AMD/HIP (gfx1100, RX 7900 XTX): engine loads the Coder pack fully, then verify: launch: unspecified launch failure on the first request — or hangs before ready. No working configuration.

Engine 0.1.29, build from source on Linux, HIP backend, RX 7900 XTX (gfx1100). The build succeeds, the pack prepares, the engine loads the whole model — and it never serves a token.

What happens

Three configurations, all on the same box and pack:

1. --spec 4 (what setup.sh --backend hip wrote). The server comes up and /v1/models reports the model loaded. The first chat request kills the engine:

strata generate: PCIe probe: 7.0 GB/s host->device -> pcie_frac 0.15 (default 0.55)
strata verify: window up to 6 tokens, 74.0 MiB of device buffers
strata verify: captured the 6-token window (upload no error, sync no error)
strata serve: verify: launch: unspecified launch failure

the server then reports the engine stopped unexpectedly (exit code 1), done: 0 tokens in 5 s (0.0 tok/s) (error, cancel=False), and a retry gets the engine stopped unexpectedly (exit code 1) again.

2. --spec 2 (the smallest window the pack accepts). Loads to completion — 644 MiB of VRAM free with everything loaded, session is up (engine 0.1.29) — then never signals ready: the API port stays closed and the log stops.

3. --spec 4 plus the flags docs/AMD_HIP.md prescribes for AMD (--prefill 512 --pool-workers 15 --vram-reserve-mib 1024 --adapt-every 0 --pcie-frac 0). Same hang as (2), with 946 MiB of VRAM free with everything loaded.

--spec 0 cannot be used to test around the verify path: the engine rejects it with

is a native (IQ) pack: it needs --native SHARD1, --spec T (T >= 2) and --prefill CHUNK

so there is no configuration on this hardware that reaches a generated token.

Why it reads as the verify kernel rather than the transfer

The window is allocated and its upload and sync both report success (upload no error, sync no error) on the line immediately before the launch failure, so the fault looks like the kernel launch itself rather than the buffer setup. This release touched exactly that area (#187 QSA block scores read each key block once for all of a verify window's queries, #148 bitwise test of the verify window's multi-token GEMV contract, #154 guards on the verify window's tables), and #228 (QSA select on tensor cores via rocWMMA, opt-in, gfx1100) is in the same neighbourhood — possibly relevant, possibly not.

Environment

• GPU: AMD Radeon RX 7900 XTX (gfx1100), 24 GB VRAM — the card docs/AMD_HIP.md names
• ROCm: 7.1.52802 (system install under /opt/rocm, hipcc + hipBLAS + hipBLASLt present), HIP compiler /opt/rocm/llvm/bin/clang++ — i.e. the ROCm version docs/AMD_HIP.md records as the one this backend was tested with
• OS: Ubuntu 24.04, kernel 7.0.0-34-generic; CPU AMD Ryzen 7 7700X (AVX-512), 16 threads; 46 GB RAM; model data on NVMe
• Build: cmake -S . -B build-hip -DCMAKE_BUILD_TYPE=Release -DSTRATA_ENABLE_HIP=ON -DSTRATA_ENABLE_CUDA=OFF -DCMAKE_HIP_ARCHITECTURES=gfx1100 → engine/strata 16.9 MB; engine/BUILD.json reports {"source":"local-hip","backend":"hip","version":"0.1.29","archs":["gfx1100"],"vision":"none","lib_dirs":["/opt/rocm/lib"]}
• Pack: family coder, size IQ1_M, prepared by ./setup.sh --backend hip --family coder --model IQ1_M --vision no --context 4096 --experimental-speed-projection off --low-ram auto
• Load lines from the engine's own log: experts loaded: 23.42 GiB at 0.19 GiB/s (143 s so far), filling the GPU's expert cache (9367 experts, 17.85 GiB of VRAM)

Reproduce

ROCM_PATH=/opt/rocm ./setup.sh --backend hip --family coder --model IQ1_M --vision no --context 4096
./run-coder-iq1_m.sh
# wait for "ready: http://127.0.0.1:<port>/v1", then send one chat completion

What I did NOT test

• The other quants (Q2_0, IQ2_XS, IQ3_XXS, IQ3_S need ~48–62 GB system RAM per the installer's own figures; this box has 46 GB, so only the Coder fits)
• The vision path (disabled), non-gfx1100 AMD cards, Windows+HIP, and multi-GPU
• Only one GPU on the box is usable for this; the integrated Radeon (gfx1036) is correctly reported as unsupported

Scope

Nothing here suggests the Coder weights, the pack, or the download are at fault — the engine loads 23.42 GiB of experts and exits inside its own GPU kernels before it reads a prompt. Your release notes say AMD was not retested for this release because the AMD test PC was unreachable; this is data from a gfx1100 machine that does have the card and the ROCm version named in docs/AMD_HIP.md, so hopefully it saves a round trip. I can re-test against a fix or a diagnostic build, and I'm happy to run any test you'd like instrumented.