DeepSeek-V4.1-Flash mixed-Q2 GGUF (IQ2_XXS gate/up + Q2_K down)

Routed-expert requantisation of deepseek-ai/DeepSeek-V4.1-Flash revision 2bc89ac599031fa673cab993f1df02fc4a98c673, produced on a 32-vCPU Runpod CPU pod with the pinned llama.cpp ggml reference quantisers (PR ggml-org/llama.cpp#28696 head cd628010bc3fc0a787d156c969d52a0789451c96). Tensor names follow that PR's deepseek4 table.

This is a weights artifact. No runtime has yet been shown to execute it end to end; the "Runtime status" and "Smoke-test findings" sections say exactly what has been measured on which hardware, and what is still blocked.

Recipe

• Expert payload 152.88 GB for 543,581,798,400 routed weights = 2.2500 bits/weight, block bytes copied verbatim from the digest-verified safetensors artifact (no re-quantisation during packing).
• Everything else keeps native precision: FP8 dense weights are dequantised to BF16 with DeepSeek-V4.1's 32x32 E8M0 block scales (V4 uses 128x128 — using the wrong value silently rescales every weight), BF16/F32 tensors are copied byte-for-byte, E8M0 scale tensors are consumed.
• Not imatrix-based. No activation calibration was collected; IQ2_XXS used a constant unit importance vector (the unweighted reference path). This is a deliberate memory-bounded choice, not a calibrated one.
• Not included: the two Engram tables (202.75 GB, stay native in the source checkpoint — SGLang dequantises them on lookup and can hold them host-resident), the MTP/DSpark blocks (no defined V4.1 GGUF slot in the runtime tables yet; skipped, not guessed), and the vision/aligner tensors (they belong in a separate mmproj file).

Files

SHA256SUMS covers the five shards. The pack is deterministic: two independent runs produced byte-identical shard digests.

Smoke-test findings (what was actually measured)

Operational findings worth more than the numbers:

• A Runpod network volume is datacenter-local and single-attach. The artifact volume is in US-NC-2; B300 stock was in US-WA-2/EU-NL-1, so the GPU could not mount it. Pre-fetch the shards with a CPU pod sharing the GPU's datacenter (a second network volume in that DC), then attach it to the GPU. Downloads at ~200 MB/s cost about 15 minutes of CPU-pod time instead of 15 minutes of GPU time, and avoid Hugging Face rate limits (anonymous curl fetches get HTTP 429 while an authenticated client is pulling the same repo).
• Ship config.json and the tokenizer next to the shards. They are now in this repository.
• SGLang's GGUF error surfaces before any tensor is read, so a config problem looks like a model problem; check AutoConfig resolution first when a start dies instantly.

Two defects were found and fixed by that fail-closed mapper before publication: MTP norm/head templates had landed on the main output_norm (duplicate tensor, rejected by any split-aware reader) and llama.cpp's router naming (ffn_gate_inp + exp_probs_b) collides in SGLang's mapper (which splits .weight/.bias). The published pack has neither problem.

Hardware that fits

Run recipes

1. SGLang — works with a 4-rule mapper patch (this is the tested path)

Requires lmsysorg/sglang:dev-dsv41 (SGLang PR #38798) and an NVIDIA B300 (288 GB) or B200 (180 GB). Point --model-path at a directory holding the five shards plus config.json, tokenizer.json and tokenizer_config.json from the source revision. Put a patched copy of srt/model_loader/deepseek4_gguf.py ahead of SGLang on PYTHONPATH and add these four rules to its _v4_checkpoint_name inner-layer branch (they fix the llama.cpp router naming and supply the V4.1 indexer tensors that the mapper does not know yet):

Patch required (two naming gaps, not a weight problem). Copy the mapper patch from this repository and apply it to the SGLang tree:

# patches/ is attached to this repository next to this README
git apply patches/sglang-deepseek4-gguf-name-map.patch      # from the SGLang repo root
# or, against an installed package:
patches/apply.sh --sglang-root /sgl-workspace/sglang        # applies in place, then verifies

It fixes the router naming collision (ffn_gate_inp + exp_probs_b -> layers.N.ffn.gate) and supplies the 48 tensor rules the mapper lacks (ffn_gate_inp.bias_vl x40, indexer.attn_k x4, indexer.k_norm x4). Measured: unpatched the mapper fails; patched it maps 245/245 tensors of shard 1. See patches/README.md for the full rationale and the packer-side --naming sglang alternative.

export SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1        # Engram lives on the host, dequantised on lookup
export PYTHONPATH=/path/to/patched-sglang:$PYTHONPATH
python3 -m sglang.launch_server \
  --model-path ./gguf-with-config/                    # shards + config.json + tokenizer files
  --quantization gguf --load-format gguf \
  --tp-size 1 --port 30000 \
  --disable-cuda-graph                                # V4.1 cells OOM in graph capture (sglang#38839, #38861)
  --max-total-tokens 8192 --context-length 4096 --mem-fraction-static 0.85
# DSpark off: it is the part that fails to start on this backend.

