FP32
FP32 is a 32-bit single-precision floating-point format. It is the baseline precision for deep learning model training and inference, consuming relatively large amounts of storage and computational resources.
Core Facts
Quantizing a model from FP32 to INT8 can improve inference speed by 3–4x, but may lead to reduced detection accuracy for small targets.
The core principle of quantization is to convert model parameters from high-precision floating-point representations (such as FP32's 32-bit floating point) to low-precision representations (such as INT8's 8-bit integer or INT4's 4-bit integer).
BF16 uses 1 sign bit, 8 exponent bits, and 7 mantissa bits, sacrificing precision while retaining the same dynamic range as FP32.
Timeline (last 90 days)
The method proposed in the study keeps selected intermediate results at FP32 precision in three key regions—RoPE (rotary position embedding), SwiGLU (activation function), and residual RMSNorm—and only rounds at their BF16 output stage.
BF16 (Brain Floating Point 16) is a 16-bit floating-point format that retains an 8-bit exponent (the same as FP32) but has only a 7-bit mantissa, thus offering a large dynamic range but low precision.
Models are often trained with FP32 or mixed precision, and are usually converted to FP16 when releasing inference versions to halve the size.
During SkyRL's weight synchronization, the FP32 router weights are fully preserved to avoid numerical instability caused by low-precision conversion.
Scoring outputs are more sensitive to numerical perturbations, and low-precision quantization easily causes shifts in score ranking, while ranking accuracy is the core metric for filtering and candidate ranking tasks.
Compute-bound means the hardware's compute units are fully saturated, at which point increasing batch size or optimizing numerical precision (such as using BF16 instead of FP32) becomes meaningful.
When the cumulative sum reaches 1.0, the smallest change representable by FP32 is 1 ULP = 2⁻²³ ≈ 0.00000012.
FP32 uses 32 bits to represent floating-point numbers, consisting of 1 sign bit, 8 exponent bits, and 23 mantissa bits.
FP32 can only maintain approximately 7 significant digits of precision.
The core principle of quantization is to convert model parameters from high-precision floating-point representations (such as FP32's 32-bit floating point) to low-precision representations (such as INT8's 8-bit integer or INT4's 4-bit integer).
3 more timeline events
All Facts (15)
Quantizing a model from FP32 to INT8 can improve inference speed by 3–4x, but may lead to reduced detection accuracy for small targets.
80%VerifiedThe core principle of quantization is to convert model parameters from high-precision floating-point representations (such as FP32's 32-bit floating point) to low-precision representations (such as INT8's 8-bit integer or INT4's 4-bit integer).
65%VerifiedBF16 uses 1 sign bit, 8 exponent bits, and 7 mantissa bits, sacrificing precision while retaining the same dynamic range as FP32.
65%UnverifiedThe method proposed in the study keeps selected intermediate results at FP32 precision in three key regions—RoPE (rotary position embedding), SwiGLU (activation function), and residual RMSNorm—and only rounds at their BF16 output stage.
50%UnverifiedBF16 (Brain Floating Point 16) is a 16-bit floating-point format that retains an 8-bit exponent (the same as FP32) but has only a 7-bit mantissa, thus offering a large dynamic range but low precision.
50%UnverifiedModels are often trained with FP32 or mixed precision, and are usually converted to FP16 when releasing inference versions to halve the size.
50%UnverifiedDuring SkyRL's weight synchronization, the FP32 router weights are fully preserved to avoid numerical instability caused by low-precision conversion.
50%UnverifiedScoring outputs are more sensitive to numerical perturbations, and low-precision quantization easily causes shifts in score ranking, while ranking accuracy is the core metric for filtering and candidate ranking tasks.
50%UnverifiedCompute-bound means the hardware's compute units are fully saturated, at which point increasing batch size or optimizing numerical precision (such as using BF16 instead of FP32) becomes meaningful.
50%UnverifiedWhen the cumulative sum reaches 1.0, the smallest change representable by FP32 is 1 ULP = 2⁻²³ ≈ 0.00000012.
50%UnverifiedFP32 uses 32 bits to represent floating-point numbers, consisting of 1 sign bit, 8 exponent bits, and 23 mantissa bits.
50%UnverifiedFP32 can only maintain approximately 7 significant digits of precision.
50%UnverifiedUnder a 7nm process, a single FP32 multiplication consumes approximately 4.6 pJ of energy, while a single FP32 addition requires only approximately 0.9 pJ.
50%UnverifiedFull-precision weights typically refer to model parameters stored in FP32 (32-bit floating point) format.
50%UnverifiedIn mixed-precision training, the forward pass uses FP16 to leverage Tensor Core acceleration while parameter updates retain FP32 to maintain numerical stability, which has become standard practice for large model training.
50%