Kimi-K3-NVFP4

NVFP4 conversion of moonshotai/Kimi-K3, produced by a closed-form, bit-exact MXFP4 → NVFP4 cast with NVIDIA TensorRT Model Optimizer.

Tooling: patronus-ai/kimi-k3-nvfp4

Cast fidelity	85,085,650,944 / 85,085,650,944 blocks bit-exact (100.0000%), max_abs_err 0.0
Layers cast	247,296 (routed MoE experts)
GPQA Diamond	92.96% (pass@1 avg-of-16)
IFBench	75.27% (pass@1 avg-of-5, prompt_loose)
SciCode	23.08% problem / 57.29% subtask
Verified on	SGLang kimi-k3 branch, 2 nodes × 8×B200, TP16
Size	~1.6 TB, 96 shards

What this is

	Source (Kimi-K3)	This checkpoint
Weight format	MXFP4 (E2M1, block 32, E8M0 scale)	NVFP4 (E2M1, block 16, E4M3 scale + FP32 per-tensor scale)
Weight values	—	bit-identical to the source
Scope	routed MoE experts (w1/w2/w3)	routed MoE experts (w1/w2/w3)

Attention, shared experts, dense MLP projections, lm_head and vision modules are carried through unquantized.

Bit-exactness

The cast is data-free: NVFP4 scales are derived from the source E8M0 exponents rather than re-estimated from data. E8M0 has no mantissa, so every source scale is already an exact power of two. With m = k_max - 8:

weight_scale_2  = 2^m                     (per tensor, FP32)
per_block_scale = 2^(k_j - m)             (per 16 elements, E4M3)
  => product    = 2^k_j                   exactly

E4M3 represents 2^k exactly for k ∈ [-9, 8], so any block within 17 binades of the tensor max reconstructs bit-for-bit. On Kimi-K3 that covered every block. No calibration data was used, and none is needed.

Note on precision. Because the per-block scale is always pinned to an exact power of two, this cast deliberately uses none of NVFP4's precision advantages at the weight level — neither the finer block-16 granularity nor E4M3's 3 mantissa bits. That is the cost of bit-exactness, and it is the right trade: quantization is irreversible, so a higher-resolution container cannot recover information MXFP4 already discarded. NVFP4's accuracy edge is only reachable by re-quantizing from BF16, which for 2.8T params means materializing ~5.6 TB. What the conversion buys is access to the NVFP4 kernel and tooling ecosystem.

Deployment (SGLang)

Verified on the SGLang kimi-k3 branch (not main), 2 nodes × 8×B200, TP16. Run the identical command on every node, varying only --node-rank.

# --- per-node environment ---
export SGLANG_HOST_IP=$(ip -o -4 addr show ens1 | awk '{print $4}' | cut -d/ -f1)
export GLOO_SOCKET_IFNAME=ens1 NCCL_SOCKET_IFNAME=ens1
export NCCL_IB_HCA=mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_9,mlx5_12,mlx5_13
export NCCL_IB_GID_INDEX=7 NCCL_IB_DISABLE=0 NCCL_NVLS_ENABLE=0 NCCL_CROSS_NIC=0
export PATH=/path/to/venv/bin:$PATH          # ninja must resolve from SHARED storage

python -m sglang.launch_server \
  --model-path /path/to/kimi-k3-nvfp4 \
  --trust-remote-code \
  --tp-size 16 --nnodes 2 --node-rank ${NODE_RANK} \
  --dist-init-addr ${HEAD_IP}:20000 \
  --moe-runner-backend marlin \
  --mem-fraction-static 0.85 \
  --mamba-full-memory-ratio 7.21 \
  --disable-flashinfer-autotune \
  --watchdog-timeout 3600 --dist-timeout 3600 \
  --reasoning-parser kimi_k3 --tool-call-parser kimi_k3 \
  --model-loader-extra-config '{"enable_multithread_load": true}' \
  --host 0.0.0.0 --port 30000

SGLANG_HOST_IP is mandatory for multi-node — without it the ranks never rendezvous and startup hangs with no error.

Five non-obvious requirements

Each of these independently prevents the model from serving:

1. --moe-runner-backend marlin. Kimi-K3 uses the situ (SiTuGlu) activation, which is not implemented in SGLang's NVFP4 FlashInfer runner (_SUPPORTED_FP4_ACTIVATIONS = {silu, relu2, gelu}). Marlin supports gated {silu, situ} and forwards both SiTu parameters (gemm1_alpha=β=4.0, clamp_limit=linear_β=25.0).
2. Patch modelopt_quant.py: _SUPPORTED_ACT_STRS += ("situ",). The assert at :2552 runs before the per-backend dispatch at :2556, so it rejects situ even when the selected backend supports it. Every backend already applies its own stricter check afterwards.
3. Raise UNBALANCED_MODEL_LOADING_TIMEOUT_S (load_model_utils.py, default 480 s, no CLI flag). 1.6 TB over network storage means per-node load skew routinely exceeds 8 minutes and aborts an otherwise healthy run.
4. --dist-timeout 3600. Marlin's prepare_moe_nvfp4_layer_for_marlin repacks 247k expert weights, exceeding NCCL's 600 s default; the first rank to finish then times out alone at SeqNum=1 _ALLGATHER_BASE.
5. Omit --dcp-size unless FlashInfer is new enough — cutedsl_mla_backend passes enable_dcp/cp_world to trtllm_batch_decode_with_kv_cache_mla, which flashinfer 0.6.15 does not accept.

