Validated fix for gfx1100 Windows ROCm inference crash + VRAM-based heterogeneous AI stack — BEATEK Holdings

Hey AMD community —

I’m Jeremy Jackson, founder of BEATEK Holdings, LLC out of Dallas. Wanted to share what we’ve been building on AMD hardware and introduce ourselves properly.

What we fixed: Diagnosed and validated a fix for the deterministic exit code 2 crash affecting every RX 7900 series user running ROCm inference on Windows. Two root causes — KV cache stream affinity and a Flash Attention memory layout issue on gfx1100 Windows — resolved in 35 lines. Validated at 110 t/s on RX 7900 GRE · Windows 11 · ROCm 7.1.

Also documented the ROCm 7.1 + VS 2026 clang header conflicts that block any Windows ROCm build right now — not written up anywhere else yet.

What we’re building: BEA_Aura — a heterogeneous inference platform combining:

  • AMD RX 7900 GRE (gfx1100 · ROCm · 110 t/s LLM inference)

  • Google Coral Dual Edge TPU (M.2 · PCIe · edge classification)

  • BEA_TidePool VRAM Architecture — Coral and GPU sharing VRAM over PCIe as co-processors, eliminating NVMe file I/O entirely. 3000ms latency → 5ms.

The whole stack runs on AMD hardware. ROCm working correctly on Windows was a production requirement — so we fixed it.

Full documentation, patches, architecture specs, and validation results: https://github.com/Beat-k/BEATEK_ROCm

Happy to connect with anyone on the ROCm Windows side. Also looking forward to working through the AMD AI courses — the right foundation matters.

-– Jeremy F. Jackson · BEATEK Holdings, LLC · jeremy.jackson0@beatek.io

1 Like

Hi, the exit code 2 crash on Windows ROCm has bitten a lot of 7900 users, so a documented fix is genuinely useful. I’m curious about the BEA_TidePool part though, how does the Coral TPU actually share GPU VRAM over PCIe? and also that 3000ms → 5ms jump is a big claim, so I’d love to see what the benchmark setup looks like. Thanks alot for sharing.

Hey Hakob, good questions — let me be straight on both and add the benchmark numbers.

On the VRAM sharing: I want to clarify the framing. The Coral TPU doesn’t directly read GPU VRAM over PCIe — ROCm HIP enforces process-private memory boundaries so hipMalloc allocations can’t be accessed cross-process. What we validated is that the zone architecture works in VRAM within a single process. We confirmed real addresses on the 7900 GRE (hot zone 0x300040000, queue 0x304000000, results 0x304a00000). The current working setup uses NVMe for job manifests between processes, with the Coral reading model weights from its own on-chip 8MB SRAM. Cross-process VRAM sharing is still open and documented as future work.

On the 3000ms to 5ms: the 3000ms is end-to-end router latency through a file-based polling cycle, not raw TPU time. Raw TPU inference is 0.147ms avg on our models. The 5ms figure is theoretical pending the cross-process solution. Should have been clearer about that.

Benchmark results (ROCm fix applied):

Model: Mistral 7B Instruct v0.3 Q4_K_M
GPU: AMD Radeon RX 7900 GRE 16GB · gfx1100
Backend: ROCm 7.1 · Windows 11
Layers: 33/33 offloaded (ngl 99)
CPU: Ryzen 7 5700X3D · RAM: 64GB

10 requests · 200 tokens each · 2000 total tokens
Avg t/s: 65.2
Peak t/s: 83.3 (generation phase)
Min latency: 2.40s · Max: 4.06s

Two tiers show up clearly. Prompt-eval requests land around 50 t/s while generation-heavy requests hit 83 t/s once the GPU is in stride. The bimodal split is expected with Mistral 7B at this context size.

Without the patch: exit code 2 crash on every single request. Two root causes fixed — KV cache stream affinity in ggml-cuda.cu and Flash Attention memory layout in common.cuh. 35 lines total.

this is really great, would love to see it acknowledged by the community

1 Like

right on ive been dealing with the same beast on my rx 7700 xt/gfx 1101. warmup exposions galore, i’m using ububtu 24.04 6.08 kernel for my end and used this:

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

will try your other fixes and update my training script to see if that further stabilizes the rocm+pytorch environment on my end.

added this to /.bashrc
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True,max_split_size_mb:256"
export HSA_OVERRIDE_GFX_VERSION="11.0.0"
export HSA_ENABLE_SDMA=1
export ROCM_FORCE_ENABLE_DP=1

thank you much for sharing this. i’ll report back

I tested your fix with a short epoch run (2) followed by a full training run. I pulled up numbers from pre-fix for comparison. This completely stabilized the training environment! Thank you so much for sharing your work. :purple_heart:

Date Model Dataset Samples Epochs Runtime (s) Steps/sec Samples/sec Final Loss ROCm RDNA3 Fix Stability Notes
2026-05-30 Qwen2.5-3B QLoRA mixed-proficiency-v0 ~500 13.33 ~1800 unstable unstable n/a 6.1 No stalls, jitter, hipBLASLt warnings
2026-05-30 Qwen2.5-3B QLoRA mixed-proficiency-v0 1007 13.33 635.7 0.315 2.517 0.4687 6.1 No completed but unstable throughput
2026-06-06 Qwen2.5-3B QLoRA mixed-proficiency-v1 1007 2 829.9 0.304 2.427 1.509 6.1 Yes fully stable, no ROCm errors
2026-06-06 Qwen2.5-3B QLoRA mixed-proficiency-v1 1007 13.33 635.7 0.315 2.517 0.4687 6.1 Yes fully stable, clean long run

happy to share more details if needed.

jillian

1 Like

Satisfaction guaranteed. I’m happy BEATEK Holdings services was able to assist. We will continue working on ROCm and HIP for future developments and will be happy to share.

1 Like