[danielhanchen]

Adds MiniMax-M3.1-Flash: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash

• Older MiniMax models are not affected (this branch vs master, CPU, same inputs):
  • MiniMax-M3: the converter writes byte-identical GGUFs (f32 and bf16) on the real MiniMax-M3 config, logits are bitwise identical at 64, 600 and 3000 tokens (top-k block selection active at 3000) for batched and one-token decode, and the same layers run MSA (the dense-lead layers stay dense)
  • MiniMax-M2 / M2.5: byte-identical GGUFs and bitwise-identical logits; the only difference is tokenization of decomposed (non-NFC) text, which now matches HF
• Converter for the mixed MXFP8 + NVFP4 checkpoint: routed experts stay native NVFP4 with a per-expert 1/global_scale tensor (or are dequantized exactly to BF16 / F16 / F32 / Q8_0 with --outtype, --fp8-as-q8 for Q8_0 dense weights); MXFP8 tensors decode exactly
• Sparse attention (MSA) on every layer, including the dense-lead layers: a layer runs MSA iff it carries indexer tensors, so MiniMax-M3 GGUFs behave exactly as before
• MSA batch path chunked over tokens so no intermediate reaches 2^31 elements (large prompts with a large -ub)
• BPE tokenizer applies tokenizer.ggml.normalizer.nfc like the Unigram one already does (M3.1 declares an NFC normalizer); GGUFs without the key are unaffected (same change as unicode : extending opt-in NFC support beyond UGM #30708)
• DSpark speculative drafter: DSparkMiniMaxDraftModel converter, swigluoai in dflash.cpp and per-layer target taps
• Fix: the Gemma 1 + w norm add runs in F32, so F16 sources no longer lose the low bits of w

Tested against transformers in float64 with exact dequantization, on fixtures with the real tokenizer and storage formats: mean KL 5.8e-7 (BF16 source) and 5.6e-7 (NVFP4 + MXFP8), top-1 100%; the MiniMax-M3 layout gives 1.8e-6. The tokenizer matches HF on 3,221 documents and 605 rendered chats. On the real weights, the native NVFP4 GGUF agrees with a reference GPU implementation on 95.5% to 99.1% of top-1 tokens over its own greedy continuations (chat, math, code), and greedy decoding with the DSpark drafter is identical to plain greedy. GGUFs: https://huggingface.co/unsloth/MiniMax-M3.1-Flash-GGUF

Run with llama-server -m MiniMax-M3.1-Flash.gguf --jinja -fa on -np 1 (sparse attention needs one slot or --no-kv-unified); for DSpark add --spec-type draft-dspark -md MiniMax-M3.1-Flash-DSpark-BF16.gguf --spec-draft-p-min 0.

AI Usage

Used Claude and Local Models for testing, iteration and code design - manual verification of model / PR usage

[CISC]

Please drop the dequant.

I understand it may be useful when there's no BF16 model, however it does not belong to model-specific conversion, please follow up with a generic conversion option instead (though this is low priority).