DeepSeek-V4.1-Flash-NVFP4

DeepSeek-V4.1-Flash-NVFP4 (community, weight-only)

NVFP4 quantization of deepseek-ai/DeepSeek-V4.1-Flash (revision fb2764a5cf32). Unofficial community artifact, not affiliated with DeepSeek or NVIDIA.

510 GB -> 415 GB (19% smaller). The first NVFP4 quant of this checkpoint.

What is quantized

Layout per converted tensor mirrors nvidia/DeepSeek-V4-Pro-0813-NVFP4 exactly: X.weight (U8, two E2M1 per byte, low nibble first), X.weight_scale (F8_E4M3, per-16), X.weight_scale_2 (F32 scalar). No input_scale (weights only; activations stay dynamic). The quantization_config keeps the source's fp8/ue8m0/expert_dtype: fp4 fields and adds quant_algo: MIXED_PRECISION plus a quantized_layers map naming every NVFP4 module (the convention the vLLM DeepSeek-V4 NVFP4 path reads).

Method

• Weight-only PTQ with NVIDIA Model Optimizer 0.46.1 (NVFP4QTensor.quantize), streaming shard-by-shard converter; no model instance, no activations.
• Global scale per tensor chosen by an MSE sweep over multipliers {0.8, 0.9, 1.0, 1.1, 1.25} around the standard amax/(448*6) default.
• No activation calibration set was used: there is currently no runtime that can execute this model (see below), so activation-aware calibration was impossible. Weight-only max/MSE calibration needs no data.
• Format verified bit-exact: modelopt's NVFP4 dequant of nvidia/DeepSeek-V4-Pro-0813-NVFP4 expert bytes reproduces the source MXFP4 weights exactly; this repo emits the same layout.

Measured quantization error (per-tensor, vs the FP8 source weights)

The engram tables are the largest component converted (203 GB FP8 -> 111 GB). Row-level rel. L2 error is 9.4%. An error-propagation simulation through the real engram gate (sampled table rows + real wkv/gate weights, synthetic hidden states) gives output cosine similarity > 0.9999 and contribution-weighted output error < 1% for hidden-state RMS >= 10 (0.7% at RMS 10, 0.07% at RMS 100). At hidden RMS ~1 the contribution ratio is large and output error reaches ~5%. End-to-end measurement (below) confirms the simulation: 32K needle-in-a-haystack retrieval, the most engram-sensitive workload tested, shows no engram-specific damage.

End-to-end measured (2026-09-16)

vllm/vllm-openai:deepseekv41-flash-0909 executes deepseek_v41 natively but has no dense-NVFP4 path: its deepseek_v4_1/quant_config.py forces FP8/UE8M0 for all LinearBase layers (experts can be NVFP4 via moe_quant_algo), and it never reads this repo's quantized_layers map. The repo therefore still cannot be served as-is; see Runtime support.

Quality was instead measured through a dequantized-FP8 twin: every NVFP4 tensor dequantized and re-encoded as FP8 E4M3 + UE8M0 block scales (the source repo's exact layout), served on 8xH200 with the stock image. The FP8 re-encode adds ~2.6% rel error on top of the NVFP4 error, so twin scores are a pessimistic bound on true NVFP4 quality. Identical 0-shot lm-eval harness for all checkpoints (MMLU 10/subject; zero-shot n=200; NIAH 32K n=10):

Weight-only NVFP4 costs 3.3pt MMLU (true value smaller — twin is a lower bound), zero-shot is a wash, and needle retrieval is intact. The MMLU loss lives in the dense attention/indexer tensors, not the engram tables. A v2 that leaves attention at FP8 (+7 GB) should recover most of it; activation-aware PTQ for the dense tensors is the follow-up, now that a runtime exists to collect activations.

If the engram tables turn out to be quality-critical, you can reconstruct an FP8-engram variant: download model-00047-of-00048.safetensors and model-00048-of-00048.safetensors from the source repo, swap them in, and remove the layers.1.engram/layers.14.engram entries from quantization_config.quantized_layers.

Runtime support (updated 2026-09-16) — read before downloading

• vLLM vllm/vllm-openai:deepseekv41-flash-0909 serves deepseek_v41 (including the FP8 source checkpoint) but cannot load this repo: dense-NVFP4 is unsupported for this model type — LinearBase layers are hard-wired to FP8/UE8M0, only MoE experts have an NVFP4 branch (moe_quant_algo: NVFP4, the convention NVIDIA's V4-Pro port uses). The per-layer quantized_layers map in this repo's config is not consumed by that loader. Serving this repo as-is needs a loader patch (route quantized_layers entries to the ModelOpt NVFP4 linear method); the engram tables additionally need an NVFP4 embedding path.
• transformers: deepseek_v41 is not in 5.17.0 (latest PyPI) nor in main. The source checkpoint ships DeepSeek's native inference/ runtime instead, which reads the source FP8/MXFP4 format, not this repo's NVFP4 dense tensors.
• SGLang: no deepseek_v41 model file.

End-to-end quality is nevertheless measured — see the section above (dequant-FP8 twin on stock vLLM, 8xH200).

Reproduction

Converter + validation harness: see the model card discussion / commit history. Method: nvfp4_shard.py (streaming converter), local_validate.py (range-fetch error harness), ModelOpt 0.46.1. Source sha fb2764a5cf32. Conversion took ~47 min on a 16-core CPU container (repo upload ~37 min more); no GPU used.

License

Inherited from the source model (mit).