Datasets:
|
Download REPRODUCE.md from ThakiCloud/B200-LowBit-Quantization-Benchmarks: direct link, hf CLI and curl.
- Browser
- Download file 1.79 kB
-
https://huggingface.co/datasets/ThakiCloud/B200-LowBit-Quantization-Benchmarks/resolve/main/REPRODUCE.md
- Command line
-
hf download hf://datasets/ThakiCloud/B200-LowBit-Quantization-Benchmarks/REPRODUCE.md
-
curl -L -o REPRODUCE.md https://huggingface.co/datasets/ThakiCloud/B200-LowBit-Quantization-Benchmarks/resolve/main/REPRODUCE.md
1.79 kB
Reproducing A81
This describes how the A81 configuration was built and measured. It is a recipe, not a guarantee: two builds of the same recipe produced different checkpoint hashes, so expect small quality differences between rebuilds.
Pinned environment
- Base model:
Qwen/Qwen3-8B, revisionb968826d9c46dd6066d109eabc6255188de91218 - Quantization:
llm-compressor 0.15.0,compressed-tensors 0.20.1(isolated Python environment) - Serving:
vllm/vllm-openai:v0.29.0image, no batch-invariance flag, no linear-backend override - Hardware: 1x NVIDIA B200
- Chat template with thinking disabled (
enable_thinking=False)
Recipe
- Quantize every decoder linear projection (q, k, v, o, gate, up, down; 36 layers) with GPTQ.
- Default scheme: NVFP4A16 (FP4 E2M1 weights, group size 16, 16-bit activations).
- Override the 46 units listed in
recipes/a81_allocation.jsonto INT8 weight-only, symmetric, group size 128. A unitL.qkvcoversq_proj,k_proj,v_projof layerL;L.gate_upcoversgate_projandup_proj. - Quantize
lm_headto INT8 weight-only, symmetric, group size 128. Embeddings stay in BF16. - Calibrate with the default (non-Korean-weighted) calibration mix. The calibration text is not distributed.
Checks before trusting a build
- Kernels. Load in vLLM and confirm 98 layers on
MarlinNvFp4LinearKernel, 46 onMarlinLinearKernel, zero float fallbacks,lm_headpacked. - Memory. Weight-load bytes / 2^30 must stay below 5.98 GiB (paper build: 5.917 GiB).
- Latency. Batch 1, 32 requests: TPOT should be at most 1.1x BF16 on the same GPU.
- Quality. Evaluate on an independent set you did not use for any tuning. Do not compare against the paper's FINAL-81 numbers as if they were the same build.