[Engine] qsa_rows_bench: --ctx below 32768 makes the dense step read outside the page table (Xid 31)

Open Beginner friendly
#6,209 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 5 days

Nobody has claimed this yet.

Assessment

Difficulty
2/5
Estimated time
1-3 hours
Newbie friendliness
75/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
cpp
Domain
testing

Research direction

Read src/kernels/qsa_rows_bench.cpp around lines 138 and 207 to trace how the dense window width and cell ids are built. Reproduce with qsa_rows_bench --rows --iters 100 --ctx 16384 --max-context 262144; verify the fix keeps the dense step within the context and that the 16K run completes without an illegal memory access.

Written by the indexing model from the issue text.

Description

Problem. qsa_rows_bench with a context under 32768 cells faults on the GPU. The dense step's window is fixed at
32768 cells, so at a shorter context its cell ids start below 0, and nothing bounds them.

Repro (RTX 5080, sm_120, driver 615.71.09, CUDA 13.3, Release build of the qsa_rows_bench target; the bench file is
identical to main):

qsa_rows_bench --rows --iters 100 --ctx 16384 --max-context 262144

Output, exit code 2:

qsa_rows_bench on device 0 (NVIDIA GeForce RTX 5080, 12.0), 100 iterations: context 16384, capacity 262144 cells (max_blocks 65538), selection cap 2051, int8 pools over 16452 cells
  selections: width 2051 a row; union of 2 rows 2360 cells (1.15x one row) union of 3 rows 2541 cells (1.24x one row) union of 4 rows 2608 cells (1.27x one row)
qsa_rows_bench: warm: an illegal memory access was encountered

Kernel log at the same moment:

NVRM: Xid (PCI:0000:01:00): 31, name=qsa_rows_bench, intr 00000000. MMU Fault: ENGINE GRAPHICS GPC6 GPCCLIENT_T1_5 faulted @ 0x220_cddf1000. Fault is of type FAULT_PDE ACCESS_TYPE_VIRT_READ

Cause (unconfirmed static analysis):

  • src/kernels/qsa_rows_bench.cpp:207 fills the dense window's ids as ctx + t + 1 - dense_win + j, with
    dense_win = 32768 (line 138).
  • At --ctx 16384 the first 16383 ids of each row are negative.
  • The batched decode attention reads p.page_table[cell / page_size] (src/kernels/cuda/qsa_decode_attn.cu:206, and
    418 and 618 in the other variants) with no lower bound, so the read lands before the page table.
  • The engine never passes negative ids, so the kernel itself is fine. The missing bounds check is in the bench.
  • The default --ctx 131072 and any context of 32768 or more stay in range.

Suggested fix: clamp the dense window to the context (min(dense_win, ctx + t + 1) cells a row, with the step
record's width to match), or skip the dense step when ctx + 1 < dense_win.

Validation:

  • Run: the 16K case above, once (it faulted).
  • Not run: 64K / 131K / 200K, --no-dense at 16K, and any fix. We will not run the bench again until it is fixed.

Measured and drafted with Claude Code.

Dominant language
C++
Stars
11.6k
Forks
1k
Avg merge
18d 13h
Merged PRs (30d)
4

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from Niko1221/Strata

All issues in Niko1221/Strata

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.