DeepSeek-V4-Flash on AMD ROCm
ModelScope DSW research baseline

Reproducible serving study

DeepSeek-V4-Flash on AMD ROCm

vLLM compatibility, long-context validation, and 8K top-k sweep evidence from a correctness-first AMD GPU baseline.

ROCm vLLM AITER / PyTorch fallback 2K / 8K / 32K needle retrieval 8K index_topk sweep
GitHub repo GitHub Pages HF dataset mirror
English cover showing the DeepSeek-V4-Flash AMD ROCm reproduction summary

Run summary

A compact view of the gates that were actually validated in the AMD ROCm DSW run.

Service gate
HTTP 200
vLLM OpenAI API server exposed /v1/models successfully.
Short completion
Paris, 13.540s
Stable short-answer sanity check with no garbling.
Best 8K point
index_topk=2048
TTFT 80.313s, effective prefill 101.914 tok/s.

Key results

Gate Result Notes
2K needle retrieval PASS, 53.376s Secret token was recovered from the long prompt.
8K needle retrieval PASS, 151.976s The 8K semantic gate remained stable.
32K needle retrieval PASS at index_topk=4096, 497.470s This was the verified 32K begin-position correctness point.
8K top-k sweep 2048 best for this probe Lower TTFT than the 4096 baseline, but not the 32K correctness setting.
This Space is correctness-first and fallback-heavy. It is a reproducible research baseline, not a claim of production-grade high-throughput serving.

Evidence

The figures below mirror the evidence set used in the writeup and publication draft.

ROCm fallback pipeline figure
ROCm fallback pipeline and service bring-up flow.
Service and correctness evidence figure
Service health and long-context correctness evidence.
Context correctness matrix figure
32K correctness matrix showing the index_topk boundary.
8K top-k performance sweep figure
8K top-k sweep with the best TTFT point at index_topk=2048.
Top-k and negative evidence figure
Top-k sweep and the negative KV-cache planning case.

Reproducibility notes

Included here

  • English cover and evidence figures.
  • Static publication page and writeup links.
  • Small artifacts that help others replay the baseline.

Excluded on purpose

  • Model weights.
  • Large runtime caches.
  • Anything that can be re-downloaded more cleanly than preserved.

Reference links