lyydfys/deepseek-v4-flash-rocm-vllm-repro

Reproducing DeepSeek-V4-Flash on AMD ROCm with vLLM: 32K Correctness and TopK Sweep

This article summarizes an engineering reproduction of deepseek-ai/DeepSeek-V4-Flash on an AMD ROCm ModelScope DSW instance. The work focuses on a practical question: can a complex, fast-moving DeepSeek-V4-Flash serving path be turned into a reproducible ROCm baseline with explicit correctness gates?

The answer from this run is yes, with an important boundary: the current setup is a fallback-heavy research baseline, not a production-grade high-throughput serving result.

Why this model

DeepSeek-V4-Flash is not a simple "download weights and start a server" model. Its high-performance path depends on a combination of FP4/FP8 behavior, FlashMLA, sparse MLA, MoE kernels, graph capture, and backend-specific kernel coverage. On AMD ROCm, this makes the model a useful stress test for the boundary between service-level compatibility and true model-path correctness.

The reproduction criteria were:

1. the vLLM OpenAI API server can start and /v1/models returns HTTP 200;
2. short prompts return stable, non-garbled answers;
3. long-context semantic retrieval works at 2K, 8K, and 32K;
4. the 8K index_topk performance/correctness tradeoff can be measured;
5. the scripts, data, and reports are small enough to preserve without storing model weights.

Baseline setup

The main service configuration used:

python3 -m vllm.entrypoints.openai.api_server \
  --model "$MODEL_DIR" \
  --served-model-name deepseek-v4-flash-amd-32k-batch8-16384 \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype auto \
  --trust-remote-code \
  --kv-cache-dtype fp8 \
  --block-size 256 \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --max-model-len 32768 \
  --max-num-seqs 8 \
  --max-num-batched-tokens 16384 \
  --gpu-memory-utilization 0.90 \
  --enforce-eager \
  --async-scheduling \
  --no-disable-hybrid-kv-cache-manager \
  --moe-backend triton \
  --disable-uvicorn-access-log

The run intentionally prioritized correctness and recovery. Several unstable ROCm-specific paths were guarded or routed through ROCm/AITER/Triton/PyTorch fallbacks before performance tuning.

Correctness results

The main correctness gate was needle retrieval: insert a secret code into a long context and ask the model to return only that code.

Gate	Result
Short completion	HTTP 200, Paris, 13.540s
2K needle retrieval	PASS, 53.376s
8K needle retrieval	PASS, 151.976s
32K needle retrieval	PASS at index_topk=4096, 497.470s restart run

The important lesson is that service startup is not enough. For long-context models, semantic retrieval tests are a better gate than "the server is running".

Top-k behavior

The index_topk value became a key correctness/performance knob. Lower values can reduce candidate work, but they may also drop long-distance information.

index_topk	32K begin-position result	Latency
2048	FAIL, missing key fragment	303.88s
3072	FAIL, incorrect token	519.05s
3584	FAIL, incomplete token	549.69s
4096	PASS	572.91s

For this run, index_topk=4096 was the first verified 32K begin-position correctness point.

8K sweep

For the 8K single-request probe, index_topk=2048 was the best point measured:

index_topk	8K needle	TTFT	Effective prefill
4096	PASS	96.589s	84.740 tok/s
1024	PASS	95.217s	85.962 tok/s
1536	PASS	94.250s	86.844 tok/s
2048	PASS	80.313s	101.914 tok/s

Compared with the index_topk=4096 baseline, the 2048 point improved TTFT by about 16.9% and effective prefill by about 20.3%. However, it should not be used as the 32K correctness setting because it failed the 32K begin-position needle case.

Negative result

Trying to raise --max-num-batched-tokens to 32768 failed during KV cache planning:

To serve max seq len 32768, 13.17 GiB KV cache is needed.
Available KV cache memory: 1.39 GiB.
Estimated maximum model length: 3448.

This was a useful negative result. It showed that long-context tuning cannot be done by changing one scheduler parameter in isolation.

Limitations

This run is not a direct comparison with high-end Nvidia serving benchmarks. The workload here was a single-request AMD ROCm DSW probe, while public serving benchmarks often use different GPU systems, different concurrency, and different scripts. Public results are useful as optimization references, not as direct apples-to-apples comparisons for this package.

Next steps

The next high-value work is to move from fallback-heavy correctness to ROCm-native performance:

• sparse MLA / sparse attention indexer profiling;
• ROCm-native top-k kernel experiments;
• mHC, qnorm, RoPE, KV fusion analysis;
• MoE and MXFP4/FP8 path validation;
• speculative decoding experiments once the baseline path is stable.

The repository intentionally keeps model weights out of version control. The small artifacts that matter for reproducibility are scripts, patch notes, CSV tables, figures, and environment reports.

Public links

• GitHub repository: https://github.com/lyydfys/deepseek-v4-flash-rocm-vllm-repro
• GitHub Pages: https://lyydfys.github.io/deepseek-v4-flash-rocm-vllm-repro/
• Chinese GitHub Pages: https://lyydfys.github.io/deepseek-v4-flash-rocm-vllm-repro/zh.html
• vLLM ROCm tracking issue comment: https://github.com/vllm-project/vllm/issues/41820#issuecomment-4529203579
• 中文 README: README.zh-CN.md