Engram is the known gap in this file. The two hash tables are not in the GGUF, so the loader must obtain them from the source revision; expect a missing-tensor error at the Engram layer if it cannot. That error is the signal to attach those tables, not to change the quantisation.

Q2_K (w2) uses the tensor-core MMQ kernel; IQ2_XXS (w1/w3) has no MMQ in SGLang and takes the MoE vector kernel at every token count, so prefill is slower than a k-quant file would be.

2. ggml kernel harness — works today on any Blackwell card, including 96 GB

git clone --filter=blob:none https://github.com/ggml-org/llama.cpp /opt/llama.cpp
git -C /opt/llama.cpp fetch --depth 1 origin refs/pull/28696/head
git -C /opt/llama.cpp checkout FETCH_HEAD
cmake -S /opt/llama.cpp -B /opt/llama.cpp/build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON \
      -DGGML_NATIVE=OFF -DCMAKE_CUDA_ARCHITECTURES=120a -DBUILD_SHARED_LIBS=ON \
      -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TOOLS=OFF -DLLAMA_BUILD_SERVER=OFF
cmake --build /opt/llama.cpp/build --target ggml -j$(nproc)          # 120a=RTX PRO 6000, 100a=B200, 103a=B300
hf download apetersson/DeepSeek-V4.1-Flash-MixedQ2-GGUF --include '*00001*' --local-dir ./gguf
# (any single shard is enough: the harness slices one expert tensor out of it)
python3 v41_quant_kernel_smoke.py --gguf ./gguf --kernel-bin ./kernel_smoke --tensors 6

This executes the real IQ2_XXS/Q2_K blocks through ggml_mul_mat and compares against the gguf-py dequantisation of the same bytes (gates: cosine ≥ 0.999, relative RMS ≤ 1%). It is the only end-to-end check of the weights that runs on a 96 GB card and needs no model runtime.

3. llama.cpp — validation only

The container parses (GGUFReader), but the deepseek4 runtime cannot load a V4.1 file yet. Use it to verify the artifact, not to run it.

4. vLLM — blocked on a DeepSeek entry in the GGUF plugin

vLLM main has the V4.1 model definitions (PR #56228, merged), and the official vllm-project/vllm-gguf-plugin has Triton IQ2_XXS GEMM + fused MoE. What is missing is a DeepSeek tensor-name adapter in vllm_gguf_plugin/weights_adapter/transformers.py. Once that exists, this is the better-performing route for prefill.

GPU and staging notes learned the hard way

• Stage everything on CPU before renting a GPU. Hugging Face returns HTTP 429 to anonymous or bursty fetches (an authenticated client pulling the same repo at the same time is enough to trigger it), and every retry burns GPU dollars. The working pattern: one CPU pod sharing the GPU's datacenter downloads the shards, the config.json/tokenizer and this repo's worker bundle into a network volume; the GPU pod then starts with zero network work.
• Runpod network volumes are datacenter-local and single-attach. The weights were produced in US-NC-2; B300/B200 stock came and went in EU-NL-1, US-WA-2 and EUR-IS-2, so the volume cannot simply be mounted by the GPU. A delivery volume in the GPU's DC is the fix (US-WA-2 does not support network volumes; EU-NL-1 does).
• AMD: AMD Instinct MI300X OAM (192 GB) exists in the catalog and would fit these weights, but SGLang's DeepSeek-V4.1 work targets MI350X and its GGUF quantisation kernels are CUDA-only, so the serve path is unproven there. The kernel harness could run via ggml's HIP backend (-DGGML_HIP=ON), which is untested for this artifact.
• Availability, not money, was the binding constraint on the smoke: B200 has been empty all afternoon; B300 appears and disappears within minutes; MI300X was unavailable when probed.

Limits, stated plainly

• The quantisation is uncalibrated (unit importance); no imatrix, no capability scores yet.
• Engram and MTP/DSpark are absent from this file, so a runtime that requires them must attach them from the source checkpoint (Engram) or run without the draft (DSpark).
• No behavioural evaluation has been run, so nothing here claims refusal behaviour, KL, multimodal quality or long-context quality.
• The 2.25 bits/weight pay for themselves only because one 180–288 GB card can then hold the model; on multi-GPU native-class hardware the native checkpoint is the safer choice.