Correctly apply global scale in nvfp4 quantization. #30467
+71
−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
Currently, the global scale is not applied consistently to the input tensor, leading to all values being rounded to zeroes in most cases when using the global scale logic from the test. The existing tests did not catch this because they use matrix multiplications on dequantized tensors as the reference. This ended up just comparing 0@0 to 0@0. Instead of adjusting the matmul tests, I opted to add test coverage for quantization logic directly, since that doesn't require carefully chosen tolerances in most cases.
jreiffers
commented
Jul 24, 2025
| SCALE_MAX = dtypes.finfo(config.scale_type).max.astype(x.dtype) | ||
|
|
||
| scales_q = jnp.clip(scales / config.global_scale, 0, SCALE_MAX) | ||
| x /= config.global_scale | ||
| scales_q = jnp.clip(get_scales_per_block(x), 0, SCALE_MAX) | ||
| scales_q = lax.optimization_barrier(scales_q.astype(config.scale_type)) | ||
| scaled_x = x / scales_q.astype(np.float32) |
There was a problem hiding this comment.
(This x is where we previously didn't apply the global scale)
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Outdated suggestions cannot be applied.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.
Currently, the global scale is not applied consistently to the input tensor, leading to all values being rounded to zeroes in most cases when using the global scale logic from the test.
The existing tests did not catch this because they use matrix multiplications on dequantized tensors as the reference. This ended up just comparing 0@0 to 0@0.
Instead of adjusting the matmul tests, I opted to add test coverage for quantization logic directly, since that doesn't require carefully chosen tolerances in most cases.