create a new accuracy eval script for official README.md eval accuracy #3449
+248
−24
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
Contributor
Author
|
Stack from ghstack (oldest at bottom): |
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3449
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 1581808 with merge base 69ce0fd ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
vkuzo
added a commit
that referenced
this pull request
Dec 5, 2025
Summary: Creates a standalone eval script for generating accuracy metrics for quantization README.md, based on the HuggingFace model definition of LLaMa 3.1 8B Why new script? 1. the current `prod` script in https://github.com/pytorch/ao/blob/main/torchao/_models/llama/eval.py uses a custom model definition, this was pre-HF integration, it's better to use HF's model definition now 2. we have HummingBird scripts in https://github.com/pytorch/ao/tree/40c4f44677ae11166c3dcfbb9189cfa78789390c/.github/scripts/torchao_model_releases, but they seem pretty verbose and hard to use/modify 3. we have https://github.com/pytorch/ao/blob/main/benchmarks/_models/eval_hf_models.py, I copy-pasted and modified this for the current PR. The script above didn't work as is for various reasons, and also seemed to be hard to use/modify, for main README.md it's important to have a very simple standalone script. We should probably do a pass on the naming before landing. Future work: 1. add metrics for `int4_weight_only_hqq` (need to run on A100) 2. add metrics for `mxfp8` and `nvfp4` (need to run on B200) 3. make the parsing of logs automated 4. also add a similar script for performance benchmarks, using vllm 5. delete https://github.com/pytorch/ao/blob/main/torchao/_models/llama/ Test Plan: ``` // debug run on small model with-proxy time ./benchmarks/quantization/eval_accuracy_for_readme.sh facebook/opt-125m // real run with-proxy time ./benchmarks/quantization/eval_accuracy_for_readme.sh ``` Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 39c1d72 ghstack-comment-id: 3618394399 Pull-Request: #3449
meta-cla
bot
added
the
CLA Signed
vkuzo
added
the
topic: not user facing
vkuzo
added a commit
that referenced
this pull request
Dec 8, 2025
Summary: Creates a standalone eval script for generating accuracy metrics for quantization README.md, based on the HuggingFace model definition of LLaMa 3.1 8B Why new script? 1. the current `prod` script in https://github.com/pytorch/ao/blob/main/torchao/_models/llama/eval.py uses a custom model definition, this was pre-HF integration, it's better to use HF's model definition now 2. we have HummingBird scripts in https://github.com/pytorch/ao/tree/40c4f44677ae11166c3dcfbb9189cfa78789390c/.github/scripts/torchao_model_releases, but they seem pretty verbose and hard to use/modify 3. we have https://github.com/pytorch/ao/blob/main/benchmarks/_models/eval_hf_models.py, I copy-pasted and modified this for the current PR. The script above didn't work as is for various reasons, and also seemed to be hard to use/modify, for main README.md it's important to have a very simple standalone script. We should probably do a pass on the naming before landing. Future work: 1. add metrics for `int4_weight_only_hqq` (need to run on A100) 2. add metrics for `mxfp8` and `nvfp4` (need to run on B200) 3. make the parsing of logs automated 4. also add a similar script for performance benchmarks, using vllm 5. delete https://github.com/pytorch/ao/blob/main/torchao/_models/llama/ Test Plan: ``` // debug run on small model with-proxy time ./benchmarks/quantization/eval_accuracy_for_readme.sh facebook/opt-125m // real run with-proxy time ./benchmarks/quantization/eval_accuracy_for_readme.sh ``` Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 174b317 ghstack-comment-id: 3618394399 Pull-Request: #3449
jerryzh168
reviewed
Dec 8, 2025
| # note: | ||
| # * `int4_groupwise_hqq_weight_float8_rowwise_activation` doesn't work with dtype_map auto: https://gist.github.com/vkuzo/6b128681b628744d445c553cdeac8a85 | ||
| # * `int4_groupwise_hqq_weight_only` only works on A100 | ||
| for quant_recipe in float8_rowwise int4_groupwise_weight_float8_rowwise_activation int4_groupwise_hqq_weight_only int8_rowwise_weight_only int8_rowwise; do |
Contributor
There was a problem hiding this comment.
nit: int4_groupwise_weight_float8_rowwise_activation --> float8_rowwise_activation_int4_groupwise_weight to match the config name order?
Contributor
Author
There was a problem hiding this comment.
I'm matching the order in https://github.com/pytorch/ao/tree/main?tab=readme-ov-file#stable-workflows
overall if we want to standardize this everywhere sounds reasonable, IMO let's do that in a separate "rename-only" PR?
jainapurva
approved these changes
Dec 8, 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.
Summary:
Creates a standalone eval script for generating accuracy metrics for
quantization README.md, based on the HuggingFace model definition of
LLaMa 3.1 8B
Why new script?
prodscript inhttps://github.com/pytorch/ao/blob/main/torchao/_models/llama/eval.py
uses a custom model definition, this was pre-HF integration, it's better to use HF's model definition now
https://github.com/pytorch/ao/tree/40c4f44677ae11166c3dcfbb9189cfa78789390c/.github/scripts/torchao_model_releases,
but they seem pretty verbose and hard to use/modify
https://github.com/pytorch/ao/blob/main/benchmarks/_models/eval_hf_models.py,
I copy-pasted and modified this for the current PR. The script above
didn't work as is for various reasons, and also seemed to be hard to
use/modify, for main README.md it's important to have a very simple
standalone script.
We should probably do a pass on the naming before landing.
Future work:
int4_weight_only_hqq(need to run on A100)mxfp8andnvfp4(need to run on B200)Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags: