sahilchachra/Leanstral-1.5-119B-A6B-NVFP4

Leanstral-1.5-119B-A6B — NVFP4

NVFP4 (4-bit) quantization of Leanstral-1.5-119B-A6B for Blackwell-class NVIDIA GPUs (native FP4 tensor cores). Derived from the BF16 HF conversion `sahilchachra/Leanstral-1.5-119B-A6B-BF16`, itself an unofficial community conversion of `mistralai/Leanstral-1.5-119B-A6B`.

Architecture

Composed `Mistral3` model = DeepSeek-V3 text backbone (MLA attention + shared-expert MoE: 128 routed experts, 4 active/token, 1 shared) + Pixtral vision encoder. (llama.cpp also detects the text stack as deepseek2 = DeepSeek-V3.)

What was quantized

• —NVFP4 (W4A4), `nvfp4-pack-quantized`, group 16, FP8-E4M3 scales: all language-model Linear layers and all MoE experts — the fused DeepSeek-V3 experts are auto-linearized by llm-compressor and packed. 13,932 expert tensors packed (verified — experts were NOT skipped).
• —Kept full precision: vision tower, multimodal projector, MoE router (mlp.gate), lm_head, token embeddings.
• —Calibration: HuggingFaceH4/ultrachat_200k, 128 samples, seq 512, text-only.

Provenance & precision

Pipeline: original FP8 consolidated → dequant to BF16 (lossless) → NVFP4. The 4-bit weights derive from an already-8-bit source, not full-precision weights.

✅ Verified / ⚠️ Not yet verified (tested 2026-07-04 on 1× B200, CUDA 13)

• —✅ Conversion is structurally exact: strict key+shape match against a real Mistral3(deepseek_v3+pixtral) model (801 tensors).
• —✅ NVFP4 produced with all experts packed (13,932 tensors); valid compressed-tensors format; loads in transformers.
• —✅ vLLM 0.23 resolves the architecture (Mistral3ForConditionalGeneration).
• —⚠️ End-to-end generation not yet confirmed on a single B200. Two engine-side gaps for this new composed arch (deepseek_v3 text inside Mistral3):
• —vLLM 0.23 raises KeyError: ...mlp.gate.e_score_correction_bias when loading weights (its DeepSeek-V3 loader in the Mistral3 wrapper expects a different key path). Needs a vLLM build that supports this composition.
• —transformers loads it but dequantizes NVFP4 to BF16 (no FP4 compute path) → ~159 GB + the naive MoE loop → OOM/too slow on one 183 GB B200.
• —⚠️ Vision path is converted (Pixtral tensors + preprocessor_config.json present) but not runtime-validated.
• —⚠️ RoPE scaling is approximate (YARN mapped; source also carries llama_4_scaling). Long-context (>8k) not guaranteed.

In short: the artifact is correct; serving it needs engine support for the composed arch that isn't in the pinned stack yet. Track vLLM support for DeepSeek-V3-backed Mistral3.

Intended usage (mirrors the base model)

Tokenizer is Tekken (tekken.json, use mistral_common); chat format is the Mistral instruct format (handled by mistral_common, no manual [INST]). The base model recommends temperature 1.0. Once engine support lands, the intended path:

vllm serve sahilchachra/Leanstral-1.5-119B-A6B-NVFP4 \
  --tokenizer-mode mistral --reasoning-parser mistral \
  --tool-call-parser mistral --enable-auto-tool-choice

Files included: config.json, weights + index, tekken.json, params.json, chat_template.jinja, preprocessor_config.json.

License

Derives from mistralai/Leanstral-1.5-119B-A6B; upstream license/terms apply. Community artifact, not affiliated with Mistral AI.