thaki-AI's picture
D81 aggregate results (private staging)
1c1dda6 verified
|
Raw History Blame Contribute Delete
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`, revision `b968826d9c46dd6066d109eabc6255188de91218`
- Quantization: `llm-compressor 0.15.0`, `compressed-tensors 0.20.1` (isolated Python environment)
- Serving: `vllm/vllm-openai:v0.29.0` image, no batch-invariance flag, no linear-backend override
- Hardware: 1x NVIDIA B200
- Chat template with thinking disabled (`enable_thinking=False`)
## Recipe
1. Quantize every decoder linear projection (q, k, v, o, gate, up, down; 36 layers) with GPTQ.
2. Default scheme: NVFP4A16 (FP4 E2M1 weights, group size 16, 16-bit activations).
3. Override the 46 units listed in `recipes/a81_allocation.json` to INT8 weight-only, symmetric, group size 128.
A unit `L.qkv` covers `q_proj`, `k_proj`, `v_proj` of layer `L`; `L.gate_up` covers `gate_proj` and `up_proj`.
4. Quantize `lm_head` to INT8 weight-only, symmetric, group size 128. Embeddings stay in BF16.
5. Calibrate with the default (non-Korean-weighted) calibration mix. The calibration text is not distributed.
## Checks before trusting a build
1. **Kernels.** Load in vLLM and confirm 98 layers on `MarlinNvFp4LinearKernel`, 46 on `MarlinLinearKernel`,
zero float fallbacks, `lm_head` packed.
2. **Memory.** Weight-load bytes / 2^30 must stay below 5.98 GiB (paper build: 5.917 GiB).
3. **Latency.** Batch 1, 32 requests: TPOT should be at most 1.1x BF16 on the same GPU.
4. **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.