fix(BACKEND-ROCM): gfx1200 -Werror build break in rocm_paged_attn.hip (unused m_split, g)
Nobody has claimed this yet.
Assessment
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Newbie friendliness
- 70/100
- Issue type
- Bug
- Clarity
- Clearly specified
- Activity status
- Active
- Tech stack
- cmake, cpp
- Domain
- backend, build-system
Research direction
Inspect src/vt/rocm/rocm_paged_attn.hip around lines 1207 and 1227, then reproduce the gfx1200 build with the provided CMake and make commands. Confirm the unused-variable diagnostics are addressed without changing the operator behavior, and verify that test_backend_cross_device builds successfully under ROCm.
Written by the indexing model from the issue text.
Description
Row: -
What's broken
main (current head, via commit b53e20f85e "Implement ROCm split-KV
attention decode (issue #845)", authored 2026-09-17) fails to build for
gfx1200 under the project's -Werror HIP arm. Two unused-variable
diagnostics in src/vt/rocm/rocm_paged_attn.hip:
src/vt/rocm/rocm_paged_attn.hip:1227:11: error: unused variable 'm_split' [-Werror,-Wunused-variable]
float m_split = -INFINITY;
^~~~~~~
src/vt/rocm/rocm_paged_attn.hip:1207:17: error: unused variable 'g' [-Werror,-Wunused-variable]
const int64_t g = h / (hq / num_kv_heads);
^
Compiler: ROCm's bundled Clang 22.0.0 (/opt/rocm/lib/llvm/bin/clang++),
target gfx1200.
How found
Discovered while independently reproducing the operator's gfx1200 toggle
matrix for PR #3036 (KERNEL-QUANT-CIQ-GEMM-ROCM's cooperative-tile WMMA
arms). That PR's own changed files are unaffected — git diff --stat between
the row branch's pre-merge tip and its current head (after merging main)
touches only src/vt/rocm/rocm_grouped_gemm.hip and
tests/vt/test_backend_cross_device.cpp. rocm_paged_attn.hip is untouched
by #3036; the break is inherited purely from the main merge.
git blame on both lines attributes them to b53e20f85e alone.
Impact
Any current gfx1200 build (this row's toggle matrix included) fails before a
single device test can run. Whatever CI runs for this repo does not build
ROCm/HIP (no such check appears on recent PRs, #3036 included), so this
never surfaced there — only a real gfx1200 box building current main
reproduces it.
Repro
git clone https://github.com/mudler/vllm.cpp && cd vllm.cpp
mkdir build-hip && cd build-hip
cmake .. -DVLLM_CPP_ROCM=ON -DCMAKE_HIP_ARCHITECTURES=gfx1200
make -j16 test_backend_cross_device
Host used: isravale, RX 9060 XT (gfx1200), ROCm 7.2.3.
Owed
No row claims this yet. Per AGENTS.md "An issue you do not fix in the same
flow has to say who owns it," this needs either an owning row's Row: line
or a spec's ## Owed entry — filed as - because neither exists yet; needs
triage to pick an owner (most likely folded into BACKEND-ROCM or a fresh
row for the gfx1200 -Werror arm).
FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-sonnet-5 [Claude Code]
- Dominant language
- C++
- Stars
- 423
- Forks
- 53
- Avg merge
- 19h 7m
- Merged PRs (30d)
- 337
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from mudler/vllm.cpp
-
Difficulty 1/5 Under an hour Newbie friendliness 88/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 84/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
-
sync(UPSTREAM-SYNC-HEADPIN): step6-c1a credits the version read to a job whose own read returned 1 Open
Difficulty 2/5 1-3 hours Newbie friendliness 84/100
-
test_cpu_x86_llamacpp_floor reds as a function of machine load, on branches that never touched it Open
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
Similar issues
-
Difficulty 1/5 Under an hour Newbie friendliness 90/100
AXERA-TECH/ax-llm#77 ·
-
Difficulty 1/5 Under an hour Newbie friendliness 90/100
games-on-whales/wolf#509 ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
-
bug-unconfirmed
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
NVIDIA/cuda-samples#453 ·