cognitivecomputations

/

DeepSeek-R1-AWQ

like 51

Follow

Cognitive Computations 1.6k

Which vLLM version are you using, what's your startup command, and what are the GPUs that you're using?

Thanks for your help! 😊
vLLM Version: 0.7.2
Startup Command: python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 12345 --max-model-len 32768 --trust-remote-code --tensor-parallel-size 8 --quantization moe_wna16 --gpu-memory-utilization 0.97 --kv-cache-dtype fp8_e5m2 --calculate-kv-scales --served-model-name deepseek-reasoner --model cognitivecomputations/DeepSeek-R1-AWQ --enable-reasoning --reasoning-parser deepseek_r1
GPU Configuration: 8 * A800

--enable-reasoning --reasoning-parser deepseek_r1 This will make the streaming output format slightly different, if you don't want to add special support for this, simply remove these 2 flags and it will work.

thanks I'll try it out

Thanks for your help! 😊
vLLM Version: 0.7.2
Startup Command: python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 12345 --max-model-len 32768 --trust-remote-code --tensor-parallel-size 8 --quantization moe_wna16 --gpu-memory-utilization 0.97 --kv-cache-dtype fp8_e5m2 --calculate-kv-scales --served-model-name deepseek-reasoner --model cognitivecomputations/DeepSeek-R1-AWQ --enable-reasoning --reasoning-parser deepseek_r1
GPU Configuration: 8 * A800

Does A100 support "--kv-cache-dtype fp8_e5m2"?

@traphix Yes, but it would be slower than H100.

thanks I'll try it out

have been solved?

thanks I'll try it out

have been solved?

Yeah.
According to the official DeepSeek documentation:
Additionally, we have observed that the DeepSeek-R1 series models tend to bypass thinking pattern (i.e., outputting "\n\n") when responding to certain queries, which can adversely affect the model's performance. To ensure that the model engages in thorough reasoning, we recommend enforcing the model to initiate its response with "\n" at the beginning of every output.

The frequency of triggering thinking is now normal.
However, there are still some issues, as the model's outputs often turn into gibberish.

as the model's outputs often turn into gibberish

Reduce temperature and top p.

Closed as the main issue is solved.

@traphix Yes, but it would be slower than H100.

Hello, I encountered this issue: if I don't add the --kv-cache-dtype fp8_e5m2, I need to reduce the max-model-len to 8192 to avoid OOM (Out of Memory) errors when deploying on 8xH20 gpu. Theoretically, it shouldn't be like this, right?

@ShiningMaker Try using the latest dev version by building from source, it contains MLA for AWQ which massively saves VRAM usage.