SERA-32B-GA-FP8

FP8 quantization of allenai/SERA-32B-GA, produced with llmcompressor and validated with vLLM.

Quantization Details

Parameter	Value
Method	FP8 (W8A8) via llmcompressor oneshot
Targets	All Linear layers except lm_head
Calibration dataset	allenai/Sera-4.5A-Lite-T2
Calibration samples	512
Calibration sequence length	2048 tokens
llmcompressor version	0.9.0.2
Hardware	AWS g6e.4xlarge (NVIDIA L40S, 48 GB VRAM)
Model size (uploaded)	~31.7 GB (7 safetensors shards)

The quantization pipeline processes one decoder layer at a time (pipeline="sequential") and offloads activations to CPU between layers, allowing 32B+ models to be quantized on a single GPU without OOM.

GPU Stats

• 1x L40 (g6e.xlarge)
• Total time: 1 hr

Usage

from vllm import LLM, SamplingParams

llm = LLM(model="ikarabulut-dev/SERA-32B-GA-FP8", max_model_len=16384)
params = SamplingParams(temperature=0.7, max_tokens=512)

outputs = llm.generate(
    [{"role": "user", "content": "Explain quantum entanglement simply."}],
    params,
)
print(outputs[0].outputs[0].text)

Note: This model was validated with --max-model-len 16384. Attempting a larger context on a single 48 GB GPU may OOM.

Validation

After quantization the model was loaded into vLLM and a test chat completion request was sent. The model became healthy in ~120 seconds and produced a well-formed thinking-style response — validation passed.

Limitations

• Quality degradation relative to the BF16 base model has not been formally benchmarked. FP8 quantization with 512 calibration samples is generally low-loss for instruction-tuned models, but edge cases may differ.
• Maximum recommended context length is 16 384 tokens on a single L40S GPU.
• The lm_head layer is kept in BF16 (not quantized) to preserve output distribution.

Related

• Base model: allenai/SERA-32B-GA
• Quantization tooling: vllm-project/llm-compressor