## Fix GPTQ producing invalid QArray when subchannel quantization not specified

[copybara-service]

Fix GPTQ producing invalid QArray when subchannel quantization not specified

Summary

Fixes a bug where gptq_core.quantize_weight creates inadvertent subchannel quantization that can produce invalid QArrays, even when the user did not request subchannel quantization.

Problem

When subchannel quantization is not specified (tiled_axes is empty), quantize_weight defaults the groupsize to rows:

groupsize = how.tiled_axes.get(1, rows)

This causes scales to be computed every rows columns, creating ceil(columns/rows) scale groups. When columns is not evenly divisible by rows, this produces an invalid QArray because the resulting scale shape violates the QArray contract (all(qvalue_dim % scale_dim == 0 for each axis))

When subchannel quantization is not specified, the scale should have shape (rows, 1) (per-channel quantization), not (rows, num_groups) (sub-channel).

Solution

This change modifies the default groupsize from rows to columns:

groupsize = how.tiled_axes.get(1, columns)

This ensures:

• If subchannel is specified: use the user's groupsize
• If subchannel is not specified: use columns as groupsize, producing per-channel quantization with scale.shape = (rows, 1)