Skip to content

Disable DIV with broadcast on OpenVINO GPU backend#164

Merged
zhaixuejun1993 merged 1 commit into
ravi9:dev_backend_openvinofrom
zhaixuejun1993:xuejun/arch-llama-mpt
May 15, 2026
Merged

Disable DIV with broadcast on OpenVINO GPU backend#164
zhaixuejun1993 merged 1 commit into
ravi9:dev_backend_openvinofrom
zhaixuejun1993:xuejun/arch-llama-mpt

Conversation

@zhaixuejun1993
Copy link
Copy Markdown
Collaborator

This pull request makes targeted improvements to the dynamic dimension handling and operator support logic in the OpenVINO backend of the codebase. The main changes focus on making error handling more robust in the decoder and adding better support for division operations, especially when broadcasting is involved on GPU devices.

Dynamic dimension handling improvements:

  • Replaced assertions with warnings in GgmlOvDecoder::compute_node_dynamic_dims() to prevent hard failures when dynamic dimension mismatches are detected, instead setting the dynamic dimension to -1 and logging a warning.
  • Added a warning and fallback for cases where the dynamic dimension for a CONT node cannot be uniquely determined, instead of asserting and halting execution.

Operator support enhancements:

  • Updated is_op_unsupported_case() to mark division operations (GGML_OP_DIV) as unsupported when broadcasting is required and the device is GPU, due to known issues with fused GPU kernels producing invalid results. This prevents problematic DIV operations from running on GPU until the kernel is reliable.## Overview

Additional information

Requirements

@zhaixuejun1993 zhaixuejun1993 merged commit dda650b into ravi9:dev_backend_openvino May 15, 2026
3 of 12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant