Fix CUDA GatherBlockQuantized block indexing

[Copilot]

Description

• Compute block IDs per quantized-axis row:

  block_id = row * blocks_per_row + column / block_size;

• Add CUDA regression coverage for a 4-bit, non-block-aligned quantized axis.

Motivation and Context

CUDA previously derived the scale and zero-point block from the flat data index. Non-block-aligned quantized axes therefore crossed row boundaries and silently produced incorrect results.

• Fixes GatherBlockQuantized CUDA kernel uses the flat index for the block id, giving wrong results when the quantized axis is not a multiple of block_size #32663

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

Packed zero-point indexing remains incorrect for rows with a non-pack-aligned block count, and the CUDA test can silently run on CPU.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Fixes CUDA block indexing for non-aligned quantized-axis rows and adds regression coverage.

Changes:

• Computes scale block IDs per quantized-axis row.
• Passes the unpacked quantized-axis dimension to the CUDA kernel.
• Adds a 4-bit CUDA regression test.

File summaries

File	Description
gather_block_quantized.cc	Supplies the logical quantized-axis dimension.
gather_block_quantized.cu	Corrects row-aware block indexing.
gather_block_quantized.cuh	Updates CUDA kernel parameters.
gather_block_quantized_op_test.cc	Adds non-aligned-axis coverage.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.