Add static FP8 attention support #1045
Merged
+338
−68
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
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR adds static FP8 attention support to the auto-round library. The implementation enables quantization of attention mechanisms using FP8 format, building on existing KV cache quantization infrastructure.
Key Changes
- Introduces
QuantizedAttentionImplmodule to handle FP8 quantized attention operations - Refactors shared quantization utilities into a new
experimental/utils.pymodule - Updates test suite to validate FP8 attention quantization and correct scale tensor shapes
Reviewed Changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated 7 comments.
Show a summary per file
| File | Description |
|---|---|
| test/test_cpu/test_export.py | Adds test for static FP8 attention, updates model path and corrects expected scale tensor shapes |
| auto_round/experimental/utils.py | New utility module consolidating shared FP8 quantization and module manipulation functions |
| auto_round/experimental/qmodules/fp8_static.py | Preserves original dtype in forward pass for FP8 linear operations |
| auto_round/experimental/kv_cache.py | Refactors utilities to new module, initializes scale parameters, removes debug logging |
| auto_round/experimental/attention.py | New module implementing hooked attention mechanism for FP8 quantization |
| auto_round/compressors/base.py | Adds static_attention_dtype parameter and integrates attention quantization context |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
n1ck-guo
reviewed
Nov 24, 2025
auto_round/experimental/utils.py
Outdated
| from auto_round.utils import logger | ||
|
|
||
|
|
||
| def fp8_per_tensor_qdq( |
Contributor
There was a problem hiding this comment.
The function name is a bit strange.I suggest changing it to a more easily understood name.
n1ck-guo
reviewed
Nov 24, 2025
| enable_deterministic_algorithms = kwargs.pop("enable_deterministic_algorithms", False) | ||
| self.momentum = kwargs.pop("momentum", 0.0) | ||
| static_kv_dtype = kwargs.pop("static_kv_dtype", None) | ||
| static_attention_dtype = kwargs.pop("static_attention_dtype", None) |
Contributor
There was a problem hiding this comment.
Does __main__.py also need add this parameter?
Contributor
|
LGTM |
n1ck-guo
approved these changes
Nov 24, 2025
chensuyue
pushed a commit
that referenced
this pull request
Nov 24, 2025
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.
resolve #938
LLMC format export is not supported for now, as vLLM loading is still a work in progress.
But it can be used on HPU.