gh0stx-nvfp4

Expert-pruned, NVFP4-quantized derivative of Qwen3.5-397B-A17B that runs on a single NVIDIA GB10 (DGX Spark, 128 GB unified) with speculative decoding. Two accel paths are supported: DFlash (fastest, recommended) and a built-in MTP head (fallback, no external drafter).

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
• Quant: NVFP4 W4A4 (compressed-tensors), ~90 GB weights
• Context: up to 131K (MTP path); 65K (DFlash path — see tradeoff below)
• Hardware: NVIDIA Blackwell only (sm_120/sm_121a) — NVFP4 is Blackwell-native and will not run on Hopper/Ampere. Validated on a single GB10 (DGX Spark class, 128 GB unified).

Speed (single-stream decode, one GB10)

Path	tok/s	mean accept-len	Context	Notes
DFlash (recommended)	~17	~3.5 tok/step	65K	external z-lab drafter; needs --mamba-cache-dtype float32
MTP head (built-in)	~11–13	~2.0 tok/step	131K	no external drafter; qwen3_5_mtp
no spec-decode	~6	—	131K	baseline

Single-stream decode on GB10 is latency-bound (unified LPDDR5X ≈ 273 GB/s; the SM clock sits low and power is far under budget) — so the win comes from accepting more tokens per pass, not from clocks. DFlash drafts ~3.5 accepted tok/step vs MTP's ~2.0.

Serve — DFlash (recommended, ~17 tok/s)

Requires the z-lab/Qwen3.5-397B-A17B-DFlash drafter (auto-downloads) and a GB10-ready vLLM build (aeon-vllm-ultimate — vLLM 0.24.x for sm_121a, carries the DFlash SWA/prefix-cache ports).

vllm serve promzeus/gh0stx-nvfp4 \
  --served-model-name gh0stx --quantization compressed-tensors --trust-remote-code \
  --language-model-only --load-format runai_streamer \
  --max-model-len 65536 --max-num-seqs 4 --max-num-batched-tokens 16384 \
  --gpu-memory-utilization 0.88 \
  --enable-chunked-prefill --enable-prefix-caching \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --mamba-cache-dtype float32 \
  --speculative-config '{"method":"dflash","model":"z-lab/Qwen3.5-397B-A17B-DFlash","num_speculative_tokens":10}'

The key flag is --mamba-cache-dtype float32. In fp16 the Mamba (GatedDeltaNet) KV page size doesn't reconcile with the DFlash drafter's sliding-window page size, and engine init fails in unify_kv_cache_spec_page_size (assert new_spec.page_size_bytes == max_page_size). fp32 changes the mamba page size so mamba + full-attn + drafter all unify. Also: n=10 (n≥12 is crash-correlated on this stack) and --max-num-batched-tokens 16384 are required for DFlash.

Context tradeoff: with mamba-fp32 + the drafter, the full 131K KV doesn't fit in 128 GB, so DFlash runs at 65K (--max-num-seqs 4, --gpu-memory-utilization 0.88). KV stays BF16 — fp8 KV is unsupported with the non-causal DFlash drafter, and TurboQuant is guarded off on hybrid GDN+Mamba. For full 131K context use the MTP path below.

Serve — MTP head (built-in, ~11–13 tok/s, 131K)

No external drafter; uses the grafted MTP head shipped in this repo.

vllm serve promzeus/gh0stx-nvfp4 \
  --served-model-name gh0stx --quantization compressed-tensors --trust-remote-code \
  --language-model-only --load-format runai_streamer \
  --max-model-len 131072 --max-num-seqs 2 --max-num-batched-tokens 8192 \
  --gpu-memory-utilization 0.87 \
  --enable-chunked-prefill --enable-prefix-caching \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":3}'

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 (thinking isn't spec-accelerated).
• --gpu-memory-utilization: 0.85 solo, up to 0.88 for DFlash headroom; never exceed 0.88 on GB10 (unified memory thrashes → slowdowns/hangs, not a clean OOM).
• The MTP head is BF16 (excluded from NVFP4 via ignore in config.json).

Credits

• Qwen for Qwen3.5-397B-A17B.
• z-lab for DFlash and the Qwen3.5-397B-A17B-DFlash drafter.
• AEON-7 for the aeon-vllm-ultimate GB10/sm_121a build and the DFlash-on-DGX-Spark recipe (the --mamba-cache-dtype float32 unlock).

Sibling

• promzeus/gh0stx-bf16 — full BF16 weights (~141B) for other stacks / your own quantization.