Add Multihead-style wrapper and backend factory #28
Conversation
This file contains hidden or 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
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
| key: torch.Tensor, | ||
| value: torch.Tensor, | ||
| key_padding_mask: Optional[torch.Tensor] = None, | ||
| need_weights: bool = False, |
Contributor
There was a problem hiding this comment.
Missing average_attn_weights parameter despite claiming to mirror MultiheadAttention’s forward signature.
Prompt for AI agents
Address the following comment on stream_attention/core/multihead_attention.py at line 67:
<comment>Missing average_attn_weights parameter despite claiming to mirror MultiheadAttention’s forward signature.</comment>
<file context>
@@ -0,0 +1,147 @@
+"""Multihead-style wrapper around FusedOnlineAttention.
+
+This module provides :class:`StreamMultiheadAttention`, mirroring the
+constructor and forward signature of :class:`torch.nn.MultiheadAttention`
+while delegating the actual attention computation to
+:class:`~stream_attention.core.fused_online_attention.FusedOnlineAttention`.
+
+A convenience factory :func:`create_stream_attention` is also provided to
+instantiate the optimal backend automatically. If Triton kernels are
</file context>
| if not self.batch_first: | ||
| attn_output = attn_output.transpose(0, 1) | ||
| if need_weights: | ||
| return attn_output, None |
Contributor
There was a problem hiding this comment.
need_weights=True returns None for weights; either compute weights or raise to avoid silent API mismatch.
Prompt for AI agents
Address the following comment on stream_attention/core/multihead_attention.py at line 109:
<comment>need_weights=True returns None for weights; either compute weights or raise to avoid silent API mismatch.</comment>
<file context>
@@ -0,0 +1,147 @@
+"""Multihead-style wrapper around FusedOnlineAttention.
+
+This module provides :class:`StreamMultiheadAttention`, mirroring the
+constructor and forward signature of :class:`torch.nn.MultiheadAttention`
+while delegating the actual attention computation to
+:class:`~stream_attention.core.fused_online_attention.FusedOnlineAttention`.
+
+A convenience factory :func:`create_stream_attention` is also provided to
+instantiate the optimal backend automatically. If Triton kernels are
</file context>
| attn_output = attn_output.transpose(0, 1) | ||
| if need_weights: | ||
| return attn_output, None | ||
| return attn_output |
Contributor
There was a problem hiding this comment.
Return type does not match the annotated Tuple when need_weights is False; return a 2-tuple consistently to avoid API inconsistency.
Prompt for AI agents
Address the following comment on stream_attention/core/multihead_attention.py at line 110:
<comment>Return type does not match the annotated Tuple when need_weights is False; return a 2-tuple consistently to avoid API inconsistency.</comment>
<file context>
@@ -0,0 +1,147 @@
+"""Multihead-style wrapper around FusedOnlineAttention.
+
+This module provides :class:`StreamMultiheadAttention`, mirroring the
+constructor and forward signature of :class:`torch.nn.MultiheadAttention`
+while delegating the actual attention computation to
+:class:`~stream_attention.core.fused_online_attention.FusedOnlineAttention`.
+
+A convenience factory :func:`create_stream_attention` is also provided to
+instantiate the optimal backend automatically. If Triton kernels are
</file context>
| query=q, | ||
| key=k, | ||
| value=v, | ||
| causal=True, |
Contributor
There was a problem hiding this comment.
Forces causal masking for all calls, diverging from MultiheadAttention’s default behavior and potentially breaking non-causal use cases.
Prompt for AI agents
Address the following comment on stream_attention/core/multihead_attention.py at line 99:
<comment>Forces causal masking for all calls, diverging from MultiheadAttention’s default behavior and potentially breaking non-causal use cases.</comment>
<file context>
@@ -0,0 +1,147 @@
+"""Multihead-style wrapper around FusedOnlineAttention.
+
+This module provides :class:`StreamMultiheadAttention`, mirroring the
+constructor and forward signature of :class:`torch.nn.MultiheadAttention`
+while delegating the actual attention computation to
+:class:`~stream_attention.core.fused_online_attention.FusedOnlineAttention`.
+
+A convenience factory :func:`create_stream_attention` is also provided to
+instantiate the optimal backend automatically. If Triton kernels are
</file context>
| key=k, | ||
| value=v, | ||
| causal=True, | ||
| attention_mask=attn_mask_inner, |
Contributor
There was a problem hiding this comment.
Attention dropout is never applied because dropout_p is not forwarded to FusedOnlineAttention during training.
Prompt for AI agents
Address the following comment on stream_attention/core/multihead_attention.py at line 100:
<comment>Attention dropout is never applied because dropout_p is not forwarded to FusedOnlineAttention during training.</comment>
<file context>
@@ -0,0 +1,147 @@
+"""Multihead-style wrapper around FusedOnlineAttention.
+
+This module provides :class:`StreamMultiheadAttention`, mirroring the
+constructor and forward signature of :class:`torch.nn.MultiheadAttention`
+while delegating the actual attention computation to
+:class:`~stream_attention.core.fused_online_attention.FusedOnlineAttention`.
+
+A convenience factory :func:`create_stream_attention` is also provided to
+instantiate the optimal backend automatically. If Triton kernels are
</file context>
Co-authored-by: cubic-dev-ai[bot] <191113872+cubic-dev-ai[bot]@users.noreply.github.com>
Owner
Author
|
@cursoragent fix those issues |
|
I updated the
5 tasks completed
Learn more about Cursor Agents |
Co-authored-by: alphacr792 <[email protected]>
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.
Summary
StreamMultiheadAttentionwrappingFusedOnlineAttentionwith ann.MultiheadAttention-like APIcreate_stream_attentionfactory that chooses Triton fused or PyTorch SDPA backendTesting
pytesthttps://chatgpt.com/codex/tasks/task_e_68ab6a1e403c8322b7d0cbace7d428ec
Summary by cubic
Adds a MultiheadAttention-style layer backed by the fused online attention kernel, plus a factory that auto-selects the best backend. This makes it easy to drop in a familiar API while using Triton when available, with a safe SDPA fallback.
New Features
Migration