Skip to content

Enable slicing for fp8 FusedSDPA #1285

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

Merged
czhu15 merged 5 commits into from
Apr 8, 2026
Merged

Enable slicing for fp8 FusedSDPA #1285

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e3f89b2
Revert "Use Boolean attention mask (#1032)"
yangulei Apr 2, 2026
f3ff913
use finite numbers for attention mask value
yangulei Apr 2, 2026
1a004e0
enable slicing for fp8 FusedSDPA
yangulei Apr 1, 2026
6c3c91a
fix gc error
yangulei Apr 1, 2026
9f186b8
use detach() instead of clone() to solve the accuracy issue
yangulei Apr 7, 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
7 changes: 7 additions & 0 deletions vllm_gaudi/attention/backends/hpu_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -454,6 +454,13 @@ def __init__(
HPUFusedSDPA = kernels.fsdpa()
self.fused_scaled_dot_product_attention = None if HPUFusedSDPA is None \
else ModuleFusedSDPA(HPUFusedSDPA)
try:
from habana_frameworks.torch.hpex.kernels import fp8_fused_sdpa
if self.enable_fp8_attn:
self.fused_scaled_dot_product_attention = ModuleFP8FusedSDPA(fp8_fused_sdpa)
except ImportError:
pass

self.prefill_impl = get_config().prompt_attn_impl
self.use_contiguous_pa = get_config().use_contiguous_pa
self.use_merged_prefill = get_config().merged_prefill
Expand Down
9 changes: 3 additions & 6 deletions vllm_gaudi/extension/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,12 +349,9 @@ def _naive_prompt_attention(query: torch.Tensor,
htcore.mark_step()
attn_weights.add_(position_bias)
if attn_bias is not None:
if attn_bias.dtype == torch.bool:
attn_weights = attn_weights.masked_fill(~attn_bias, float("-inf"))
else:
if attn_weights.dtype != attn_bias.dtype:
attn_bias = attn_bias.to(dtype=attn_weights.dtype)
attn_weights.add_(attn_bias)
if attn_weights.dtype != attn_bias.dtype:
attn_bias = attn_bias.to(dtype=attn_weights.dtype)
attn_weights.add_(attn_bias)
if sinks is not None:
sink = sinks.reshape(1, -1, 1, 1).expand(query.shape[0], -1, query.shape[-2], -1)
if query_heads != kv_heads:
Expand Down
Loading