DeepSeek-V2-Lite-FP8-BLOCK-padded

DeepSeek-V2-Lite-FP8-BLOCK-padded

This model is a FP8 block-quantized version of deepseek-ai/DeepSeek-V2-Lite with padding support for non-divisible dimensions.

Overview

• Base Model: deepseek-ai/DeepSeek-V2-Lite (16B parameters)
• Quantization: FP8_BLOCK (128x128 block structure)
• Purpose: Demonstrates FP8 block quantization with weight padding for models with dimensions not evenly divisible by block size

Key Feature: Block Quantization Padding

DeepSeek-V2-Lite has intermediate_size=10944, which is not divisible by the block size of 128. This model uses weight padding to handle this:

• Original intermediate_size: 10944
• Padded intermediate_size: 11008 (86 × 128)

The padding is applied during quantization and the config.json reflects the padded dimensions for vLLM compatibility.

Usage with vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="Etelis/DeepSeek-V2-Lite-FP8-BLOCK-padded",
    trust_remote_code=True,
    tensor_parallel_size=1,
)

sampling_params = SamplingParams(max_tokens=100, temperature=0.7)
output = llm.generate(["Hello, world!"], sampling_params)
print(output[0].outputs[0].text)

Requirements: H100 or newer GPU (SM 8.9+) for FP8 block quantization support.

Quantization Recipe

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier

MODEL_ID = "deepseek-ai/DeepSeek-V2-Lite"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype="auto", 
    trust_remote_code=True, 
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

# FP8 block quantization - ignore layers with composite dimensions
recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_BLOCK",
    ignore=["lm_head", "re:.*kv_a_proj_with_mqa.*"]
)

oneshot(model=model, recipe=recipe)

model.save_pretrained("DeepSeek-V2-Lite-FP8-BLOCK-padded")
tokenizer.save_pretrained("DeepSeek-V2-Lite-FP8-BLOCK-padded")

Created With

• llm-compressor (with padding support)
• compressed-tensors (PR #547)

Ignored Layers

• lm_head: Not quantized (standard practice)
• kv_a_proj_with_mqa: Has composite dimensions (512 + 64 = 576) that cannot be safely padded

License

This model inherits the DeepSeek Model License from the base model.