Patches 2–3 are scripted in serving/patch_situ_nvfp4.py.

Querying

K3 is a reasoning model — it emits thinking tokens before the answer. Use a generous max_tokens (≥512) and read choices[0].message.content; the thinking is in reasoning_content. reasoning_effort accepts low / high / max.

curl $URL/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "k3",
  "messages": [{"role":"user","content":"What is 17 * 23? Reply with only the number."}],
  "max_tokens": 1200, "temperature": 0, "reasoning_effort": "low"}'

With too small a budget you get content: "" and finish_reason: "length" — that is truncated thinking, not a broken model.

This is a W4A16 deployment

Marlin dequantizes the 4-bit weights to BF16 inside the GEMM, so weights stay 4-bit in HBM (the memory and bandwidth win is kept) but the math runs at BF16 tensor-core rate. Activations are never quantized.

Consequences: no activation calibration is needed or used and this deployment setup does not deliver FP4 compute throughput. Accuracy should be at least that of native MXFP4, since the weights are bit-identical and the activations are higher precision than the source's MXFP8.

Evaluation

All three are judge-free and scored with the official harnesses.

Benchmark	Score	Setup
GPQA Diamond	92.96%	pass@1 avg-of-16, sd 1.24; 198 questions
IFBench	75.27%	pass@1 avg-of-5 prompt_loose, sd 0.86; 300 prompts
SciCode	23.08% / 57.29%	problem / subtask; 65 problems, 288 subtasks

GPQA per-seed: 96.0 / 93.4 / 91.9 / 91.4 / 92.9 / 93.4 / 90.9 / 93.9 / 92.4 / 91.4 / 93.9 / 92.4 / 92.9 / 93.9 / 93.4 / 92.9.

IFBench: prompt_loose counts a response only if every constraint on it is met; prompt_strict was 71.07% and instruction-level loose 78.20%.

SciCode: AA test split, prompt_config=eval/scicode/background, greedy.

Harness: nemo-skills with prompt eval/aai/mcq-4choices and ++eval_type=multichoice, matching the Artificial Analysis methodology; 198 Diamond questions, greedy for seed 0 and temperature=0.6, top_p=0.95 thereafter.

tokens_to_generate=65536 is load-bearing: at 16384 the model exhausted its budget inside reasoning on ~10% of questions and emitted an empty answer, scoring as wrong for a purely harness reason — worth ~8.6 points on the greedy pass (87.4% → 96.0%). Truncation fell from 9.6% to 0.58%.

Layout

Per quantized module <base>:

<base>.weight           uint8   (out, in/2)    NVFP4-packed E2M1 nibbles
<base>.weight_scale     E4M3    (out, in/16)   per-block scale
<base>.weight_scale_2   fp32    scalar         per-tensor scale

config.json reports format: nvfp4-pack-quantized, quant_method: modelopt, quant_algo: NVFP4, group_size: 16, and an exclude_modules list of 55 globs derived from the modules that actually kept an unpacked 2-D float .weight.

Branches. main is weight-only (input_activations: null). w4a4 adds input_scale placeholders of 1.0; they are unused by the marlin path.

Bundled fix

modeling_kimi_linear.py here includes a fix to KimiDeltaAttention: A_log is stored at head_dim (128) but the upstream release builds it at num_heads (96), so from_pretrained fails to load its own weights.

The checkpoint holds 96 trained decay values zero-padded to 128 — verified across all 69 A_log tensors here: elements [96:128] are exactly zero in every one. The FLA kernel indexes by num_heads, so the padding is inert.

The bundled file follows discussion #150: it sizes A_log at num_heads and narrows the checkpoint tensor in _load_from_state_dict, raising if the discarded tail is ever non-zero.

An earlier version of this file followed our own discussion #144, which enlarges the parameter to head_dim instead. That approach loads but cannot run: fla/ops/kda/gate.py computes A_log.view(H, 1) with H = g.shape[-2] (the head count), so a 128-element parameter raises shape '[96, 1]' is invalid for input of size 128 on the first forward. If you pulled this checkpoint before 2026-07-31, re-fetch modeling_kimi_linear.py.

This matters only for transformers-based loading — SGLang narrows to the first num_heads elements independently, which the zero-padding makes lossless.

Known limitation: the FP4 fast path

The trtllm-gen cubin pool ships SiTu kernels at NVFP4's 16-element scale block (57 × BF16-activation, 45 × E2M1-activation), but they are unreachable from SGLang's NVFP4 runner, which resolves activations via FlashInfer's ActivationType (no SiTu member) and reads a different cubin store. Wiring the dispatch to sglang.kernels.ops.moe.trtllm_gen_moe reaches the JIT build stage and then fails on CUDA 12.8 with namespace "cuda" has no member "maximum" — cuda::maximum requires CUDA 13's CCCL. Untested beyond that point.

Reproducing

python cast_mxfp4_to_nvfp4_offline.py \
  --source_ckpt /path/to/Kimi-K3 \
  --output_ckpt /path/to/Kimi-K3-NVFP4 --verify

Runs on CPU without instantiating the model (~7 h, bounded by shard I/O).

License

Derivative of moonshotai/Kimi-K3; the upstream model's license and terms apply.