Darkstar-Qwen3.8-27B-Base-ModelOpt-W4A16-NVFP4-Mixed-FP8

Darkstar-Qwen3.8-27B-Base-ModelOpt-W4A16-NVFP4-Mixed-FP8

Summary

Clean, unedited NVIDIA ModelOpt quantization of Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The selected mixed-precision artifact uses W4A16 NVFP4 group 16 for language MLP projections and lm_head, FP8 e4m3 for self-attention and GatedDeltaNet projections, and BF16 for protected components and runtime KV. It was selected over the rejected uniform W4A4 candidate on single-stream throughput.

ModelOpt is pinned to 0.46.0rc2 at 43fd41a58d52c4e6e5dec1d1ff5989ecc737ae1a; the selected operator recipe SHA-256 is 90fc6b37c00334debd49f1975ab406b5e20667f07e4be0be3e463a648abac642. Calibration used cnn_dailymail plus nemotron-post-training-dataset-v2, 512+512 samples, sequence length 2048, seed 1234. Vision, all 15 MTP tensors, conv1d, norms, embeddings, and runtime KV remain BF16.

Full provenance, protocol, and caveats:

• Source card
• Artifact lineage
• Benchmark matrix
• GPQA protocol

Evaluation

Safety and limitations

This is the unedited control quantization. The refusal measurements are behavior observations, not a safety endorsement. Quantization can shift behavior relative to upstream BF16. Only single-stream throughput is characterized here; aggregate concurrent throughput is separate.

Release reference

Engineering release: darkstar-qwen3.8-27b-v1.0.0. This immutable tag exists and the release contract is published.

Runtime

Validated with vLLM 0.27.1, compiled mode, Flash Attention, BF16 KV cache, context 126,144, MTP depth 4, 32K scheduler budget, and max_num_seqs=16:

vllm serve HangGlidersRule/Darkstar-Qwen3.8-27B-Base-ModelOpt-W4A16-NVFP4-Mixed-FP8 \
  --served-model-name darkstar-qwen38-base-nvfp4 \
  --kv-cache-dtype bf16 \
  --max-model-len 126144 \
  --max-num-seqs 16 \
  --max-num-batched-tokens 32768 \
  --enable-chunked-prefill \
  --compilation-config 2 \
  --speculative-config '{"method": "mtp", "num_speculative_tokens": 4}'