Issues / #579
#579 AMD/HIP RX 9070 XT (gfx1201) stalls during batched prompt ingestion on ROCm 10.0
open · @s91402001 · 7 comentarios · En GitHub
Setup & installServer & APIAMD / HIPModels & quants
Descripción
## Summary
On an AMD Radeon RX 9070 XT (`gfx1201`) under Ubuntu 24.04, Strata's HIP engine stalls during batched prompt ingestion. The 60-second watchdog reports no progress and aborts the engine; OpenCode then receives:
```json
{"message":"the engine stopped unexpectedly (exit code -6); the next request restarts it","type":"server_error"}
```
This was still reproducible after upgrading Strata to v0.1.38 and installing AMD ROCm 10.0.0 for gfx1201. A short OpenAI-compatible request succeeds; a long prompt hangs during batched prefill.
## Environment
- Strata repository/tag: `v0.1.38`, commit `99f3dbd`
- Engine: `0.1.38`, locally built HIP engine, target `gfx1201`
- GPU: AMD Radeon RX 9070 XT, 16 GB VRAM, `gfx1201`
- Secondary GPU: Ryzen 7 8745HS integrated Radeon 780M, `gfx1103` (Strata config selects GPU 0, the RX 9070 XT)
- CPU: AMD Ryzen 7 8745HS, 16 logical CPUs; AVX2 and AVX-512 reported
- RAM: 30.7 GiB usable
- OS: Ubuntu 24.04.4 LTS
- Kernel: `6.17.0-40-generic`
- ROCm user-space: AMD ROCm package `10.0.0-4` for gfx1201
- HIP runtime: `7.15.26333-0000000`
- Installed AMD driver package: `amdgpu-dkms 7.1.3.31500000`, firmware `31.50`
- **Kernel module currently loaded:** `amdgpu 6.16.13` (the new DKMS driver has been installed but not yet tested after reboot)
- Model: Qwen3.8-Flash-Next IQ2_XS, native pack, PLE shard and MTP draft
- Backend/config: HIP, GPU 0, `--expert-cache auto --prefill 512 --spec 4 --spec-min-p 0.5 --max-context 131072 --kv int8 --kv-resident 32768 --mmap-experts`
- `STRATA_PLE_BATCH` is unset for the current ROCm 10 build.
## Reproduction
1. Start the configured IQ2_XS server.
2. Send an OpenAI-compatible chat completion with a long prompt (roughly 8K tokens or more; an OpenCode request also reproduces).
3. The request stalls in `reading the prompt (batched)` and the watchdog aborts the engine after 60 seconds without progress.
A short API request succeeds. A synthetic prompt of about 21.5K tokens was also tried.
## Observed failure
With the ROCm 10 build, the server log reported:
```text
[strata] the engine stopped unexpectedly (exit code -6).
The engine stopped itself because it had stopped making progress
...
strata serve: no progress for 60 s during a request
(reading the prompt (batched): layer 5 of the prompt chunk from token 0)
```
The engine log contains the corresponding stall report. At the time, it reported about 20 GiB of RAM available and effectively no swap use; this does not look like ordinary system-RAM exhaustion.
The earlier ROCm 7.2.4 log shows the same stall class at multiple layers (for example layers 5, 14, 16, 20, 23, 27 and 38), also during batched prompt ingestion. That older log predates the ROCm 10 cleanup/reinstall, so the attached logs should be read with their timestamps in mind.
## What I tried
- Checked the OpenCode endpoint and OpenAI-compatible API; short requests succeed.
- Tried `--prefill auto` and `--prefill 512`; the stall persisted.
- Upgraded Strata from engine 0.1.35 to 0.1.38 and rebuilt for `gfx1201`; the stall persisted.
- Tried `STRATA_PLE_BATCH=0`; it did not prevent the reported failure.
- Installed ROCm 10.0.0 gfx1201 packages, rebuilt the engine against `/opt/rocm/core-10.0`, and verified the binary loads ROCm 10 HIP/hipBLAS/rocBLAS libraries. A short API request succeeded.
- The long-prompt ROCm 10 run in `/tmp/strata-rocm10.log` occurred before I noticed that cleaning up ROCm 7 had removed ROCm 10 files; ROCm 10 was subsequently reinstalled. Please treat that run as a useful but not fully clean A/B result.
- ROCm 10 DKMS driver 7.1.3 is installed, but this run has not been rebooted, so the loaded module is still 6.16.13.
## Attachments
- `strata-iq2_xs.log` — engine log with repeated stall reports.
- `/tmp/strata-rocm10.log` — server output from the ROCm 10 run, including the 21.5K-token request and watchdog error.
- The configured server arguments are listed above; the config's model paths and profile can be redacted if desired.
[strata-iq2_xs.log](https://github.com/user-attachments/files/32989179/strata-iq2_xs.log)
[strata-rocm10.log](https://github.com/user-attachments/files/32989184/strata-rocm10.log)En el sitio
Enlaces a install, modelos, releases.