[XNNPACK] Serialize weights as fp16 rather than fp32 #9753
Merged
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9753
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit cb31420 with merge base ce74f8e ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
digantdesai
reviewed
Mar 31, 2025
| @@ -368,7 +368,7 @@ def define_tensor( # noqa: C901 | |||
| constant data. If used along with convert_to_nhwc, this | |||
| swap will happen before converting to nhwc. | |||
| quant_params: Quantization meta data for this tensor, None if it is not quantized | |||
| fp32_static_weights: XNN_FLAG_FP32_STATIC_WEIGHTS for fp16 conv | |||
| force_fp32: forces tensor to be serialize as fp32, used for bias of dynamically quantized ops | |||
There was a problem hiding this comment.
s/fp32_static_weight/force_fp32 - seems a little too vague if you ask me.
digantdesai
approved these changes
Mar 31, 2025
kirklandsign
pushed a commit
that referenced
this pull request
Apr 11, 2025
### Summary Previously we've used FP32_STATIC_WEIGHTS flag in xnnpack to coerce fp32 weights into fp16 for linear and conv. This allowed us to mimc fp16 computation because the weights would be converted and packed as fp16 at runtime. However, this means we lose the benefit of the smaller .pte file because the weights are serialized as fp32 rather than fp16. Additionally, we still have to load the weights as fp32, since they are converted at runtime. This has some poor effects on performance ### Test plan ``` python -m unittest backends.xnnpack.test.ops.test_linear.TestLinear.test_fp16_linear python -m unittest backends.xnnpack.test.ops.test_linear.TestLinear python -m unittest backends.xnnpack.test.ops.test_conv2d.TestConv2d ``` Llama 3.2 with bf16 weights: Before: ``` -rw-r--r-- 1 maxren staff 5468937344 Mar 28 17:00 llama3_2_fp16_direct_convert_runtime.pte ``` After: ``` -rw-r--r-- 1 maxren staff 2997443712 Mar 28 16:57 llama3_2_fp16_direct_convert_runtime.pte ```
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
Previously we've used FP32_STATIC_WEIGHTS flag in xnnpack to coerce fp32 weights into fp16 for linear and conv. This allowed us to mimc fp16 computation because the weights would be converted and packed as fp16 at runtime. However, this means we lose the benefit of the smaller .pte file because the weights are serialized as fp32 rather than fp16. Additionally, we still have to load the weights as fp32, since they are converted at runtime. This has some poor effects on performance
Test plan
Llama 3.2 with bf16 weights:
Before:
After: