_flags():
return to_dict(flags)


def unified_attn_dev_flags():
flags = [
Value('unified_attn_dense_shared_bias', True),
Value('unified_attn_chunked_shared_attn', True),
Value('unified_attn_online_merge', True),
Value('unified_attn_shared_attn_chunk_size', 64),
Value('unified_attn_split_graphs', Enabled('unified_attn_online_merge')),
Value(
'unified_attn_softmax_fa2',
All(VersionRange(">=1.24.0.279"), Enabled('unified_attn'), Kernel(softmax_fa2), Hardware('gaudi3'),
Not(Enabled('unified_attn_chunked_shared_attn')))),
]
return flags


def get_features():
supported_attn_impls = ['flex_impl', 'fsdpa_impl', 'naive_impl']
bucketing_strategies = ['exponential_bucketing', 'linear_bucketing']
Expand Down Expand Up @@ -90,12 +105,11 @@ def get_features():
Value('dynamic_shapes_compilation', True, env_var='VLLM_T_COMPILE_DYNAMIC_SHAPES', env_var_type=boolean),
Value('fullgraph_compilation', False, env_var='VLLM_T_COMPILE_FULLGRAPH', env_var_type=boolean),
Value('unified_attn', False),
Value('unified_attn_softmax_fa2',
All(VersionRange(">=1.24.0.279"), Enabled('unified_attn'), Kernel(softmax_fa2), Hardware('gaudi3'))),
*unified_attn_dev_flags(),
Value('scale_adjustment', True, env_var='VLLM_SCALE_ADJUSTMENT', env_var_type=boolean),
Value('flatten_input', Any(ModelType('qwen3_moe'), ModelType('granitemoe'), ModelType('glm4_moe'))),
Value('unified_attn_shared_cache_ratio',
0.8,
1,
env_var='VLLM_UNIFIED_ATTENTION_SHARED_CACHE_RATIO',
env_var_type=float),
Value('high_level_profiler_enabled', False, env_var='VLLM_PROFILER_ENABLED', env_var_type=boolean),
Expand Down
Loading
Loading