fixing reference implementation.

[jjsjann123]

Shouldn't have sliced swizzled & padded data when copying it back to the original buffer. The issue was noticed when I try to validate the layout op in #5198

This unfortunately didn't affect the threshold for result validation.

[jjsjann123]

!test

[github-actions]

Description

• Fixed incorrect slicing of swizzled and padded data

• Corrected block scale offset calculation for alignment

• Ensured full tensor is returned without truncation

• Aligned scale factor layout with hardware requirements

---

Changes walkthrough 📝

Relevant files

Bug fix

narrow_precision.py Fix tensor slicing and block scale alignment                          tests/python/direct_utils/narrow_precision.py Removed slicing of reshaped swizzled data to retain full tensor Fixed block scale range calculation to align with 128-byte boundaries Updated linear_to_swizzled_128_4 to return full padded tensor Corrected r_sf computation using ceiling division for proper alignment +2/-2

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Incorrect Slice Handling The removal of the slice [:mn, :sf_k] when returning the reshaped tensor may lead to incorrect data being used downstream, as it no longer restricts the output to the original dimensions. This could propagate padded values into contexts expecting only valid data. return tmp.transpose(1, 3).reshape(mn_padded, k_padded) Block Scale Range Miscalculation The new calculation for r_sf using (r - l + 127) // 128 * 128 may over-allocate the block scale range, potentially writing beyond intended bounds when used with linear_to_swizzled_128_4, especially if the original r_sf was tightly bounded. r_sf = l_sf + (r - l + 127) // 128 * 128