Soofi-S-Isar-Preview-FP8

⚠️ Preview / internal checkpoint. Weights and metadata may still change.

FP8 (W8A8 dynamic) quantization of Soofi-Project/Soofi-S-Isar-Preview for space-efficient serving with vLLM. This is the Isar reasoning variant — it emits explicit <think> traces before the final answer.

Quantized from bf16 safetensors with llm-compressor to the compressed-tensors FP8 format, which vLLM loads natively at roughly half the weight memory of the bf16 checkpoint.

Architecture support: SOOFI-S is a custom hybrid Mamba-2/MoE model and ships with its own modeling code (trust_remote_code). FP8 serving requires a vLLM build that understands this architecture — verify against the actual checkpoint before relying on this artifact.

Quantization details

Property	Value
Scheme	FP8_DYNAMIC (W8A8)
Weights	FP8 E4M3, per-channel static scales
Activations	FP8 E4M3, per-token dynamic (quantized at runtime)
Calibration	none (data-free)
Kept in full precision	lm_head, Mamba-2 in_proj/out_proj

Why dynamic / data-free? Dynamic per-token activation scales need no calibration dataset and are robust for MoE, where a single static activation scale across experts is a poor fit. The MoE router is not an nn.Linear, so it stays full precision automatically; the Mamba-2 in_proj/out_proj (the recurrent SSM path) are kept bf16 as the most quantization-sensitive layers.

Size scales with the total 30B parameters (not the 3.5B active), so the FP8 weights are ~half the bf16 size minus the few full-precision tensors above.

Usage with vLLM

# OpenAI-compatible server
vllm serve Soofi-Project/Soofi-S-Isar-Preview-FP8 --trust-remote-code

from vllm import LLM, SamplingParams

llm = LLM(model="Soofi-Project/Soofi-S-Isar-Preview-FP8", trust_remote_code=True)
out = llm.chat(
    [{"role": "user", "content": "How many r's are in strawberry?"}],
    # reasoning models spend output tokens on the <think> trace — budget generously
    SamplingParams(temperature=0.6, top_p=0.95, max_tokens=2048),
)
print(out[0].outputs[0].text)

--trust-remote-code is required: the custom hybrid Mamba-2/MoE modeling code travels with the checkpoint. FP8 needs a GPU with hardware FP8 support (NVIDIA Hopper/Ada/Blackwell — e.g. H100, L40S, RTX 4090) for the fast path; on older GPUs vLLM falls back to a slower Marlin kernel.

Identity and tool calling work out of the box. vLLM's chat endpoint applies the model's own embedded Jinja chat_template, so the identity default system prompt and the native tool-calling format are used verbatim — unlike Ollama, you do not need to supply a SYSTEM block manually.

Reasoning output: the model emits <think> … </think> blocks inline, which consume output tokens — give it a generous max_tokens budget and context.

Architecture note

This is a hybrid Mixture-of-Experts model designed from scratch: 23 Mamba-2/MoE layers + 6 attention layers, 128 routing experts + 1 shared expert per MoE layer, 6 experts active per token (30B total / 3.5B active). FP8 is applied to the Linear layers (attention/MoE expert projections); the SSM (Mamba-2) recurrent parameters and the router stay in higher precision. A recent version of vLLM is recommended.

Related models

• Base (bf16): Soofi-Project/Soofi-S-Isar-Preview
• Other reasoning variant: Soofi-Project/Soofi-S-Rhine-Preview
• GGUF (llama.cpp/Ollama): Soofi-Project/Soofi-S-Isar-Preview-GGUF

License & provenance

Released under a custom license ("Other"), following the base model Soofi-Project/Soofi-S-Isar-Preview. TODO: mirror the full license text once the base model card defines it.