Nemotron-3.5-Lightning DFlash — local BF16 reconstruction (Intel XPU)
⚠️ Draft only. Not a standalone chat model. Pair with the GPTQ INT4 G64-sym target.
This is a local NVFP4 → BF16 reconstruction of
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash
so vLLM XPU can load the official DFlash speculator. NVIDIA ships the draft as
ModelOpt NVFP4 (E2M1 + E4M3 scales). The pinned XPU image does not consume that
packing for this draft.
Publisher: SergiioB.
📖 Cookbook — setup, patches, methodology, evidence
- Repo: intel-arc-pro-b70-inference-cookbook
- Nemotron speed card + capacity ladder: NEMOTRON-DFLASH-B70.md
- Model + launch recipe: NEMOTRON-B70.md
- Image + patch matrix: IMAGE-AND-PATCH-MATRIX.md
- Copy-paste launch lines: FULL-SETUP-COMMANDS.md
Launch with the GPTQ target
vllm serve /path/to/gptq-target \
--quantization gptq --dtype float16 \
--max-model-len 16384 --max-num-seqs 1 \
--no-enable-prefix-caching --async-scheduling \
--speculative-config '{"method":"dflash","model":"/path/to/this-draft","num_speculative_tokens":7}'
Set VLLM_XPU_ENABLE_XPU_GRAPH=1. On B70 also apply the cookbook patches
(native grouped-topk v2 + SSU B8/W4).
Conversion
- Attention, embeddings, and norms were already BF16 and were copied.
- MLP +
fcNVFP4 tensors were dequantized: E2M1 low-nibble-first ×float8_e4m3fn× F32weight_scale_2, group 16, linear (not swizzled). Matches vLLMnvfp4_emulation_utilsdecode. quantization_configwas removed fromconfig.json. If a leftoverhf_quant_config.jsonis present, rename it — vLLM will otherwise treat the draft as NVFP4 and fail on XPU.- 19 matrices dequantized, 51 tensors copied. See
CONVERSION.txt.
This is not a published official BF16 draft. It is a reconstruction.
Measured with the GPTQ target on Arc Pro B70
Real measurements from the publisher's card: isolated C1, n=5 medians, cache
off, 150 W cap, DFlash n_spec=7. Commands and raw logs are in the cookbook
(links above):
- representative decode 186.61 t/s at exact p2048/g128 (174.60–201.83)
- p8192/g128 157.92 t/s (1.81× vs matched no-spec 87.25 on that cell)
- p8192/g1 cold input 7160 t/s (prompt tokens / client TTFT — not isolated engine prefill)
- window acceptance 1830/3521 = 52.0%
Do not headline the p512 194.6 median (family range 140–220).
License
OpenMDW-1.1, same as the NVIDIA DFlash source.