DJLougen / DeepSeek-Coder-V2-Lite-Instruct-NVFP4

DeepSeek-Coder-V2-Lite-Instruct-NVFP4

NV_FP4 (4-bit floating point) quantized version of deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct .

Model Details

NV_FP4 Format

NV_FP4 is a 4-bit floating point format optimized for NVIDIA hardware:

• Format : 1 sign bit, 2 exponent bits, 1 mantissa bit (E2M1)
• Representable values : ±{0, 0.5, 1, 1.5, 2, 3, 4, 6}
• Block-wise scaling : 32 elements per scale factor
• Target hardware : NVIDIA GPUs including Jetson/Spark

Quantization Stats

• Tensors quantized : 5207
• Tensors preserved (FP16) : 84 (embeddings, norms, biases)

Files

• model_nvfp4_quantized.safetensors - Quantized weight tensors (packed uint8 + fp16 scales)
• model_nvfp4_unquantized.safetensors - Preserved tensors (embeddings, norms, biases)
• nvfp4_metadata.json - Quantization metadata

Usage

from safetensors import safe_open
import torch

# FP4 E2M1 lookup table
FP4_VALUES = torch.tensor([0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0])

def dequantize_nvfp4(packed, scales, shape, block_size=32):
    sign1 = (packed >> 7) & 1
    idx1 = (packed >> 4) & 7
    sign2 = (packed >> 3) & 1
    idx2 = packed & 7

    num_blocks = packed.shape[0]
    unpacked = torch.zeros(num_blocks, block_size)
    unpacked[:, 0::2] = FP4_VALUES[idx1] * (1 - 2 * sign1.float())
    unpacked[:, 1::2] = FP4_VALUES[idx2] * (1 - 2 * sign2.float())

    dequantized = (unpacked * scales.unsqueeze(-1)).flatten()
    return dequantized[:torch.prod(torch.tensor(shape))].view(shape)

# Load and dequantize
with safe_open("model_nvfp4_quantized.safetensors", framework="pt") as f:
    for key in f.keys():
        if key.endswith(".packed"):
            name = key.replace(".packed", "")
            packed = f.get_tensor(f"{name}.packed")
            scales = f.get_tensor(f"{name}.scales")
            shape = tuple(f.get_tensor(f"{name}.shape").tolist())
            weight = dequantize_nvfp4(packed, scales, shape)
            print(f"{name}: {weight.shape}")

Base Model

DeepSeek-Coder-V2-Lite-Instruct - MoE code model:

• 16B total / 2.4B active parameters
• 64 experts , 6 activated + 2 shared
• 128K context length

License

Inherits license from base model. See original .