gh0stx-bf16

gh0stx-bf16

Expert-pruned BF16 derivative of Qwen3.5-397B-A17B. Full-precision weights — portable to any CUDA stack; quantize to whatever your hardware needs.

Specs

• Base: Qwen3.5-397B-A17B (hybrid GatedDeltaNet + sparse MoE + periodic full-attention)
• Prune: expert-pruned 512 → 174 experts with a domain-calibrated keep-set (routing-frequency over a balanced corpus, not raw router-norm) → ~141B total / ~17B active
• Format: BF16, ~280 GB (16 shards)
• Context: up to 131K (native up to 256K)

Hardware

BF16 needs ~280 GB → multi-GPU, or quantize first:

• 8× H100/A100 80GB (TP=8) — runs BF16 directly.
• Fewer / smaller cards — quantize: AWQ / GPTQ (INT4) ≈ 75–90 GB, or FP8 (H100) ≈ ~140 GB (llm-compressor / AutoAWQ).
• Single NVIDIA Blackwell card (GB10 / DGX Spark) → use the NVFP4 sibling promzeus/gh0stx-nvfp4 (fits in 128 GB, DFlash ~18.5 tok/s).

Serve (vLLM, BF16, tensor-parallel)

vllm serve promzeus/gh0stx-bf16 \
  --served-model-name gh0stx --tensor-parallel-size 8 \
  --trust-remote-code --max-model-len 131072 \
  --gpu-memory-utilization 0.90 \
  --enable-chunked-prefill --enable-prefix-caching \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

Set --tensor-parallel-size to your GPU count. Requires a vLLM build with Qwen3.5-MoE support.

Speculative decoding (recommended: DFlash)

DFlash is the fastest path and is format-agnostic — the drafter is an external ~0.8B model, no file surgery on this repo. Add to the serve command:

  --mamba-cache-dtype float32 \
  --speculative-config '{"method":"dflash","model":"z-lab/Qwen3.5-397B-A17B-DFlash","num_speculative_tokens":10}'

• Drafter: z-lab/Qwen3.5-397B-A17B-DFlash (auto-downloads, matches this model's base → healthy acceptance).
• --mamba-cache-dtype float32 is required with DFlash on this hybrid (GatedDeltaNet) model: in fp16 the Mamba KV page size doesn't reconcile with the drafter's sliding-window page size and engine init fails in unify_kv_cache_spec_page_size. Use n=10 (n≥12 is crash-correlated on the validated stack).
• On a single Blackwell card this yields ~18.5 tok/s (see the NVFP4 sibling); on multi-GPU BF16 the acceptance win carries over, absolute t/s depends on your interconnect.

Alternative: built-in MTP head. A grafted qwen3_5_mtp head is shipped in the NVFP4 sibling in NVFP4 form (Blackwell-only). For BF16/other hardware, DFlash above is the recommended accelerator.

Notes

• KV cache: keep BF16. Do not set --kv-cache-dtype fp8 — unstable with the linear-attention + spec-decode path (and mandatory BF16 for the DFlash drafter).
• Thinking mode is ON by default (Qwen3.5). Disable per request with chat_template_kwargs: {"enable_thinking": false} — recommended for agentic/tool use.

Credits

• Qwen for Qwen3.5-397B-A17B.
• z-lab for DFlash and the Qwen3.5-397B-A17B-DFlash drafter.
• AEON-7 for the DGX-Spark / sm_121a DFlash recipe (the --mamba-cache-dtype float32 unlock).

Sibling

• promzeus/gh0stx-nvfp4 — NVFP4 (W4A4) for a single Blackwell card (GB10), with DFlash (~18.5 tok/s @ 65K) and a built-in NVFP4 MTP head (~13 tok/s @ 131K).