deepseek-v4-flash-rocm-vllm-repro
Reproducing DeepSeek-V4-Flash on AMD ROCm with vLLM
Subtitle: 32K correctness, 8K index_topk sweep, and a fallback-heavy ROCm
research baseline.
Live Space: https://lyydfys-deepseek-v4-flash-rocm-vllm-space.static.hf.space
This page is a public-facing summary of the reproducibility package. It is written for GitHub Pages, so it keeps the narrative short and links back to the repository reports for details.
Summary
This project reproduces DeepSeek-V4-Flash serving on a ModelScope DSW AMD ROCm instance through vLLM and fallback-safe ROCm/AITER/PyTorch paths. The project does not claim upstream ROCm enablement. It preserves an engineering baseline with scripts, notebook, data, figures, and reports.
Verified gates:
- OpenAI-compatible vLLM service health;
- short generation sanity check;
- 2K, 8K, and 32K semantic needle retrieval;
- 8K
index_topksweep; - negative scheduler/KV cache finding.
Results
| Result | Value |
|---|---|
| 32K correctness top-k | index_topk=4096 |
| 32K restart needle latency | 497.470s |
| Best 8K top-k | index_topk=2048 |
| Best 8K TTFT | 80.313s |
| Best 8K effective prefill | 101.914 prompt tok/s |
Figures
Boundary
The baseline is useful for reproducibility and future ROCm kernel work. It is not a production-grade serving result and should not be presented as an apples-to-apples Nvidia comparison.
Repository
See the repository README and reports for the full commands, scripts, data, and evidence.