Eval hf models using lm_eval #2179
Open
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/ao/2179
Note: Links to docs will display an error until the docs builds have been completed. ❌ 7 New FailuresAs of commit e14c186 with merge base 7854249 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
jerryzh168
reviewed
May 6, 2025
torchao/_models/llama/eval_hf.py
Outdated
| # print("Response:", output_text[0][len(prompt):]) | ||
|
|
||
|
|
||
| def model_throughput( |
There was a problem hiding this comment.
nit: maybe you can add a memory measurement as well
jainapurva
added
topic: not user facing
HDCharles
reviewed
May 22, 2025
|
|
||
| ## Eval on Llama 3.1 8B and Llama 3.2 3B | ||
|
|
||
| We use lm-eval tasks for evaluating TorchAO Quantization APIs on HuggingFace models. The results are in the table below: |
There was a problem hiding this comment.
would specify what task this is for.
HDCharles
reviewed
May 22, 2025
| | Llama 3.1 8B | int8dq | 60.01 | 78.82 | 7.45 | 8.03 | | ||
| | Llama 3.1 8B | float8wo | 59.83 | 78.61 | 7.37 | 8.03 | | ||
| | Llama 3.1 8B | float8dq (PerRow) | 59.86 | 78.57 | 7.41 | 8.04 | | ||
| | Llama 3.1 8B | float8dq (PerTensor) | 59.95 | 78.66 | 7.42 | 8.03 | |
There was a problem hiding this comment.
per tensor is more accurate than per row quantization?
HDCharles
reviewed
May 22, 2025
| @@ -291,6 +292,15 @@ def string_to_config( | |||
| else: | |||
| granularity = PerTensor() | |||
| return Float8DynamicActivationFloat8WeightConfig(granularity=granularity) | |||
| if "gemlitewo" in quantization: | |||
There was a problem hiding this comment.
gemlite can be 4 bit or 8 bit, should probably specify that this is for 4 bit
https://github.com/pytorch/ao/blob/main/torchao/quantization/quant_api.py#L968
HDCharles
approved these changes
May 22, 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.
Using the huggingface and torchao integration, we can directly quantize hf models using torchao quantization techniques. The current PR adds a flow for quantizing an hf model using transformer's TorchAOConfig module with torchao's quant techniques, followed by evaluation of the quantized model on different lm-eval tasks.
Eval results for llama3.1-8B and 3.2-3B are added to torchao/_models/README.md