#5953 [Engine] Qwen3.6-35B-A3B fails at startup on gfx1150 (Linux, ROCm 7.15): prefill gemm cublasGemmEx f16 - cuBLAS status 6
open · @masahitojp · 5 Kommentare · Auf GitHub
Server & APIAMD / HIPNVIDIA / CUDAModels & quantsWindowsLinux
Beschreibung
## Summary
On **Linux + Radeon 890M (gfx1150, Ryzen AI 9 HX 370)** with **ROCm 7.15 / hipBLASLt 1.4.0.1**, Qwen3.6-35B-A3B (Unsloth UD-Q4_K_XL) loads but the engine exits during startup at the first prefill GEMM:
```
prefill gemm: cublasGemmEx f16: cuBLAS status 6
Warning: Resource leak detected by SharedSignalPool, 89 Signals leaked.
```
Qwen3.8-Flash-Next (IQ3_XXS) runs fine on the same engine and same machine, so the engine itself works on gfx1150.
## Environment
| | |
|---|---|
| OS | Ubuntu 26.10 (kernel 7.4.0-070400rc2-generic) |
| CPU/GPU | AMD Ryzen AI 9 HX 370 / Radeon 890M (Strix Point, gfx1150) |
| Memory | 90 GB visible, TTM pages_limit = 64 GB |
| ROCm | 7.15.0 (`hipcc`), hipBLASLt runtime 100401 |
| Strata | 0.2.13.1 (main `13eedf55`) |
| Engine | 0.1.59.1, `source: local-hip`, `archs: [gfx1150]`, built with `STRATA_EXPERIMENTAL_GFX1150=1` |
| Model | `pfnet`? no — `unsloth/Qwen3.6-35B-A3B-MTP-GGUF`, UD-Q4_K_XL (22.9 GB) |
| Context | 32768 |
## Steps
```sh
STRATA_EXPERIMENTAL_GFX1150=1 .venv/bin/python strata.py launch --family qwen36 --model UD-Q4_K_XL --context 32768 --no-tui --yes
# the generated config referenced component:strata-engine, so it was pointed at the local engine binary:
# exe = /data/masahito/llm/Strata/engine/strata
.venv/bin/python serve/server.py --engine strata --config strata-qwen36-ud-q4_k_xl.json --port 8080
```
## Observed
The model loads (4.40 GiB dense + 18.75 GiB routed experts on the GPU, 54.97 GiB VRAM free), then:
```
[strata] engine started: .../engine/strata --serve --native .../Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf --max-context 32768 --prefill auto --mtp 2
Secondary CUID not available: AMDCUID support not enabled.
strata-q36: ... 753 tensors (layout checked), 4.40 GiB on the GPU in 0.6 s
strata-q36: prompt experts on MMQ (int8)
prefill gemm: hipBLASLt version mismatch: file=100200 runtime=100401; using hipBLASEx
strata-q36: context 32768 cells, prompt chunk 4096, MTP block loaded, 54.97 GiB of VRAM free
strata-q36: every routed expert in VRAM (18.75 GiB) in 12.0 s
prefill gemm: cublasGemmEx f16: cuBLAS status 6
Warning: Resource leak detected by SharedSignalPool, 89 Signals leaked.
```
## Ruled out
| Variant | Result |
|---|---|
| engine 0.1.59.1 | fails |
| engine 0.1.57 (previous local build) | fails, same line |
| `STRATA_HIPBLASLT_TUNING` set / unset | fails both (with the file: "hipBLASLt version mismatch ... using hipBLASEx") |
| `--mtp 2` / no `--mtp` | fails both ("MTP block loaded" is printed either way) |
| `STRATA_HIPBLASLT=0` | fails |
| `STRATA_DENSE_F32=1` | fails (still calls the f16 path) |
| Qwen3.8-Flash-Next IQ3_XXS, same engine | **works** |
## Question
* Is the `cublasGemmEx f16` dense path expected to work on gfx1150 with ROCm 7.15 (hipBLASLt 1.4.0.1)? The tuning-table file for gfx1150 here is scoped to hipBLASLt 100200, so the runtime mismatch may be part of it.
* Is there an env switch to force the dense projections through a path that works on gfx1150 (bf16, MMQ/quantized, or rocBLAS), or is this a known limitation of the experimental gfx1150 opt-in?
Related: #5951 (gfx1150 Windows success with Qwen3.8), #5715 (Qwen3.6 exit code 1).
Mehr auf der Site
Links zu Install, Modellen, Releases.