Triton Dequant Kernel Out-of-Bounds Read Fix (PR #2949)

Summary of Key Vulnerability Information Vulnerability Overview Vulnerability Type: Out-of-Bounds Read Location: Triton dequantization kernel (the kernel in ) Root Cause: The kernel resolves each group as , then indexes into and without an upper-bound check. When the for any of the groups exceeds , the kernel wraps around and reads past-the-buffer , causing the CUDA context (D55) to crash and potentially reading adjacent GPU memory. Trigger Condition: When loading an unquantized GPTQ checkpoint, falls outside the range. Scope of Impact Affected Component: Triton dequantization path ( kernel) Impact: - Crashes the CUDA context (D55) - Potential reading of adjacent GPU memory (information disclosure risk) - Affects only the Triton path; the Torch backend fails safely on the same input due to the gather operation on , meaning only the Triton path reads out-of-bounds memory. Fix Fix Strategy: Validate at load time, rejecting values outside the range. Specific Implementation: - Validate in , rejecting values outside the range. - This range is exactly the range where the kernel's negative wrap-around can be resolved safely and where the Torch backend's gather operation will accept. - Since the guard rejects inputs that already fail safely in the Torch path, it is reasonable to place the check at load time (rather than in every forward pass). - Placing the check in a guard on the loader side (rather than inside the kernel) is because rejecting at load time prevents bad from reaching the device; a mask on the kernel side may be too deep in the defense. POC / Test Code Other Information PR Number: #2949 Status: Merged Committer: Anai-Guo Reviewer: Qubitum (Contributor) Merge Commit:

Goal Reached Thanks to every supporter — we hit 100%!

Triton Dequant Kernel Out-of-Bounds Read Fix (PR #2949)

More from this source